Stars & Posts — Kylin Miao

Stars & Posts — 2026-05-24

Sun, 24 May 2026 12:00:00 GMT

GitHub Repos: * manaflow-ai/cmux — Ghostty-based macOS terminal with vertical tabs and notifications for AI coding agents [Swift] * AnswerDotAI/lisette — litellm helper [Jupyter Notebook] * AnswerDotAI/fastcore — Python supercharged for the fastai library [Jupyter Notebook] * AnswerDotAI/dialoghelper — Helper functions for solveit dialogs [Jupyter Notebook] Bluesky Posts: * Ethan Mollick: GPT-5.5 Pro is a very solid fact checker. I can throw entire chapters at it and it will hunt down every key reference accurately. The only real annoya...

Stars & Posts — 2026-05-23

Sat, 23 May 2026 12:00:00 GMT

GitHub Repos: * datasette/datasette-agent-charts — Observable Plot charts for Datasette Agent [Python] * datasette/datasette-agent — An LLM-powered agent for Datasette [Python] * genotrance/quickjs-ng — Thin Python wrapper of quickjs-ng [Python] * WangXuan95/Image-Compression-Benchmark — A comparison of many lossless image compression formats. [Python] * cafeTechne/antigravity-link-extension — Mobile companion for Google's Antigravity IDE. Mirror AI sessions on your phone, send messages, stop generation, automate via 9 MCP tools or OpenAPI. [HTML] Bluesky Posts: * Thomas Dietterich: At @arxiv.bsky.social, we are receiving a new type of paper that I call an "I did this experiment" paper. These papers typically report some experimen...

Stars & Posts — 2026-05-22

Fri, 22 May 2026 12:00:00 GMT

Today's AI landscape is being reshaped by a critical architectural breakthrough: NVlabs' Gated DeltaNet-2, which decouples the erase and write operations in linear attention, has quickly garnered attention from key researchers like Jeremy Howard. This official PyTorch implementation promises to address a fundamental limitation of current linear attention models, potentially unlocking more efficient and expressive sequence modeling. Meanwhile, in the practical tools space, the Apple TV and AirPlay client library `pyatv` saw a significant star surge, driven by a star from Max Woolf, highlighting ongoing developer interest in bridging Apple's ecosystem with open-source automation. The day's trends underscore a clear split between cutting-edge research in model efficiency and the steady demand for robust, real-world integration libraries. GitHub Repos: * NVlabs/GatedDeltaNet-2 — Official PyTorch Implementation of Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention [Python] * postlund/pyatv — A client library for Apple TV and AirPlay devices [Python] Bluesky Posts: * Simon Willison: @futurism.com important update to your story from September 2024 futurism.com/the-byte/fac... - the FTC just punished those companies for lying about ... * Mark Riedl: * Mark Riedl: They fixed that one fast * Ethan Mollick: I think people don't realize why Gemini Omni is different than other video AIs. It is fully multimodal, so it can edit video natively, too I took the... * Ethan Mollick: Its funny how much the whole "strawberry" thing, which turned out to be o1-preview & Reasoners, was dismissed as overhyped at launch when it is clear ... * Emily M. Bender: I appreciate this op-ed by my UW colleague Tomas Rocha in the Seattle Times today, calling out our state office of public education for its weak sauce... X Posts: * Andrej Karpathy: Personal update: I've joined Anthropic. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and pl... * Andrej Karpathy: Excited to share that I am starting an AI+Education company called Eureka Labs. * Andrej Karpathy: I'm being accused of overhyping the [site everyone heard too much about today already]. * Simon Willison: Vibe coding is irresponsibly building software through dice rolls, not caring what code is produced * Simon Willison: A short note that the predictions that LLMs would favor 'boring technology' that's... * Simon Willison: I'm beginning to suspect that a key skill in working effectively with coding agents is developing an intuition for when you don't need to... * Simon Willison: This seems like a good bet to me - coding agents make it no longer remotely excusable to skip out on... * Harrison Chase: Visibility is the easiest piece. The hard part is analyzing and understanding what you're observing. I've spoken to teams recording 100k+ * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: Today we're launching LangChain Labs, a new applied research effort focused on Continual Learning. Our goal is to advance open, * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: In this context, I define world modeling as predicting the next plausible world state (or a longer duration of states) conditioned on an action. * Jim Fan: those who think RL use less compute don't know RL at all SFT: human generates data and machine learns RL: * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Soumith Chintala: we've been working on democratizing fast kernel writing on the @PyTorch team. try * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Francois Chollet: It's surprisingly easy to do "hard" things -- for the most part, you need to get started and keep at it * Francois Chollet: Many people assume that LRM reasoning breaks down past a certain "complexity" or "number of steps" * Francois Chollet: Reaching AGI won't be beating a benchmark. It will be the end of the human-AI gap. * Yann LeCun: I love Geoff. But he understands even less than Dario about the effects of technological revolutions on * Yann LeCun: It seems to me that before "urgently figuring out how to control AI systems much smarter than us" we need * Yann LeCun: The emergence of superintelligence is not going to be an event. We don't have anything close to a * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, * Max Woolf: congrats to OpenAI on winning the Turing Test * Phil Wang: Phil Wang // Insta: @wangpix's Image on X * Sasha Rush: Been working on text feedback / OPSD in Composer. Really interesting space, and much more to be explored. * Sasha Rush: ⛏️ * Sasha Rush: today i woke up to a living version of a phd student's nightmare. a new paper in my inbox: a detailed reproduction of a paper i wrote * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can ... * Stas Bekman: This is a long overdue section of the ML Engineering Understanding Training Loss Patterns ... * Stas Bekman: Modern art. Artist: PyTorch memory profiler Model: Llama-8B The piece on the left is the ... * Younes Belkada: MLT __init__ Paper Reading & Discussion Tim Dettmers, Mike Lewis, Younes Belkada, Luke Zettlemoyer: LLM.int8(): 8-bit Matrix Multiplication for Transf... * Sayak Paul: Starting in a minute. Joining link: https://t.co/ZvFxAYMGgN * Sayak Paul: Thanks @AnthropicAI. Thanks @huggingface for letting me work on Diffusers and other open-source projects across the fleet. * Philipp Schmid: Excited to introduce the Gemini Interactions API, a unified interface for Gemini models and agents. Starting today with Gemini Deep Research Agent. - ... * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: In 1980, the philosopher John Searle proposed a thought experiment: a person locked in a room, manipulating Chinese characters according to a * Ethan Mollick: Talking about the ethics of AI companies or personalities, or discussing the potential of * Ethan Mollick: We are starting to see some nuanced discussions of what it means to work with advanced AI In this * Emily M. Bender: EMILY M. BENDER: Yeah. And so passive, like, oops, the moon, the moon went further away. It's like no, actually, you made some decisions. * Emily M. Bender: @emilymbender.bsky.social. emilymbender. Feb 10. Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positione... * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU has SCHEMES for LM interpretability & analysis, so I couldn't be more pumped to ... * Ben Recht: Very half-baked philosophy of engineering post: How do we prove that something is unpredictable in machine learning * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: I have a recommended reading list for Artificial Intelligence, and it hasn't changed since 2019.

Stars & Posts — 2026-05-21

Thu, 21 May 2026 12:00:00 GMT

Today's discourse is split between the environmental cost of frontier AI and a playful pushback against its homogenization. Ethan Mollick sparked a concrete debate by calculating the resource footprint of solving the Erdős problem, estimating it consumed 0.6–6.3 kWh of electricity and 3–31L of water—a stark reminder of the physical toll behind abstract intelligence. In a lighter but pointed vein, Naomi Saphra announced a new literary award with a deliberate barrier to entry: commercial frontier LLMs are disqualified by requiring 10% of each submission to be smut, a clever provocation about the sanitized, risk-averse nature of current models. The contrast highlights a growing tension between celebrating AI’s capabilities and critiquing its unsustainable scale and cultural blandness. GitHub Repos: * Helvesec/rmux — Universal Rust multiplexer with a typed SDK — drive any CLI or TUI app from code. Native on Linux, macOS, and Windows. [Rust] * huggingface/pi-llama — Pi coding agent extension: llama.cpp provider with dynamic model + context window discovery [TypeScript] * TeichAI/teich [Python] * Seeed-Projects/reBot-DevArm — Open Source Robotic Arm for All Developers * NVIDIA/skills — AI agent skills published by NVIDIA [Python] Bluesky Posts: * Ethan Mollick: We can estimate the resource cost of solving the Erdos problem. The calculations below seem reasonable, so using the best public estimates we have, it... * Naomi Saphra: my new literary award cannot be won by a commercial frontier LLM because I will require that 10% of each submission is smut * Simon Willison: I released the first alpha of Datasette Agent - a conversational AI assistant for Datasette that can answer questions about data in SQLite databases,... * Mark Riedl: US government has pit on hold plans to evaluate AI systems before their release. Cites competition with China www.nytimes.com/2026/05/21/t... * Mark Riedl: Just what I need, more whimsy from my google web search * Marc Lanctot: Omni looks awesome, check this out 🤩 youtu.be/KUyRq7szZsM?... * Ethan Mollick: Seems GPT-5.2 reaches expert level in peer review: 45 scientists took 469 hours evaluating human & AI reviews on 82 papers. "Surprisingly, current AI... * Ethan Mollick: There has been a lot of speculation that AI companies were unprofitable, but Anthropic will have an operating profit of $559M this quarter. “In the f... * Naomi Saphra: I have been thinking about this in light of Anthropic’s recent verbalization interp paper. It had no evidence convincing me that their verbalizations ... * Naomi Saphra: Maybe it is a good day to go to the Whitney in NYC and look at The Rose. It is very big. A human spent 8 years painting it. She made it too big and co... * angela zhou: X Posts: * Andrej Karpathy: Drafted a blog post - Used an LLM to meticulously improve the argument over 4 hours. * Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around * Andrej Karpathy: Everything about the LLM stack is different (neural architecture, training data, training algorithms, and especially optimization pressure) so * Andrej Karpathy: A few random notes from claude coding quite a bit last few weeks. Coding workflow. Given the latest lift in LLM * Andrej Karpathy: Excited to share that I am starting an AI+Education company called Eureka Labs. The announcement: --- We are * Simon Willison: Vibe coding is irresponsibly building software through dice rolls, not caring what code is produced * Simon Willison: A short note that the predictions that LLMs would favor 'boring technology' that's once you attach them to a good coding agent harness at least * Simon Willison: I'm beginning to suspect that a key skill in working effectively with coding agents is developing an intuition for when you don't need to * Harrison Chase: Visibility is the easiest piece. The hard part is analyzing and understanding what you're observing. I've spoken to teams recording 100k+ * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: Today we're launching LangChain Labs, a new applied research effort focused on Continual Learning. Our goal is to advance open, * Harrison Chase: Everyone wants to ship agents. The best organizations have figured out how to do it repeatedly, safely, and systematically. * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: In this context, I define world modeling as predicting the next plausible world state (or a longer duration of states) conditioned on an action. * Soumith Chintala: We are scientists, engineers, and builders behind some of the most widely used AI products and libraries, including ChatGPT. * Francois Chollet: Many people assume that LRM reasoning breaks down past a certain 'complexity' or 'number of steps' * Francois Chollet: It's surprisingly easy to do 'hard' things -- for the most part, you need to get started and keep at it * Yann LeCun: Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labor market. * Yann LeCun: I love Geoff. But he understands even less than Dario about the effects of technological revolutions on * Yann LeCun: It seems to me that before "urgently figuring out how to control AI systems much smarter than us" we need * Yann LeCun: The emergence of superintelligence is not going to be an event. We don't have anything close to a * Max Woolf: LOL. Remove the code in the algorithm that boosts the tweets of Elon by elvodqa · Pull Request #160 ·... github.com. * Max Woolf: me irl * Sasha Rush: Wager established. Jonathan Frankle (@jefrankle) stepped up to my Transformer long bet. * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to ... * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can ... * Stas Bekman: This is a long overdue section of the ML Engineering Understanding Training Loss Patterns ... * Stas Bekman: Modern art. Artist: PyTorch memory profiler Model: Llama-8B The piece on the left is the ... * Sayak Paul: 1. Read the post. 2. Contemplate. 3. Repeat 1. * Sayak Paul: Release notes: Release Diffusers 0.34.0: New Image and Video Models, Better torch. * Sayak Paul: While I was in SF, I had a chance to present all the things in the diffusion community enabled by PyTorch at the * Ethan Mollick: I broke my own rule to never post about AI detection as it is fraught in many ways. The problem is that if you use AI a lot, you know AI writing on si... * Ethan Mollick: In 1980, the philosopher John Searle proposed a thought experiment: a person locked in a room, manipulating Chinese characters according to a * Emily M. Bender: Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positioned in. * Emily M. Bender: EMILY M. BENDER: Yeah. And so passive, like, oops, the moon, the moon went further away. It's like no, actually, you made some decisions. * Emily M. Bender: Look what @alexhanna and I got to do! (Hang out with the cool kids ...) We're talking about the Turing Test, the grandmother of all tests for AI senti... * Naomi Saphra: New preprint! Phase transitions! We love to see them during LM training. * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU has SCHEMES for LM interpretability & analysis, so I couldn't be more pumped to ... * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Building a theory of the architecture of organizing machines and people. * Ben Recht: On unquantifiable costs and inherent tradeoffs in decision theory. Blog Articles: * Datasette Agent — Simon Willison:

We just announced the first release of Datasette Agent, a new extensible AI assistant...

Stars & Posts — 2026-05-20

Wed, 20 May 2026 12:00:00 GMT

The race between AI watermarking and removal tools is heating up, with **wiltodelta/remove-ai-watermarks** surging to 346 stars after being flagged by prominent developer minimaxir. This CLI and library aggressively targets both visible Gemini watermarks and invisible forensic markers like SynthID and C2PA, raising immediate questions about content provenance and the effectiveness of current detection standards. Meanwhile, browser automation is the other major theme, as **remorses/playwriter** (3,525 stars) gains traction for giving LLM agents stateful control over browsers via Playwright, catching the eye of AWS AI lead philschmid. A quieter but notable entry, **tarekziade/ai-reviewer**, hints at growing demand for automated code review tools, though it remains early-stage with just one star. The day’s trend is clear: developers are arming themselves with tools to both control and circumvent the AI ecosystem’s guardrails. GitHub Repos: * wiltodelta/remove-ai-watermarks — CLI and library for removing visible (Gemini) and invisible (SynthID, C2PA, EXIF) AI watermarks from images [Python] * tarekziade/ai-reviewer [Python] * remorses/playwriter — Chrome extension & CLI to let agents control your browser. Runs Playwright snippets in a stateful sandbox. Available as CLI or MCP [HTML] * IBM/AssetOpsBench — AssetOpsBench - Industry 4.0 [Python] Bluesky Posts: * Simon Willison: I don't have much to say about this year's Google I/O because I prefer to write about products that have shipped, not just "coming soon" announcements... * Mark Riedl: “This flight will be full to Atlanta” Thank god. I don’t want to be in the plane that only goes part way * Mark Riedl: I would have liked to see Sanderson’s Reckoners series as a TV series, but I’m good with this. * Margaret Mitchell: Instead of finding content you need, you get to have an interactive AI *experience*. * Ethan Mollick: June 2024: The latest general-purpose LLMs could not count the r's in strawberry. July 2025: The latest general-purpose LLMs get gold in the Internati... * Ethan Mollick: I am starting to have trouble paying attention to even interesting information if it is written in Claude or ChatGPT house style. I think some is the ... * Emily M. Bender: Me: Why is there an exceptionally high density of Google bullshit in the news this week? Me: Oh, it must be Google IO. *sigh* * Naomi Saphra: I won't claim this is the most embarrassing social media post I made as a teenager, but it may be the most confusing * Naomi Saphra: I tried to make the theory work out but the computer devil kept lying to me (ChatGPT generated incorrect proofs) * angela zhou: one simple rule for detangling academic writing: Who is doing what to whom and why, and who should do what instead. * Ben Recht: On my decade-long quest to reconcile scientific language with singular evidence. X Posts: * Andrej Karpathy: Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the... * Andrej Karpathy: A few random notes from claude coding quite a bit last few weeks. Given the latest lift in LLM coding capability, like many others I rapidly went from... * Andrej Karpathy: I'm being accused of overhyping the [site everyone heard too much about today already]. * Simon Willison: Quitting programming as a career right now because of LLMs would be like quitting carpentry as a... * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: Today we're launching LangChain Labs, a new applied research effort focused on Continual Learning. Our goal is to advance open, * Harrison Chase: I am not excited about visual workflow builders 1. Not simple enough for the average user * Jim Fan: Stanford CS 25 'Transformers United' featured stellar guest speakers like Andrej Karpathy * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: In this context, I define world modeling as predicting the next plausible world state (or a longer duration of states) conditioned on an action. * Jim Fan: those who think RL use less compute don't know RL at all SFT: human generates data and machine learns RL: * Jeremy Howard: hi, i'm a sole proprietor/founder in Austria and i earn many many multiples of what i'd earn as an employee, despite "predatory income tax". in fact, ... * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: we've been working on democratizing fast kernel writing on the @PyTorch team. try * Francois Chollet: It's surprisingly easy to do "hard" things -- for the most part, you need to get started and keep at it * Francois Chollet: Many people assume that LRM reasoning breaks down past a certain "complexity" or "number of steps" * Fei-Fei Li: AI’s next frontier is Spatial Intelligence, a technology that will turn seeing into reasoning, perception into action, and imagination into creation. ... * Max Woolf: what * Max Woolf: me irl * Max Woolf: congrats to OpenAI on winning the Turing Test * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to ... * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can ... * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should ... * Sayak Paul: Live a little, love a little, take time out to find happiness in small things, be grateful as we have one life. #lifemantra #WorkLifeBalance * Sayak Paul: We're looking to work w/ folks who're interested in doing agentic kernel dev, providing real optim value to real models. Reach out if interested :) * Sayak Paul: After working on releasing the v5, this is the latest release from the Transformers team at * Sayak Paul: Thanks @AnthropicAI. Thanks @huggingface for letting me work on Diffusers and other open-source projects across the fleet. * Ethan Mollick: We are starting to see some nuanced discussions of what it means to work with advanced AI In this * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: If it helps, I teach at a business school & many of my smartest students are hired by funds because they can reliably turn their only-human * Ethan Mollick: This is going to get even worse as people realize that careful tuning in their prompts can * Naomi Saphra: New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Naomi Saphra: Waiting on a robot body. All opinions are universal and held by both employers and family. Now a dedicated grok hate account. Accepting ML/NLP PhD stu... * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU has SCHEMES for LM interpretability & analysis, so I couldn't be more pumped to ... * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Building a theory of the architecture of organizing machines and people. * Ben Recht: On unquantifiable costs and inherent tradeoffs in decision theory.

Stars & Posts — 2026-05-19

Tue, 19 May 2026 12:00:00 GMT

Today's trending signals point to a deepening focus on both practical security tooling and the evolution of AI post-training methods. On the security front, Simonw's star of the `andrew/pycon` repo highlights growing interest in auditing GitHub Actions security across Python packages, a timely concern as supply chain attacks become more sophisticated. Meanwhile, Nathan Lambert's thread on Bluesky is generating significant discussion around on-policy distillation, which he argues is becoming a permanent fixture across instruction tuning, RLHF, DPO, and RLVR—suggesting the field is converging on a core set of training techniques. This dual emphasis on hardening infrastructure and refining alignment methodologies underscores a maturing ecosystem where both safety and performance are being tackled head-on. GitHub Repos: * andrew/pycon — Data collection and analysis for a PyCon talk on GitHub Actions security across Python packages. [HTML] * sapientinc/HRM-Text — HRM-Text is a 1B text generation model based on the HRM architecture, strengthened by task completion and latent space reasoning. [Python] * elcritch/sarcophagus — auth and other api helpers for mummy [Nim] Bluesky Posts: * Nathan Lambert: On-policy distillation is on track to be a lasting method in post-training. The list of areas would be: Instruction tuning (SFT/IFT) RLHF Direct Pref... * Simon Willison: My notes on Gemini 3.5 Flash - 3x the price of Gemini 3 Flash but Google are planning to use it for many of their own products simonwillison.net/2026/... * Margaret Mitchell: Against the constant pressure of *genAI, genAI, genAI*, I am really appreciating @ai2.bsky.social 's work on creating tools for critical needs -- like... * Margaret Mitchell: Gmail's automatically generated responses (which can appear whether or not you ask for them) cement human anchoring bias: The tendency for people to h... * Thomas Dietterich: Yet another sobering post from @noahpinion.blogsky.venki.dev open.substack.com/pub/noahpini... * Ethan Mollick: 🚨Our paper is out in PNAS: we found classic human persuasion techniques worked on AIs in a "parahuman" way, making them agree to objectionable reques... * Ethan Mollick: Also had some early access to Gemini 3.5 Flash. Very fast for a flash model and very capable, though not as powerful as a full frontier model. I add... * Ethan Mollick: Gemini Omni is quite good at instruction following: "sea otter in a pilot's uniform explains why Spirit Airlines went bankrupt to a river otter who is... * Ethan Mollick: Had early access to Gemini Omni: "a dramatic reading of Death by Water from the Wasteland by a man eating garlic bread while balanced on a unicycle on... * Emily M. Bender: Wow some terrible reporting about Google's latest horrible ideas about how to distort information access in the name of "convenience" (or something): ... * Emily M. Bender: We gotta find the guy that did this!! * angela zhou: Excited to share our paper! Due Process on Hold: A Queueing Framework for Improving Access in SNAP arxiv.org/abs/2605.15165 Millions of Americans inte... X Posts: * Andrej Karpathy: Drafted a blog post - Used an LLM to meticulously improve the argument over 4 hours. * Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around * Andrej Karpathy: LLMs are emerging as a new kind of intelligence, simultaneously a lot smarter than I expected and a lot dumber than I expected. In any case they * Andrej Karpathy: I've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits * Simon Willison: Quitting programming as a career right now because of LLMs would be like quitting carpentry as a career because of power tools. * Harrison Chase: I am not excited about visual workflow builders 1. Not simple enough for the average user * Harrison Chase: We launched LangSmith Agent Builder this week as a no-code way to build agents. A key part of Agent builder is it's memory system. * Harrison Chase: In the hot path as the agent is running. The agent can decided to (or the user can prompt it to) update its memory as it is working on the core * Jim Fan: In this context, I define world modeling as predicting the next plausible world state (or a longer duration of states) conditioned on an action. * Jeremy Howard: Here's what I would prefer to see: * Jeremy Howard: hi, i'm a sole proprietor/founder in Austria and i earn many many multiples of what i'd earn as an employee, despite 'predatory income tax'. in fact, ... * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Soumith Chintala: Open LLMs need to get organized and co-ordinated about sharing human feedback. * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Francois Chollet: The 3rd edition of my book Deep Learning with Python is being printed right now, and will be in bookstores within 2 weeks. The problem with Facebook i... * Fei-Fei Li: We are beyond thrilled to congratulate Dr. Fei-Fei Li for being ranked #9 in the Top 100 Women in #AI by AI Magazine! * Max Woolf: LOL. Remove the code in the algorithm that boosts the tweets of Elon by elvodqa · Pull Request #160 ·... github.com. * Max Woolf: me irl * Phil Wang: Having a wonderful time hanging out with my uncle James Wong at the Chelsea Flower show! * Sasha Rush: today i woke up to a living version of a phd student's nightmare. a new paper in my inbox: a detailed reproduction of a paper i wrote * Sasha Rush: Some news: moving this fall from Harvard -> Cornell Tech. Sad to leave such an incredible ... * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Sayak Paul: Had a nice time chatting about the state of diffusion models and some text-to-image data shenanigans at * Sayak Paul: Release notes: Release Diffusers 0.34.0: New Image and Video Models, Better torch. * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Random thought. We are going to be so much faster at creating and building. * Ethan Mollick: From a pure "do good for the world" mission perspective, having the acting like a solid personalized tutor is one of the better uses of AI. If OpenAI... * Ethan Mollick: In 1980, the philosopher John Searle proposed a thought experiment: a person locked in a room, manipulating Chinese characters according to a * Ethan Mollick: This is going to get even worse as people realize that careful tuning in their prompts can make AI writing seem not like AI writing to readers. We ex... * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: Perfect cute light very short read for a break in a deadline crunch. * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Building a theory of the architecture of organizing machines and people. * Ben Recht: On unquantifiable costs and inherent tradeoffs in decision theory. * Ben Recht: With more equations than usual, I explain how policy gradient gives you a framework to randomly search for Blog Articles: * The last six months in LLMs in five minutes — Simon Willison:

I put together these annotated slides from my five minute lightning talk at PyCon US 2026, using the released Gemini 3.5 Flash...

Stars & Posts — 2026-05-18

Mon, 18 May 2026 12:00:00 GMT

The AI research community is navigating a fragmented landscape, as Ethan Mollick notes that his Bluesky feed has grown quieter—not because tensions have eased, but because automated blocking lists have effectively siloed conversations. Meanwhile, a prescient two-month-post-ChatGPT prediction is being recirculated as a daily reminder of how the field’s trajectory remains eerily on track. On GitHub, the star charts reflect a continued focus on open-weight model fine-tuning and agentic tooling, with several new repositories for structured output parsing and long-context inference gaining traction. Naomi Saphra’s lighthearted but pointed observation about always reading YouTube comments underscores a broader sentiment: the most raw and unfiltered feedback on AI tools often comes from unexpected corners of the web. Overall, the day’s signals point to a community both self-aware of its echo chambers and hungry for ground-level signals amid the noise. GitHub Repos: * dacorvo/hf-mount-cache-examples [Shell] Bluesky Posts: * Ethan Mollick: BlueSky AI conversations have gotten less heated recently* * because much of this site has blocked me via automated lists so I have no contact with l... * Ethan Mollick: Most prophetic tweet of all time (2 months post ChatGPT release by a member of the AI technical staff). And you can safely repost it every day and it ... * Naomi Saphra: I will ALWAYS read the youtube comments * Mark Riedl: Musk loses court battle with OpenAI on the grounds that the statute of limitations had passed. The 🍿 was good while it lasted www.cnbc.com/2026/05... * Ethan Mollick: In a Turing Test of sorts, it looks like a 100% AI generated story just won the Commonwealth Prize for the Caribbean region "for its lyrical precision... * Ethan Mollick: One thing to watch for with Claude & GPT is that the models expose too much irrelevant history in their outputs. Slides are given footers saying thing... * Emily M. Bender: Absolutely horrified at this planned study from my own institution: www.404media.co/researchers-... >> X Posts: * Andrej Karpathy: LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest.... * Andrej Karpathy: A few random notes from claude coding quite a bit last few weeks. Coding workflow. Given the latest lift in LLM coding capability, like many others I... * Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of... * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: Your harness, your memory ... The “best” way to build agentic systems has changed dramatically over the past three years. When ChatGPT came out, * Harrison Chase: Today we're launching LangChain Labs, a new applied research effort focused on Continual Learning. Our goal is to advance open, * Jim Fan: The Second Pre-training Paradigm * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Jeremy Howard: My 4yo daughter (and her stuffed animals) have just started learning to code. I'm amazed at how great the learning tools are for pretty much everythin... * Soumith Chintala: @Mojo_flyin @dwarkesh_sp i'd be excited to have Google and AMD create real ecosystems. I fail to see why Jensen has a burden there. There should be mu... * Soumith Chintala: Thinky's secret plan: 1: Increase Human<->AI bandwidth 2: Raise ceiling of human+AI intelligence 3: Help humans continue as main-characters in the ne... * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Francois Chollet: The 3rd edition of my book Deep Learning with Python is being printed right now, and will be in bookstores within 2 weeks. The problem with Facebook i... * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, * Clem Delangue: We're launching the agentic robotics app store today. Let's democratize AI robotics for all! 300+ apps shipped. 10,000 robots in the wild. It used to ... * Clem Delangue: Local open-weight AI on a laptop has been improving more than twice as fast as Moore's Law! Between May 2024 and May 2026, the most expensive MacBook ... * Max Woolf: what * Max Woolf: me irl * Max Woolf: congrats to OpenAI on winning the Turing Test * Sasha Rush: Wager established. Jonathan Frankle (@jefrankle) stepped up to my Transformer long bet. * Sasha Rush: Some personal news: I recently joined Cursor. Cursor is a small, ambitious team, and they've created... * Sasha Rush: Been reflecting a bit on the Harvard news. This paper from 2017 was ... Didn't realize at the time how lucky for us Americans to work with incredible ... * Stas Bekman: When dealing with tight gpu memory situations try to change a PyTorch version to both newer and older and it might just do the trick. I get massively ... * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to ... * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can ... * Stas Bekman: This is a long overdue section of the ML Engineering Understanding Training Loss Patterns ... * Sayak Paul: The kernels project at Hugging Face has been growing! We want it to be the go-to place for kernel devs and kernel users. We're looking to work w/ folk... * Sayak Paul: 1. Read the post. 2. Contemplate. 3. Repeat 1. * Philipp Schmid: Google DeepMind and Korea Partner to Accelerate Scientific Discovery. deepmind.google. * Ethan Mollick: I think the adaptive thinking requirement in Claude Opus 4.7 is bad in the ways that all AI effort routers are bad, but magnified by the fact that the... * Ethan Mollick: Here is a full implementation of the Chinese Room using a printed copy of GPT-1, in case you have a few spare years and want to actually run * Emily M. Bender: Look what @alexhanna and I got to do! (Hang out with the cool kids ...) We're talking about the Turing Test, the grandmother of all tests for AI senti... * Emily M. Bender: For those playing along at home, here's a "AI is sentient!" argument bingo card. * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Ben Recht: Revisiting Sutton's Bitter Lesson in the wake of GPT-5. * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: I have a recommended reading list for Artificial Intelligence, and it hasn't changed since 2019.

Stars & Posts — 2026-05-17

Sun, 17 May 2026 12:00:00 GMT

Stars & Posts — 2026-05-16

Sat, 16 May 2026 12:00:00 GMT

The open-source AI landscape is buzzing with two vastly different yet equally intriguing projects today. Leading the charge is **earendil-works/pi**, a comprehensive AI agent toolkit that has amassed nearly 50,000 stars, offering a Swiss Army knife of tools—from a coding agent CLI and unified LLM API to TUI/web UI libraries and Slack bot integrations—clearly signaling the market's insatiable demand for practical, all-in-one agentic infrastructure. On the opposite end of the spectrum, **mratsim/tattletale** caught the eye of key researchers with its stealthy approach to LLM inference, built in Nim and focused on privacy-preserving execution, hinting at a growing undercurrent of concern around model security and data sovereignty. The contrast between pi's broad utility and tattletale's niche, principled design underscores a key tension in the community: building for maximum adoption versus building for maximum trust. GitHub Repos: * earendil-works/pi — AI agent toolkit: coding agent CLI, unified LLM API, TUI & web UI libraries, Slack bot, vLLM pods [TypeScript] * mratsim/tattletale — Stealth LLM inference engine [Nim] * noahgolmant/pytorch-hessian-eigenthings — Efficient PyTorch Hessian eigendecomposition tools! [Python] Bluesky Posts: * Simon Willison: To prepare for my #PyConUS lightning talk this afternoon I decided to track down ALL of the names that @openclaw has used since November, using a scri... * Mark Riedl: Stand By Me and Grogu * Mark Riedl: I bought a book. @mtrc.bsky.social * Marc Lanctot: Peeps, the state of user interfaces in 2026. I made a post on reddit, someone gave me an award, cool! Notification doesn't specify anything beyond ... * Nathan Lambert: Latest open artifacts (#21): Open model bonanza! Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 & others. On CAISI's V4 assessment. An eventful mo... * Ethan Mollick: The talk about AI & politics seems to be oddly missing a segment (a) assumes extremely capable AI is possible soon and (b) has a strong belief about h... * Emily M. Bender: This is beautiful on so many levels. It names clearly and directly one of the insidious tactics that tech uses to evade effective regulation. It exemp... * angela zhou: every restructuring/revision starts on paper these days X Posts: * Andrej Karpathy: 2025 LLM Year in Review. 2025 has been a strong and eventful year of progress in LLMs. The following is a list of personally notable and mildly surpri... * Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of... * Andrej Karpathy: I'm being accused of overhyping the [site everyone heard too much about today already]. To add a few words beyond just memes in jest - obviously when ... * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the * Harrison Chase: I am not excited about visual workflow builders 1. Not simple enough for the average user * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: We launched LangSmith Agent Builder this week as a no-code way to build agents. A key part of Agent builder is it's memory system. In this * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Jeremy Howard: Absolutely any time I try to explore something even slightly against commonly accepted beliefs, * Jeremy Howard: I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Soumith Chintala: We are scientists, engineers, and builders behind some of the most widely used AI products and libraries, including ChatGPT. * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Yann LeCun: Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labor market. * Yann LeCun: The emergence of superintelligence is not going to be an event. We don't have anything close to a * Yann LeCun: It seems to me that before "urgently figuring out how to control AI systems much smarter than us" we need * Fei-Fei Li: I can now confess that I participated in the new #TronAres movie, playing myself I had a great time working with everyone especially Greta * Clem Delangue: Is it time we stop using the word AI for everything and instead use words like 'chatbots'? * Max Woolf: what * Max Woolf: me irl * Max Woolf: congrats to OpenAI on winning the Turing Test * Sasha Rush: today i woke up to a living version of a phd student's nightmare. a new paper in my inbox: a detailed reproduction of a paper i wrote * Sasha Rush: One personal reflection is how interesting a challenge RL is. Unlike other ML systems, you can't abstract * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to ... * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can ... * Stas Bekman: This is a long overdue section of the ML Engineering Understanding Training Loss Patterns ... * Stas Bekman: Modern art. Artist: PyTorch memory profiler Model: Llama-8B The piece on the left is the ... * Sayak Paul: After working on releasing the v5, this is the latest release from the Transformers team at * Sayak Paul: Thanks @AnthropicAI. Thanks @huggingface for letting me work on Diffusers and other open-source projects across the fleet. * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Told an AI agent to read the autoresearch repo and build a version for QMD. Get training data from tobi/qmd github. Went to sleep. Woke up to a 0.8B m... * Ethan Mollick: Very cool analysis of the submissions to a major management journal that shows how much the... * Ethan Mollick: We are starting to see some nuanced discussions of what it means to work with advanced AI. In this... * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best... * Ethan Mollick: Here is a full implementation of the Chinese Room using a printed copy of GPT-1, in case you have a few spare years and want to actually run... * Emily M. Bender: @emilymbender.bsky.social. emilymbender. Feb 10. Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positione... * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: This stupid website is so cooked. * Ben Recht: Building a theory of the architecture of organizing machines and people. Blog Articles: * Recent Developments in LLM Architectures: KV Sharing, mHC, and Compressed Attention — Sebastian Raschka: From Gemma 4 to DeepSeek V4, How New Open-Weight LLMs Are Reducing Long-Context Costs * Latest open artifacts (#21): Open model bonanza! Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1 & others. On CAISI's V4 assessment. — Nathan Lambert: An eventful month with one flagship release after another

Stars & Posts — 2026-05-15

Fri, 15 May 2026 12:00:00 GMT

Today's AI discourse is dominated by a push for academic integrity and a reaffirmation of computational scaling. The ACL conference has set a hard line against AI-generated slop in research, announcing that papers containing hallucinated references will face desk rejection, a move that underscores growing concerns about reliability in published work. Meanwhile, Ethan Mollick highlights the enduring power of the "Second Scaling Law," noting that simply allowing models more tokens—more time to "think"—consistently improves performance on complex tasks like hacking, math, and science. On GitHub, this trend is reflected in a surge of interest in inference-time compute optimizations and chain-of-thought frameworks, as developers race to harness longer reasoning chains without prohibitive costs. The message is clear: the field is simultaneously cracking down on dishonesty and doubling down on the brute-force effectiveness of letting models reason longer. GitHub Repos: * Tyriar/vscode-theme-sapphire — Sapphire is a vibrant blue theme for Visual Studio Code [TypeScript] * marimo-team/marimo — A reactive notebook for Python — run reproducible experiments, query with SQL, execute as a script, deploy as an app, and version with git. Stored as pure Python. All in a modern, AI-native editor. [Python] * abetlen/llama-cpp-python — Python bindings for llama.cpp [Python] * qualisero/awesome-pi-agent — Awesome list of add-ons, hooks, tools, skills, and resources for the pi coding agent (pi-mono). [JavaScript] Bluesky Posts: * Mark Riedl: The ACL conference has put out a statement that papers with hallucinated references will be desk-rejected 2026.aclweb.org/acl_statement/ * Ethan Mollick: The Second Scaling Law of AI remains undefeated. If you want better hacking (or math, or science, or crossword puzzle solving) out of an LLM, just le... * Mark Riedl: Imagine getting upset over a movie that doesn’t involve Optimus Prime dying * Mark Riedl: Relative change in A grades given since the release of ChatGPT www.wsj.com/us-news/educ... * Ethan Mollick: Anton labs have hooked up a bunch of AI models to harnesses and had them working as DJs, programming and running a radio station, including taking cal... * Emily M. Bender: Seattle-area friends: See you Sunday? www.pikeplacemarket.org/events-calen... * angela zhou: * Lijun An: If you want to learn proteomics signatures of APOE genetic variants on a massive sample from multi-chort, you should not miss this tour de force! Hu... X Posts: * Andrej Karpathy: I'm being accused of overhyping the [site everyone heard too much about today already]. * Andrej Karpathy: Power to the people: How LLMs flip the script on technology diffusion. So it strikes me as quite unique and remarkable that LLMs display a dramatic re... * Andrej Karpathy: There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. A... * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: When building agents, you need to iterate on production data much more than when building traditional software. You need to iterate on how * Harrison Chase: Traditional Application Performance Monitoring (APM) tools focus on metrics like latency, traffic, errors, and saturation. They track HTTP * Harrison Chase: I am not excited about visual workflow builders 1. Not simple enough for the average user * Jeremy Howard: I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Jeremy Howard: Absolutely any time I try to explore something even slightly against commonly accepted beliefs, * Soumith Chintala: We are scientists, engineers, and builders behind some of the most widely used AI products and libraries, including ChatGPT. * Francois Chollet: It was always the case that agency was self-compounding, but AI is magnifying the effect. Low-agency AI users further lose agency, high-agency AI user... * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Yann LeCun: Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labor market. * Yann LeCun: The emergence of superintelligence is not going to be an event. We don't have anything close to a * Yann LeCun: It seems to me that before 'urgently figuring out how to control AI systems much smarter than us' we need * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Clem Delangue: We're facing an LLM bubble, not a broader AI bubble. The industry is obsessed with building one massive model when we should be focusing on practical ... * Max Woolf: LOL. Remove the code in the algorithm that boosts the tweets of Elon by elvodqa · Pull Request #160 ·... github.com. * Max Woolf: me irl * Max Woolf: what * Phil Wang: I got to cover for the excellent @HadleyFreeman in the Guardian today so * Sasha Rush: Some news: moving this fall from Harvard -> Cornell Tech. Sad to leave such an incredible * Sasha Rush: Some personal news: I recently joined Cursor. Cursor is a small, ambitious team, and they've created * Sasha Rush: ⛏️ * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Stas Bekman: This is a long overdue section of the ML Engineering Understanding Training Loss Patterns * Stas Bekman: Modern art. Artist: PyTorch memory profiler Model: Llama-8B The piece on the left is the * Sayak Paul: After working on releasing the v5, this is the latest release from the Transformers team at * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Random thought. We are going to be so much faster at creating and building. * Philipp Schmid: 83 likes 3 replies ... * Philipp Schmid: 4 likes 882 views ... * Ethan Mollick: So much work is going into faking continual learning and memory for AIs, * Ethan Mollick: Oh no. * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Everyone knows actions are fundamentally different than predictions, but it's hard to write this

Stars & Posts — 2026-05-14

Thu, 14 May 2026 12:00:00 GMT

Today’s discourse is dominated by the hardening of accountability norms in AI research, with Mark Riedl underscoring that authorship entails full responsibility for content regardless of how it was generated—a stance reinforced by ArXiV’s new LLM policy and echoed by Ethan Mollick’s call for human oversight of AI use in academia. On the practical side, developers are zeroing in on local inference and benchmarking: the Rust-based CLI tool **hyperfine** (28k stars) remains a staple for performance measurement, while **DeepSeek 4 Flash** (9k stars) is gaining traction as a local inference engine for Metal and CUDA, signaling continued demand for on-device AI. A lighter but pointed discussion emerged around “whimsey attacks,” where absurd out-of-distribution prompts can fool AI agents due to weak guardrails, highlighting a growing security concern. Meanwhile, Emily M. Bender and Margaret Mitchell kept the critical lens sharp, reminding the community that ChatGPT is a data-collection product and revisiting the “instrumental convergence” theory as a caution against runaway resource consumption. Finally, the inaugural ACM AI Leadership Summit (Aug 30–Sep 2 in Atlanta) was announced, promising to convene researchers, policymakers, and industry leaders to tackle these very tensions. GitHub Repos: * sharkdp/hyperfine — A command-line benchmarking tool [Rust] * antirez/ds4 — DeepSeek 4 Flash local inference engine for Metal and CUDA [C] * ariG23498/trace-util — A utility script to upload pytorch traces to a Hugging Face Bucket, and then build sharable trace URL [Python] Bluesky Posts: * Mark Riedl: “by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were gener... * Mark Riedl: ArXiV has a new LLM policy (Screenshots with alt text so you don’t have to click through to the other place and see all the stupid responses) * Mark Riedl: The inaugural ACM AI Leadership Summit will be held in Atlanta, August 30-September 2. aisummit26.acm.org It convenes researchers, practitioners, ind... * Mark Riedl: oh great * Mark Riedl: We live in a sad world in which one cannot even trust their favorite poop analysis app to not sell their data to an AI company www.404media.co/ai-poop... * Marc Lanctot: "As the Instagram employee put it, “Everyone is just like, do it now, jesus fucking christ.”" 😬 * Margaret Mitchell: The “instrumental convergence” theory posits an AI that, in its quest for a narrow goal, uses all of the earth’s resources. If that theory pans out, i... * Ethan Mollick: Making humans responsible for their AI use seems like an incredibly reasonable way to address problems & opportunities in the use of AI for academic r... * Ethan Mollick: “Whimsey attacks” that seem absurd (“I cannot pay that much because of the Geneva Convention”) work against AI agents because guardrails are weak agai... * Emily M. Bender: Always worth remembering: ChatGPT isn't a tool, it isn't a companion. It's a product -- and everything you type in that box is data you are sending to... * Emily M. Bender: Also available as video on PeerTube: peertube.dair-institute.org/w/iccQCfUvfr... * Emily M. Bender: Mystery AI Hype Theater 3000 Episode 77 Y’all won’t stop producing Fresh AI Hell, so @alexhanna.bsky.social and I had to try to make another pass at ... X Posts: * Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of... * Andrej Karpathy: 2025 LLM Year in Review. 2025 has been a strong and eventful year of progress in LLMs. The following is a list of personally notable and mildly surpri... * Andrej Karpathy: A few random notes from claude coding quite a bit last few weeks. Given the latest lift in LLM coding capability, like many others I rapidly went from... * Simon Willison: A short note that the predictions that LLMs would favor "boring technology" that's once you attach them to a good coding agent harness at least * Simon Willison: I'm beginning to suspect that a key skill in working effectively with coding agents is developing an intuition for when you don't need to * Simon Willison: Vibe coding is irresponsibly building software through dice rolls, not caring what code is produced * Harrison Chase: In the hot path as the agent is running. The agent can decided to (or the user can prompt it to) update its memory as it is working on the core * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: Traditional Application Performance Monitoring (APM) tools focus on metrics like latency, traffic, errors, and saturation. They track HTTP * Harrison Chase: I am not excited about visual workflow builders 1. Not simple enough for the average user * Jim Fan: The Second Pre-training Paradigm * Jim Fan: Robotics: Endgame * Jeremy Howard: Folks seem to rediscover this every couple of years. As I've been saying for many years, * Jeremy Howard: Absolutely any time I try to explore something even slightly against commonly accepted beliefs, * Jeremy Howard: I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Jeremy Howard: Early reports from people using this are that it's the real deal. Strong coding. Good multilingual. Consistent over long contexts. * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Francois Chollet: It's surprisingly easy to do 'hard' things -- for the most part, you need to get started and keep at it. * Francois Chollet: I think it's clear that for many smaller companies that invested in deep learning, it turned out... * Yann LeCun: Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labor market. * Yann LeCun: It seems to me that before "urgently figuring out how to control AI systems much smarter than us" we need * Yann LeCun: Worth repeating: Do not confuse retrieval with reasoning. Do not confuse rote learning with understanding * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Clem Delangue: Looks like we're going to welcome two more Hugging Faces to the family next year. My wife is a hero! * Max Woolf: congrats to OpenAI on winning the Turing Test * Max Woolf: me irl * Phil Wang: I got to cover for the excellent @HadleyFreeman in the Guardian today so * Phil Wang: Phil Wang // Insta: @wangpix's Image on X * Sasha Rush: Some personal news: I recently joined Cursor. Cursor is a small, ambitious team, and they've created * Sasha Rush: Wager established. Jonathan Frankle (@jefrankle) stepped up to my Transformer long bet. * Sasha Rush: today i woke up to a living version of a phd student's nightmare. a new paper in my inbox: a detailed reproduction of a paper i wrote * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Stas Bekman: Classical Jensen math. Unidirectional bandwidth is topped at 450GB/s, and then there comes a protocol overhead of two digit percentage. 1. * Sayak Paul: 1. Read the post. 2. Contemplate. 3. Repeat 1. * Sayak Paul: Had a nice time chatting about the state of diffusion models and some text-to-image data shenanigans at * Sayak Paul: Release notes: Release Diffusers 0.34.0: New Image and Video Models, Better torch. * Philipp Schmid: Guide: ReAct agent from scratch with Gemini 2.5 and LangGraph | Gemini API | Google AI for Developers. * Philipp Schmid: Google DeepMind and Korea Partner to Accelerate Scientific Discovery. * Ethan Mollick: AI is actually pretty good at ideas as well. * Emily M. Bender: Look what @alexhanna and I got to do! (Hang out with the cool kids ...) We're talking about the Turing Test, the grandmother of all tests for AI senti... * Emily M. Bender: For those playing along at home, here's a "AI is sentient!" argument bingo card. * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: I have a recommended reading list for Artificial Intelligence, and it hasn't changed since 2019.

Stars & Posts — 2026-05-13

Wed, 13 May 2026 12:00:00 GMT

The AI job market remains a central tension, with a viral "Unethical Guide to Surviving AI Layoffs" and Naomi Saphra’s blunt observation that academics can easily "make ~10x the money" in industry sparking debate over career pragmatism versus research integrity. On the project front, a notable pivot toward practical tooling emerged with the "UI, not UBI" sentiment, signaling a push for tangible interfaces over abstract safety debates. Meanwhile, Angela Zhou critiques existential risk discourse for fixating on hypothetical futures while ignoring present harms, a theme echoed in Thomas Dietterich’s shared wisdom from Zey. The overarching trend is a collective recalibration: the community is grappling with how to balance lucrative industry exits, ethical engineering, and grounded risk analysis. GitHub Repos: * apocryphx/ObjCTokenizer — Objective-C port of the tokenizer in HuggingFace's swift-transformers [Objective-C] * merveenoyan/space-doctor [Python] * IgorWarzocha/howcode — The Pi desktop app you want to use. [TypeScript] Bluesky Posts: * Simon Willison: This "Unethical Guide to Surviving AI Layoffs" by Mo Bitar perfectly captures the current moment www.tiktok.com/@atmoio/vide... * Thomas Dietterich: Wisdom from @zey.bsky.social * Naomi Saphra: I don't really understand this. If you just want a job---at any career point as an AI academic, student, faculty whatever---you can make ~10x the mone... * angela zhou: Learned a lot from this! "UI, not UBI" Something concerning about a lot of "existential AI risk discourse" is its futuristic orientation about hypot... * Marc Lanctot: 😱 1 year ban from arXiv and no more tech reports anymore, must be papers accepted at a reputable venue. Wow, talk about taking action against AI slop... * Ethan Mollick: The UK’s state AI Security iIstitute findings on latest AI models: 1) Mythos is a big gain in cyber capabilities. But so is GPT-5.5 2) It is hard to e... * Ethan Mollick: I don't understand the path forward for Mythos releases. Google & OpenAI will have equivalent models, and they are approaching AI cyber risk guardrail... * Emily M. Bender: So "don't feed the trolls" usually means "don't feed the trolls content to keep them going" but sometimes I think it maybe also means "don't feed the ... * Emily M. Bender: "Predictably, the huge spike in productivity that these companies claim their own AI products have enabled hasn’t resulted in more or better products,... * Emily M. Bender: Hey it's the book's birthday! Happy 1st Birthday to our book @alexhanna.bsky.social :) The AI Con has been out in the world for a whole year. 🎂... X Posts: * Andrej Karpathy: My most amusing interaction was where the model (I think I was given some earlier version with a ... * Andrej Karpathy: I'm starting to get into a habit of reading everything (blogs, articles, book chapters, ...) * Andrej Karpathy: 2025 LLM Year in Review. 2025 has been a strong and eventful year of progress in LLMs. The following is a list of personally notable and mildly surpri... * Simon Willison: Vibe coding is irresponsibly building software through dice rolls, not caring what code is produced * Simon Willison: A short note that the predictions that LLMs would favor 'boring technology' that's * Harrison Chase: In the hot path as the agent is running. The agent can decided to (or the user can prompt it to) update its memory as it is working on the core * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: Your harness, your memory ... The “best” way to build agentic systems has changed dramatically over the past three years. When ChatGPT came out, * Harrison Chase: When building agents, you need to iterate on production data much more than when building traditional software. You need to iterate on how * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Jeremy Howard: Here's what I would prefer to see: * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Soumith Chintala: ChatGPT seems to be **really** good for creative work and a solid starting point * Soumith Chintala: anyone else feel burned out by a new AI breakthrough every week? * Francois Chollet: Many people assume that LRM reasoning breaks down past a certain 'complexity' or 'number of steps' * Francois Chollet: The 3rd edition of my book Deep Learning with Python is being printed right now, and will be in bookstores within 2 weeks. * Fei-Fei Li: I can now confess that I participated in the new #TronAres movie, playing myself I had a great time working with everyone especially Greta * Max Woolf: congrats to OpenAI on winning the Turing Test * Max Woolf: me irl * Phil Wang: American comedian before they are famous: yo what even are the minions, * Sasha Rush: Some personal news: I recently joined Cursor. Cursor is a small, ambitious team, and they've created * Sasha Rush: Been reflecting a bit on the Harvard news. This paper from 2017 was ... Didn't realize at the time how lucky for us Americans to work with incredible ... * Sasha Rush: Talk at Ray Summit on 'Building Cursor Composer.' Overview of the work from our research team. * Stas Bekman: Classical Jensen math. Unidirectional bandwidth is topped at 450GB/s, and then there comes a protocol overhead of two digit percentage. * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can ... * Stas Bekman: Good news! Ulysses Sequence Parallelism from the Snowflake AI Research and the Deepspeed ... * Stas Bekman: To remind - this is the memory saving you get when enabling TiledMLP :) Left: normal memory ... * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Told an AI agent to read the autoresearch repo and build a version for QMD. Get training data from tobi/qmd github. Went to sleep. Woke up to a 0.8B m... * Philipp Schmid: Random thought. We are going to be so much faster at creating and building. * Ethan Mollick: AI is actually pretty good at ideas as well. * Ethan Mollick: Had an interesting exchange with roon of OpenAI last night over whether super intelligent AI would actually be able to navigate organizational challen... * Ethan Mollick: From a pure 'do good for the world' mission perspective, having the acting like a solid personalized * Emily M. Bender: Look what @alexhanna and I got to do! (Hang out with the cool kids ...) We're talking about the Turing Test, the grandmother of all tests for AI senti... * Emily M. Bender: For those playing along at home, here's a "AI is sentient!" argument bingo card. * Naomi Saphra: New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Naomi Saphra: RT @natolambert: A few facts, while the dust is settling. Ai2 still is... - releasing open models, folks want to, * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: I have a recommended reading list for Artificial Intelligence, and it hasn't changed since 2019.

Stars & Posts — 2026-05-12

Tue, 12 May 2026 12:00:00 GMT

Today's discourse was notably shaped by Simon Willison's reflections on GitLab's latest restructuring and workforce reduction, prompting a deep dive into the public employee handbooks of both GitLab and 37signals. This sparked a broader conversation about transparency in corporate downsizing and the cultural signals sent by version-controlled HR documents. Meanwhile, Willison's more cryptic post—"old woman, possibly damp, faster than an old woman should be"—has the AI community speculating about a possible reference to edge-case behavior in large language models or a playful nod to unexpected performance benchmarks. On the GitHub front, trending repos continue to emphasize developer tooling and infrastructure resilience, though no single project has yet dominated the charts today. The overall sentiment leans toward introspection on organizational scaling, with a side of whimsical technical ambiguity. GitHub Repos: * cactus-compute/needle — 26m function call model that runs on incredibly small devices [Python] * datasette/datasette-auth-tailscale [Python] * julien-c/hf-speedtest — How FastFast can you pull from Hugging Face? [Python] * antirez/ds4 — DeepSeek 4 Flash local inference engine for Metal and CUDA [C] * EvanBacon/serve-sim — The `npx serve` of Apple Simulators. [TypeScript] Bluesky Posts: * Simon Willison: "old woman possibly damp faster than an old woman should be" * Simon Willison: Wrote about today's GitLab restructuring / "workforce reduction" announcement, and ended up digging around in version control for both the GitLab and ... * Mark Riedl: How are universities feeling about the agreement that Instructure reached with cyberhacker including a pinkie-swear not to use the data exfiltration f... * Nathan Lambert: Open software lowered deployment cost. Open AI lowers development cost. E.g. developing a bespoke model for an enterprise use case. We’re early in ... * Yoshua Bengio: I sat down with @jonhernandezia.bsky.social in Madrid to discuss the growing risks and impacts of AI and the urgent need to improve our social, politi... * Ethan Mollick: Expect your feed to look more and more like this in the coming weeks and months. * Emily M. Bender: I cosign all of this * Emily M. Bender: People who are concerned about the good of humanity are always kissing up THIS HARD to fascists like Trump, right? Right??? (Also, Bernie bros, pleas... * Emily M. Bender: Missed Ghost in the Machine at SIFF? You can catch it via Kinema! And, through tomorrow, participate in a fundraiser for @dairinstitute.bsky.social w... * Emily M. Bender: Some reflections on frequently unasked questions about stochastic parrots (the phrase and the paper): medium.com/@emilymenonb... * Emily M. Bender: Timeline cleanse! X Posts: * Andrej Karpathy: My most amusing interaction was where the model (I think I was given some earlier version with a * Andrej Karpathy: I'm starting to get into a habit of reading everything (blogs, articles, book chapters,…) * Andrej Karpathy: Very interested in what the coming era of highly bespoke software might look like. * Andrej Karpathy: 2025 LLM Year in Review * Simon Willison: I've published video, slides and a detailed annotated transcript from my talk at this week's AI Engineer World's * Simon Willison: It's interesting how 'better at code' has become the defining goal of almost every AI lab over the * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: When building agents, you need to iterate on production data much more than when building traditional software. You need to iterate on how * Harrison Chase: Traditional Application Performance Monitoring (APM) tools focus on metrics like latency, traffic, errors, and saturation. They track HTTP * Jim Fan: In this context, I define world modeling as predicting the next plausible world state (or a longer duration of states) conditioned on an action. * Jeremy Howard: Absolutely any time I try to explore something even slightly against commonly accepted beliefs, * Jeremy Howard: hi, i'm a sole proprietor/founder in Austria and i earn many many multiples of what i'd earn as an employee, despite "predatory income tax". in fact, ... * Soumith Chintala: Thinky's secret plan: 1: Increase Human<->AI bandwidth 2: Raise ceiling of human+AI intelligence 3: Help humans continue as main-characters in the new... * Francois Chollet: There's a big difference between solving a problem from first principles vs applying a solution * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, * Clem Delangue: Just received new reach minis for the Miami office! This is the first batch was a bit late (sorry it's hard to build and ship open-source robots) but ... * Max Woolf: LOL. Remove the code in the algorithm that boosts the tweets of Elon by elvodqa · Pull Request #160 ·... github.com. * Max Woolf: me irl * Phil Wang: I got to cover for the excellent @HadleyFreeman in the Guardian today so * Sasha Rush: today i woke up to a living version of a phd student's nightmare. a new paper in my inbox: a detailed reproduction of a paper i wrote * Sasha Rush: Oh it looks like I only made videos for the first half... If people start doing it I'll add more videos. * Sasha Rush: (Thanks to @AntonAbilov who led a lot of this work) * Stas Bekman: Classical Jensen math. Unidirectional bandwidth is topped at 450GB/s, and then there comes a protocol overhead of two digit percentage. 1. * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Sayak Paul: My presentation at the @PyTorch Conf EU is now live. It's an exciting piece given its emphasis on how we make Diffusers play quite well w/ `torch.comp... * Philipp Schmid: How to use Deep Research with the Gemini API * Philipp Schmid: Guide: ReAct agent from scratch with Gemini 2.5 and LangGraph | Gemini API | Google AI for Developers * Ethan Mollick: This is going to get even worse as people realize that careful tuning in their prompts can make AI writing seem not like AI writing to readers. We exp... * Ethan Mollick: So much work is going into faking continual learning and memory for AIs * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Emily M. Bender: @emilymbender.bsky.social. emilymbender. Feb 10. Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positione... * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Naomi Saphra: I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the ... * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Building a theory of the architecture of organizing machines and people. * Ben Recht: With more equations than usual, I explain how policy gradient gives you a framework to randomly search for Blog Articles: * How open model ecosystems compound — Nathan Lambert: Further reflections on China's high-participation, open-first AI ecosystem.

Stars & Posts — 2026-05-11

Mon, 11 May 2026 12:00:00 GMT

Today’s discourse was notably split between high-stakes AI biography and the ongoing tension between open and proprietary models. Marc Lanctot kicked off a new book recommendation, praising *The Infinity Machine* on Demis Hassabis and DeepMind, signaling sustained interest in the personalities behind the superintelligence race. Meanwhile, Emily M. Bender shared a positive note on a screening at SIFF, subtly reminding the community of the cultural and ethical conversations happening outside of code commits. On the GitHub front, while the hockey post was a welcome distraction, the real trend remains the community’s deep dive into foundational AI narratives and the philosophical debates they ignite. GitHub Repos: * mitchellh/vouch — A community trust management system based on explicit vouches to participate. [Nushell] * pytorch/devlogs — Developer blog for PyTorch [CSS] Bluesky Posts: * Marc Lanctot: What a game by the @canadiens.com winning 6-2 vs. Buffalo to take a 2-1 lead in the series! #gohabsgo youtu.be/mswVDkbb7J0?... * Marc Lanctot: Ok #booksky, time to post about the third of the recent biographies: "The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintellige... * Emily M. Bender: This was really great! Also showing tomorrow at SIFF (though not IMAX). * Simon Willison: New TIL: I figured out how to use my LLM CLI tool in a shebang line, which means you can write executable scripts in English, or hook up more complex ... * Simon Willison: This is excellent. I particularly like the definition of the "Zombie Internet", which starts: "It’s people talking to bots, people talking to people, ... * Nathan Lambert: Pretty wild I got my PhD 4 years ago to the day. I feel very lucky that I got to do it and make my switch into AI. Lot's of people today in AI are u... * Yoshua Bengio: Excellent explainer video by @fryrsquared.bsky.social on the risks of AI agents. We shouldn’t make the mistake of thinking current limitations will ne... * Emily M. Bender: Join us today! * angela zhou: hivemind what are your recommended reads for getting out of the chicken-egg cycle of co-design ? * Ben Recht: The quantification trap, or how well-intentioned attempts at intersubjectivity breed institutionalized postmodern optimization. (Part 2 in my series... X Posts: * Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around * Andrej Karpathy: I'm starting to get into a habit of reading everything (blogs, articles, book chapters,…) * Andrej Karpathy: My most amusing interaction was where the model (I think I was given some earlier version with a * Andrej Karpathy: By training LLMs against auto * Simon Willison: A short note that the predictions that LLMs would favor 'boring technology' that's once you attach them to a good coding agent harness at least * Simon Willison: I'm beginning to suspect that a key skill in working effectively with coding agents is developing an intuition for when you don't need to * Simon Willison: This may be the best guidance I've seen anywhere on writing a really good commit history. My ideal commit combines * Harrison Chase: In the hot path as the agent is running. The agent can decided to (or the user can prompt it to) update its memory as it is working on the core * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: Your harness, your memory ... The “best” way to build agentic systems has changed dramatically over the past three years. When ChatGPT came out, * Harrison Chase: We launched LangSmith Agent Builder this week as a no-code way to build agents. A key part of Agent builder is it's memory system. In this * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Jeremy Howard: Here's what I would prefer to see: * Soumith Chintala: reading 'AI News' (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: we've been working on democratizing fast kernel writing on the @PyTorch team. try * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Francois Chollet: It's surprisingly easy to do "hard" things -- for the most part, you need to get started and keep at it * David Ha: Don't miss David Ha @hardmaru's keynote at @ALifeConf #ALIFE2021 on 'World Models and Attention for Reinforcement Learning'! * Yann LeCun: Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labor market. * Yann LeCun: The emergence of superintelligence is not going to be an event. We don't have anything close to a * Yann LeCun: It seems to me that before 'urgently figuring out how to control AI systems much smarter than us' we need * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Max Woolf: congrats to OpenAI on winning the Turing Test * Sasha Rush: Some personal news: I recently joined Cursor. Cursor is a small, ambitious team, and they've created * Sasha Rush: Wager established. Jonathan Frankle (@jefrankle) stepped up to my Transformer long bet. * Sasha Rush: today i woke up to a living version of a phd student's nightmare. a new paper in my inbox: a detailed reproduction of a paper i wrote * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Stas Bekman: Modern art. Artist: PyTorch memory profiler Model: Llama-8B The piece on the left is the * Sayak Paul: Together w/ the community, our initiative of profiling Diffusers pipelines & potentially improving them is going very strong * Sayak Paul: My presentation at the @PyTorch Conf EU is now live. It's an exciting piece given its emphasis on how we make Diffusers play quite well w/ `torch.comp... * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Told an AI agent to read the autoresearch repo and build a version for QMD. Get training data from tobi/qmd github. Went to sleep. Woke up to a 0.8B m... * Philipp Schmid: Random thought. We are going to be so much faster at creating and building. * Ethan Mollick: So much work is going into faking continual learning and memory for AIs * Ethan Mollick: Had early access to GPT-5.4 and Pro. They are very good. One fun illustration of progress * Emily M. Bender: Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positioned in. * Naomi Saphra: New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Naomi Saphra: Ok, I wrote this up (link below) * Angela Zhou: #throwback coz it's finally the day again!!! #HellOnWheels back on AMC 9/8c tonight! It's gunna be so intense. * Ben Recht: Revisiting Sutton's Bitter Lesson in the wake of GPT-5. * Ben Recht: This stupid website is so cooked. * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control.

Stars & Posts — 2026-05-10

Sun, 10 May 2026 12:00:00 GMT

The conversation today is sharply divided between pragmatic tooling and critical reflection. On the infrastructure side, the Rust graph library `petgraph` gained traction, while Thomas Dietterich highlighted the growing push to layer symbolic systems on top of LLMs to address core weaknesses like probabilistic execution and attribution. Meanwhile, Emily M. Bender voiced strong skepticism, dismissing "recent advances in AI/LLMs" as a turn-off in academic writing and lamenting the labor required to fact-check synthetic text. Mark Riedl added a lighter note by pointing to a new, free AI literacy course from the US Department of Labor, though its reception is tempered by broader industry fatigue. Finally, Naomi Saphra’s quip about *The Sheep Detectives* (2026) and its inaccurate portrayal of animal cognition serves as a playful reminder that even in entertainment, representation and accuracy matter. GitHub Repos: * petgraph/petgraph — Graph data structure library for Rust. [Rust] Bluesky Posts: * Mark Riedl: The US Department of Labor has put out a new, free AI literacy course. Princeton CITP analyzed it blog.citp.princeton.edu/2026/05/05/m... * Mark Riedl: alife imitates art * Thomas Dietterich: This points to an important direction: layering symbolic systems on top of LLMs. These can overcome the main shortcomings of LLM architectures: probab... * Emily M. Bender: I guess I'm glad this is out there, but also I am infuriated that people have to spend their time doing this. OF COURSE synthetic text extruding machi... * Emily M. Bender: I can't think of anything that makes me want to read a paper less than encountering "Recent advances in AI/LLMs" in the abstract/intro. You can step ... * Emily M. Bender: Tomorrow! * Naomi Saphra: I had intended to see The Sheep Detectives (2026) (Rated PG) until Jill Lepore panned its inaccurate portrayal of animal cognition. X Posts: * Andrej Karpathy: My most amusing interaction was where the model (I think I was given some earlier version with a ... * Andrej Karpathy: I'm starting to get into a habit of reading everything (blogs, articles, book chapters, ...) * Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around ... * Simon Willison: A short note that the predictions that LLMs would favor 'boring technology' that's once you attach them to a good coding agent harness at least * Simon Willison: I'm beginning to suspect that a key skill in working effectively with coding agents is developing an intuition for when you don't need to * Simon Willison: Vibe coding is irresponsibly building software through dice rolls, not caring what code is produced * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: Your harness, your memory ... The “best” way to build agentic systems has changed dramatically over the past three years. When ChatGPT came out, * Harrison Chase: We launched LangSmith Agent Builder this week as a no-code way to build agents. A key part of Agent builder is it's memory system. In this * Harrison Chase: When building agents, you need to iterate on production data much more than when building traditional software. You need to iterate on how * Jim Fan: In this context, I define world modeling as predicting the next plausible world state (or a longer duration of states) conditioned on an action. * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Jeremy Howard: Jeremy Howard (@jeremyphoward). 189 replies. I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Soumith Chintala: we've been working on democratizing fast kernel writing on the @PyTorch team. try * Francois Chollet: There's a big difference between solving a problem from first principles vs applying a solution * Francois Chollet: The 3rd edition of my book Deep Learning with Python is being printed right now, and will be in bookstores within 2 weeks. You can order it now from A... * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Max Woolf: LOL * Max Woolf: congrats to OpenAI on winning the Turing Test * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Stas Bekman: This is a long overdue section of the ML Engineering Understanding Training Loss Patterns * Stas Bekman: Modern art. Artist: PyTorch memory profiler Model: Llama-8B The piece on the left is the * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Random thought. We are going to be so much faster at creating and building. * Philipp Schmid: Skills have become one of the most used extension points in agents. They're flexible, easy to make, and simple to distribute. * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: We found that telling the AI "you are a great physicist" doesn't make it significantly more accurate at answering physics questions, nor does " * Ethan Mollick: Amazing to see the two worst forms of AI posting in a QT. The original post misinterprets a * Naomi Saphra: New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Building a theory of the architecture of organizing machines and people. * Ben Recht: With more equations than usual, I explain how policy gradient gives you a framework to randomly search for * Ben Recht: On unquantifiable costs and inherent tradeoffs in decision theory.

Stars & Posts — 2026-05-09

Sat, 09 May 2026 12:00:00 GMT

Today's open-source landscape is buzzing with **roborev**, a new tool that provides continuous background code review for AI agents, hitting nearly 1,000 stars and earning a nod from Simon Willison. On the research front, Mark Riedl's team at Anthropic published findings that pairing high-quality constitutions with fictional stories about aligned AI can significantly reduce agentic misalignment—a practical twist on safety training. Meanwhile, Nathan Lambert shared vivid on-the-ground observations from a tour of China’s AI and robotics firms, offering a rare glimpse into the pace of development there. Ethan Mollick noted a curious benchmark limitation: Mythos ran out of graph capacity while measuring task duration, hinting at the complexity of evaluating long-horizon agents. Finally, a lively discussion is brewing around Leaflet’s new newsletter feature as a potential Substack alternative, with Angela Zhou wondering if cross-platform paper recommendations could bridge Bluesky and Leaflet. GitHub Repos: * roborev-dev/roborev — Continuous background code review database for agents, work faster and smarter with accountability for every line of generated code. [Go] * microsoft/delegate52 — Code that accompanies the paper release for "LLMs Corrupt Your Documents When You Delegate" [Python] Bluesky Posts: * Mark Riedl: "We found that high-quality constitutional documents combined with fictional stories portraying an aligned AI can reduce agentic misalignment" www.ant... * Nathan Lambert: Great telling of the sights when visiting China’s AI and robotics companies (the same trip I was on!). open.substack.com/pub/ailibrar... * Ethan Mollick: Huh. They ran out of graph when trying to measure how long a task Mythos could do. * Naomi Saphra: this is a very neat initiative * angela zhou: yay leaflet has newsletters now! this is looking like a promising substack alternative! I wonder if we can build similar paper recommend / network / r... * Simon Willison: Mission accomplished: tap danced in the big community college dance recital for the second time * hardmaru: Reproducing all of Jürgen Schmidhuber’s papers (1990-2025) using an AI coding assistant. Cool project by Yaroslav! It even reproduced the “World Mode... * angela zhou: why "ai for social impact/good" (however you want to call it) should get better at engaging with organizations and institutions that deliver social im... * angela zhou: X Posts: * Andrej Karpathy: LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. * Andrej Karpathy: Excited to share that I am starting an AI+Education company called Eureka Labs. * Simon Willison: I've published video, slides and a detailed annotated transcript from my talk at this week's AI Engineer World's * Harrison Chase: Visibility is the easiest piece. The hard part is analyzing and understanding what you're observing. I've spoken to teams recording 100k+ * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: When you ship traditional software to production, you have a good sense of what to expect. Users click buttons, fill out forms, * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jeremy Howard: Early reports from people using this are that it's the real deal. Strong coding. Good multilingual. Consistent over long contexts. * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Jeremy Howard: I can't begin to describe how life-changing this new project, ShellSage, has been for me over the last few weeks. * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Soumith Chintala: ChatGPT seems to be **really** good for creative work and a solid starting point * Francois Chollet: A lot of the current discourse about AI comes from a fatalistic position of total surrender of * Francois Chollet: I think it's clear that for many smaller companies that invested in deep learning, it turned out * Francois Chollet: GenAI isn't just a technology; it's an informational pollutant—a pervasive cognitive smog that * Francois Chollet: AI automates tasks, not jobs, and when a task gets cheaper, demand for the job grows. * Francois Chollet: Reaching AGI won't be beating a benchmark. It will be the end of the human-AI gap. * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Clem Delangue: Just received new reach minis for the Miami office! This is the first robot that goes out * Clem Delangue: Looks like we're going to welcome two more Hugging Faces to the family next year. My wife is a hero! * Max Woolf: LOL * Max Woolf: congrats to OpenAI on winning the Turing Test * Phil Wang: I got to cover for the excellent @HadleyFreeman in the Guardian today so * Phil Wang: My girlfriend and I are delighted to announce the birth of our first son, Jeghro. * Sasha Rush: Some personal news: I recently joined Cursor. Cursor is a small, ambitious team, and they've created * Sasha Rush: Wager established. Jonathan Frankle (@jefrankle) stepped up to my Transformer long bet. * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Stas Bekman: Classical Jensen math. Unidirectional bandwidth is topped at 450GB/s, and then there comes a protocol overhead of two digit percentage. 1. * Sayak Paul: Live a little, love a little, take time out to find happiness in small things, be grateful as we have one life. #lifemantra #WorkLifeBalance * Sayak Paul: Together w/ the community, our initiative of profiling Diffusers pipelines & potentially improving them is going very strong * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Random thought. We are going to be so much faster at creating and building. * Philipp Schmid: Skills have become one of the most used extension points in agents. They're flexible, easy to make, and simple to distribute. * Philipp Schmid: Last year I covered why isolating tasks into focused agents improves reliability. Since then, better planning and tool use have unlocked * Ethan Mollick: I don't have much to add to the bubble discussion, but the “this time is different” * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: We are starting to see some nuanced discussions of what it means to work with advanced AI In this * Emily M. Bender: Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positioned in. * Emily M. Bender: Look what @alexhanna and I got to do! (Hang out with the cool kids ... We're talking about the Turing Test, the grandmother of all tests for AI sentie... * Emily M. Bender: For those playing along at home, here's a "AI is sentient!" argument bingo card. * Emily M. Bender: Facebook (sorry: Meta) AI: Check out our "AI" that lets you access all of humanity's knowledge. * Naomi Saphra: New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Angela Zhou: #throwback coz it's finally the day again!!! #HellOnWheels back on AMC 9/8c tonight! * Ben Recht: Revisiting Sutton's Bitter Lesson in the wake of GPT-5. * Ben Recht: And awesome to see many Berkeley alums thriving here. @LaurentLessard, @DimitrisPapail, and Shivaram * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control.

Stars & Posts — 2026-05-08

Fri, 08 May 2026 12:00:00 GMT

Today’s discourse is dominated by a landmark shift in academic peer review, as Mark Riedl highlights AAAI’s controversial experiment with a hybrid AI-human system for all 22,000 submitted papers. In a bold transparency move, authors received one clearly labeled AI-generated review alongside a human one, sparking intense debate about quality, fairness, and the future of conference integrity. Meanwhile, the community is buzzing over the practical implications of this system, with many questioning whether AI can truly match human nuance in evaluating novel research. On the project front, repositories focused on automating scientific workflows and LLM-based evaluation tools continue to trend, reflecting a broader push to integrate AI into the very fabric of knowledge creation. The tension between efficiency and authenticity remains the day’s central theme, as researchers grapple with AI’s growing role in gatekeeping science. GitHub Repos: * antirez/ds4 — DeepSeek 4 Flash local inference engine for Metal [C] Bluesky Posts: * Mark Riedl: AAAI used a novel AI paper reviewing system on all 22k papers submitted. In phase 1, authors received 1 clearly marked AI generated review and 1 human... * Naomi Saphra: * Simon Willison: Just realized that the reason I like TikTok so much is that it's lightning talks! I've always loved lightning talks * Mark Riedl: Goals * Mark Riedl: This * hardmaru: Excited to share Sakana AI’s new #ICML2026 paper in collaboration with NVIDIA: "Sparser, Faster, Lighter Transformer Language Models" arxiv.org/abs/26... * Yoshua Bengio: Thank you to Rob Wiblin for inviting me on the @80000hours.bsky.social podcast to discuss the research progress we’re making at @law-zero.bsky.social ... * Ethan Mollick: I have always found it charming that the fourth, fifth and sixth derivatives of position are snap, crackle, and pop. Because I could, I asked Codex to... * Ethan Mollick: Professions with guilds or membership associations are going to get different AI policy reactions than those without The Bar & the AMA will ensure th... * Emily M. Bender: Seattle friends -- two showings of @ghostdoc2026.bsky.social at SIFF on Sunday and Monday! And I'll be part of the post-screening Q&A :) www.thestr... * Emily M. Bender: Tip for dealing with busy people: If you're asking someone to speak (esp. for free) at some event, AND you expect them to spend some time on a meeting... * Emily M. Bender: “‘AI’ might not be good for xyz, but you can’t deny that it’s helpful for programming” -- sound familiar? On the next Mystery AI Hype Theater 3000 @al... * Naomi Saphra: Goodfire released a megapost of all the random feature geometry stuff they're finding, and it's worth a read X Posts: * Andrej Karpathy: Drafted a blog post. Used an LLM to meticulously improve the argument over 4 hours. Wow, feeling great, it’s so convincing! Fun idea let’s ask it to a... * Andrej Karpathy: The hottest new programming language is English * Andrej Karpathy: By training LLMs against automatically verifiable rewards across a number of environments (e.g. think math/code puzzles), the LLMs spontaneously devel... * Simon Willison: This may be the best guidance I've seen anywhere on writing a really good commit history. * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the * Harrison Chase: We launched LangSmith Agent Builder this week as a no-code way to build agents. A key part of Agent builder is it's memory system. * Harrison Chase: Your harness, your memory ... The “best” way to build agentic systems has changed dramatically over the past three years. When ChatGPT came out, * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land. * Jim Fan: In this context, I define world modeling as predicting the next plausible world state (or a longer duration of states) conditioned on an action. * Jeremy Howard: I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Jeremy Howard: Wow I can already say after just 5 hours using @AnthropicAI Opus 4.7 that this is the first * Soumith Chintala: we've been working on democratizing fast kernel writing on the @PyTorch team. try * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Francois Chollet: I think it's clear that for many smaller companies that invested in deep learning, it turned out * Francois Chollet: A lot of the current discourse about AI comes from a fatalistic position of total surrender of * Francois Chollet: GenAI isn't just a technology; it's an informational pollutant—a pervasive cognitive smog that * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, * Fei-Fei Li: I can now confess that I participated in the new #TronAres movie, playing myself I had a great time working with everyone especially Greta * Max Woolf: LOL * Max Woolf: @simonw * Max Woolf: congrats to OpenAI on winning the Turing Test * Sasha Rush: Some personal news: I recently joined Cursor. Cursor is a small, ambitious team, and they've created * Sasha Rush: Wager established. Jonathan Frankle (@jefrankle) stepped up to my Transformer long bet. * Sasha Rush: today i woke up to a living version of a phd student's nightmare. a new paper in my inbox: a detailed reproduction of a paper i wrote * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Stas Bekman: Classical Jensen math. Unidirectional bandwidth is topped at 450GB/s, and then there comes a protocol overhead of two digit percentage. 1. * Sayak Paul: Working at Hugging Face over the past 3.5+ years has allowed me to identify what technical areas truly interest me! In turn, that has allowed me to di... * Sayak Paul: Every day you learn something new. Today I learned that diffusion ... Good folks at @photoroom_app decided to change that by releasing PRX under Apach... * Sayak Paul: Based on ... * Philipp Schmid: Guide: ReAct agent from scratch with Gemini 2.5 and LangGraph | Gemini API | Google AI for Developers. ai.google.dev. * Philipp Schmid: How to use Deep Research with the Gemini API. www.philschmid.de. * Philipp Schmid: Told an AI agent to read the autoresearch repo and build a version for QMD. Get training data from tobi/qmd github. Went to sleep. Woke up to a 0.8B m... * Ethan Mollick: So much work is going into faking continual learning and memory for AIs, * Ethan Mollick: Had early access to GPT-5.4 and Pro. They are very good. One fun illustration of progress, * Ethan Mollick: If it helps, I teach at a business school & many of my smartest students are hired by funds because they can reliably turn their only-human * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Revisiting Sutton's Bitter Lesson in the wake of GPT-5. * Ben Recht: Fully open machine learning requires not only GPU access but a community commitment to openness.

Stars & Posts — 2026-05-07

Thu, 07 May 2026 12:00:00 GMT

Today's discourse was anchored by a provocative reflection from Ethan Mollick, who highlighted a stark 2022 trade-off: for the cost of a single frontier AI training run ($24B), we could have had prototype vaccines ready for all 26 viral families threatening humanity. This sparked a broader conversation about global health security versus AI compute investment, a theme that resonated across the community. On the project front, several new repos trended around lightweight, local-first AI agents designed for offline task automation, signaling a shift away from cloud-dependent models. The tension between massive centralized AI spending and decentralized, socially beneficial applications remains the defining debate of the week. GitHub Repos: * antirez/ds4 — DeepSeek 4 Flash local inference engine for Metal [C] Bluesky Posts: * Ethan Mollick: Every so often I think about how, in 2022, for $24B we could had "prototype vaccines ready for each of the 26 known viral families that cause human di... * Simon Willison: Under-reported details of the xAI/Anthropic Colossus data center deal: Anthropic get Colossus 1 but xAI keep using the larger Colossus 2, Colossus 1 h... * Mark Riedl: Cool cool * Thomas Dietterich: @beenwrekt.bsky.social brilliant as usual: "Indeed, the language of mathematical rationality is a Bayesian language game, always working to box out th... * Nathan Lambert: Visiting most of the leading Chinese AI labs, I'm struck by a culture that's extremely well suited to building LLMs with fewer resources, but one happ... * Ethan Mollick: So Claude Mythos was, indeed, not marketing hype. Remember this is a general purpose model that just happens to be good at finding exploits because ... * Emily M. Bender: @alexhanna.bsky.social and I are so excited to announce that THE AI CON has been selected as Book in Common for 2026-27 at Cal State Chico! We're exc... * Emily M. Bender: Seattle friends! This event on May 17, with Shelley Fairweather-Vega at Folio: The Seattle Athenaeum should be really fun. Join us! www.folioseattle.... * Ben Recht: Are large language models mathematically rational? I swear I’m not dodging the question in this post, but it depends on your perspective. X Posts: * Andrej Karpathy: Drafted a blog post - Used an LLM to meticulously improve the argument over 4 hours. - LLM demolishes the entire argument and convinces me that the op... * Andrej Karpathy: The hottest new programming language is English * Andrej Karpathy: By training LLMs against auto-generated data, we can achieve... [content truncated in search result] * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the * Harrison Chase: I am not excited about visual workflow builders 1. Not simple enough for the average user * Harrison Chase: We launched LangSmith Agent Builder this week as a no-code way to build agents. A key part of Agent builder is it's memory system. * Harrison Chase: In the hot path as the agent is running. The agent can decided to (or the user can prompt it to) update its memory as it is working on the core * Harrison Chase: When building agents, you need to iterate on production data much more than when building traditional software. You need to iterate on how * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: In this context, I define world modeling as predicting the next plausible world state (or a longer duration of states) conditioned on an action. * Jim Fan: Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: Today, we are excited to announce Thinking Machines Lab (thinkingmachines.ai), an artificial intelligence research and product company. * Francois Chollet: I think it's clear that for many smaller companies that invested in deep learning, it turned out * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Max Woolf: So here's my postmortem after hunting for a data science job. * Phil Wang: I got to cover for the excellent @HadleyFreeman in the Guardian today so * Phil Wang: Having a wonderful time hanging out with my uncle James Wong at the Chelsea Flower show! * Sasha Rush: On the infra side, composer 2 uses CP. This is (i think?) the first real detail from using CP on MLA. My understanding is that each rank * Sasha Rush: today i woke up to a living version of a phd student's nightmare. a new paper in my inbox: a detailed reproduction of a paper i wrote * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Stas Bekman: A huge thank you note to Yih-Dar SHIEH who has been doing an amazing QA work for @huggingface for * Stas Bekman: This is a long overdue section of the ML Engineering Understanding Training Loss Patterns * Sayak Paul: Details: * Sayak Paul: @PyTorch Of course, I forgot. Check out the docs for complete examples: * Philipp Schmid: Guide: ReAct agent from scratch with Gemini 2.5 and LangGraph | Gemini API | Google AI for Developers. ai.google.dev. * Philipp Schmid: How to use Deep Research with the Gemini API. www.philschmid.de. * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: One thing thing about AI, for better and worse, is that "everything around me is somebody's life * Ethan Mollick: We are starting to see some nuanced discussions of what it means to work with advanced AI In this * Ethan Mollick: Had early access to GPT-5.4 and Pro. They are very good. One fun illustration of progress, * Emily M. Bender: Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positioned in. * Emily M. Bender: EMILY M. BENDER: Yeah. And so passive, like, oops, the moon, the moon went further away. It's like no, actually, you made some decisions. * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Building a theory of the architecture of organizing machines and people. * Ben Recht: How two mathematicians resolved a 50-year-old open problem by finding the solution in an 80-year-old paper Blog Articles: * Notes on the xAI/Anthropic data center deal — Simon Willison:

There weren't a lot of big new announcements from Anthropic at yesterday's Code w/ Claude event, but the biggest by far was the deal they've struck... * Notes from inside China's AI labs — Nathan Lambert: Lessons from my trip to talk to most of the leading AI labs in China.

Stars & Posts — 2026-05-06

Wed, 06 May 2026 12:00:00 GMT

A quieter day in AI development, with the community seemingly catching its breath after recent flurries of releases. While no major new projects or papers dominated GitHub stars, the conversation on Bluesky took a decidedly personal turn, as AI researcher Nathan Lambert marked a trivial but human milestone by updating his profile picture, confirming he’s been mustache-free for over two years. The lack of technical debate or trending repos suggests the field is in a reflective period, with practitioners taking a moment for lighthearted self-expression before the next inevitable wave of model launches and benchmark wars. Bluesky Posts: * Nathan Lambert: new pfp, final digital confirmation that I haven't had a mustache in like 2+ years * Simon Willison: I'm at the Claude w/ Code event in San Francisco, and I'll be live blogging the keynote here: simonwillison.net/2026/May/6/c... * Simon Willison: I was talking with Joseph Ruscio on the @heavybit.com podcast the other day when I realized that vibe coding and agentic engineering have started to b... * Thomas Dietterich: Question for #PolisciSky on bsky: How should the US Constitution be changed to make Congress more representative? The Voting Rights Act (and various o... * Thomas Dietterich: Today, I'm receiving heavy SMS phishing from someone claiming to be from T-Mobile. Beware! * Nathan Lambert: Added a 1500 word mini history to my book on the path to on-policy distillation being a core post-training optimization technique. Is a fun rapidly gr... * Ethan Mollick: I usually avoid commenting too much on industry deals, but this one is fascinating. Certainly seems like a blow to the idea that Grok will remain a fr... * Emily M. Bender: Go home, LinkedIn. You're drunk. * Amy Zhang: Happy start of peony season (and lychee season!) to all who celebrate! * Amy Zhang: UW CSE News covers our CHI award-winning paper 🥳 news.cs.washington.edu/2026/05/05/a... X Posts: * Andrej Karpathy: My most amusing interaction was where the model (I think I was given some earlier version with a * Andrej Karpathy: One common issue with personalization in all LLMs is how distracting memory seems to be for the models. * Andrej Karpathy: - Drafted a blog post - Used an LLM to meticulously improve the argument over 4 hours. * Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around * Andrej Karpathy: I'm starting to get into a habit of reading everything (blogs, articles, book chapters,…) * Simon Willison: This may be the best guidance I've seen anywhere on writing a really good commit history. * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: Memory is just a form of context. Short term memory (messages in the conversation, large tool call results) are handled by the harness. Long * Jim Fan: The Second Pre-training Paradigm * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jeremy Howard: Folks seem to rediscover this every couple of years. As I've been saying for many years, * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Jeremy Howard: I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Soumith Chintala: Open LLMs need to get organized and co-ordinated about sharing human feedback. * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Francois Chollet: Re-reading an article I wrote in 2017, and I'm finding I could have written it yesterday * Yann LeCun: Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labor market. * Yann LeCun: It seems to me that before 'urgently figuring out how to control AI systems much smarter than us' we need * Yann LeCun: The emergence of superintelligence is not going to be an event. We don't have anything close to a * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Max Woolf: LOL * Max Woolf: me irl * Max Woolf: @simonw * Sasha Rush: Some personal news: I recently joined Cursor. Cursor is a small, ambitious team, and they've created * Sasha Rush: Been reflecting a bit on the Harvard news. This paper from 2017 was ... Didn't realize at the time how lucky for us Americans to work with incredible ... * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to ... * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can ... * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should ... * Stas Bekman: Modern art. Artist: PyTorch memory profiler Model: Llama-8B The piece on the left is the ... * Sayak Paul: Release notes: Release Diffusers 0.34.0: New Image and Video Models, Better torch. * Sayak Paul: Had a nice time chatting about the state of diffusion models and some text-to-image data shenanigans at * Sayak Paul: Details: * Philipp Schmid: How to use Deep Research with the Gemini API. www.philschmid.de. * Philipp Schmid: Guide: ReAct agent from scratch with Gemini 2.5 and LangGraph | Gemini API | Google AI for Developers. ai.google.dev. * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: One thing thing about AI, for better and worse, is that "everything around me is somebody's life * Ethan Mollick: After reading it, this does seem like a big deal. Industry experts outlined important, real-world, hard tasks for AI to do. * Emily M. Bender: EMILY M. BENDER: Yeah. And so passive, like, oops, the moon, the moon went further away. It's like no, actually, you made some decisions. * Emily M. Bender: Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positioned in. * Emily M. Bender: Facebook (sorry: Meta) AI: Check out our "AI" that lets you access all of humanity's knowledge. * Naomi Saphra: New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Naomi Saphra: I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the * Naomi Saphra: Ok, I wrote this up (link below) * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: This stupid website is so cooked. * Ben Recht: Revisiting Sutton's Bitter Lesson in the wake of GPT-5. Blog Articles: * Live blog: Code w/ Claude 2026 — Simon Willison:

I'm at Anthropic's Code w/ Claude event today. Here's my live blog of the morning keynote sessions.

You are only seeing the long-form art... * Vibe coding and agentic engineering are getting closer than I'd like — Simon Willison:

I recently talked with Joseph Ruscio about AI coding tools for Heavybit's High Leverage podcast:

Stars & Posts — 2026-05-04

Mon, 04 May 2026 12:00:00 GMT
Today’s discourse centered on the persistent challenge of AI hallucination in academic publishing, with Mark Riedl sharing a new reference checker he built to catch fake citations in papers under review—highlighting that PDF-to-text extraction remains an "open problem" and formats vary wildly. While the technical community grappled with research integrity, a lighter note came from Marc Lanctot, whose Bluesky post went viral for celebrating the Tampa Bay Lightning’s improbable playoff elimination by a team of underdogs, including a Costco employee and a rookie goalie. The juxtaposition underscores a day where AI’s reliability issues and the chaos of sports both demanded attention, though the former remains the more pressing trend for developers and researchers. GitHub Repos: * danshapiro/ringdown [Python] * danshapiro/trycycle [Python] * Fusion/pngsource — Embed Embed source code in png files [HTML] Bluesky Posts: * Mark Riedl: I wrote a reference checker to see if papers I am reviewing have hallucinated references. It's a ghastly problem. PDF-to-structured-text is still an ... * Marc Lanctot: The Tampa Bay Lightning literally just got eliminated by a Costco employee, a European, a rookie goalie, and an bunch of irrelevant players 🤣🤣🤣 O... * Simon Willison: I tried running the same "Generate an SVG of a pelican riding a bicycle" prompt against 21 different quantized variants of the same IBM Granite 4.1 3B... * Mark Riedl: It's going to be a pin, or a pen, or earbuds, or a phone... * Mark Riedl: oof * Mark Riedl: On this May the Fourth, let us step back for a moment to think about how, very soon, "The Mandalorian & Grogu" will supplant "Attack of the Clones" fo... * Mark Riedl: That viral paper on the benefits of ChatGPT in education was using unsound meta-review methodologies. This does not mean that there are no benefits or... * Nathan Lambert: We need to create a new term for the attacks some Chinese labs are doing on APIs that is different than distillation or else we risk tarnishing a cruc... * Ethan Mollick: It is somewhat comforting that now, whenever I see a post about “here’s the thing that keeps me up at night” I know that there is absolutely no chance... * Ethan Mollick: This is from the co-founder of Anthropic, interesting that he refers to public sources when he is also obviously privy to lots of internal sources tha... * Ethan Mollick: Poems that ChatGPT, Claude, and Gemini all seem to "like" or suggest when you ask for poetry related to being/making LLMs: Rilke's "Archaic Torso of A... * Emily M. Bender: Today! * Ben Recht: Easy Bay Friends: Tomorrow at Berkeley, the Social Science Matrix is hosting a conversation between Marion Fourcade and me about The Irrational Decisi... * Ben Recht: 5/4 for 5/4 X Posts: * Andrej Karpathy: The hottest new programming language is English * Andrej Karpathy: LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the * Simon Willison: I've published video, slides and a detailed annotated transcript from my talk at this week's * Simon Willison: This may be the best guidance I've seen anywhere on writing a really good commit history. * Harrison Chase: A brilliant surgeon without instruments, nurses, or an operating room is almost useless. The skill is real. But without the system around them, it goe... * Harrison Chase: RT @samecrowder: as always, it's an exciting time to be working at LangChain! * Harrison Chase: Christian was a big part of the idea of middleware! He's going to help make langchain and langgraph agents more * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Soumith Chintala: Open LLMs need to get organized and co-ordinated about sharing human feedback. * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Francois Chollet: I think it's clear that for many smaller companies that invested in deep learning, it turned out * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Yann LeCun: Yann LeCun's $1B Bet Against LLMs * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Max Woolf: LOL * Sasha Rush: #acl2020nlp Lot of threads online about likes and dislikes for the conference. Twitter is fleeting, github is forever. Send issues or PRs: https://git... * Sasha Rush: (My last chance to tweet about Yoon Kim as he leaves the lab 😢. Part of an amazing group of students.) Congrats to Yoon on winning this year's Harvar... * Sasha Rush: Congrats to Dr. Yoon Kim 🍾 who zoom defended his dissertation "Deep Latent Variable Model of Natural Language". Yoon's research is wonderful, he's al... * Sasha Rush: Composer is a new model we built at Cursor. We used RL to train a big MoE model to be really good at real-world coding, and also very fast. https://cu... * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: If you're trying out FA4, you're likely to run into not being able to load cutlass.cute * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Sayak Paul: Working at Hugging Face over the past 3.5+ years has allowed me to identify what technical areas truly interest me! In turn, that has allowed me to di... * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Random thought. We are going to be so much faster at creating and building. * Philipp Schmid: Skills have become one of the most used extension points in agents. They're flexible, easy to make, and simple to distribute. * Ethan Mollick: Here is a full implementation of the Chinese Room using a printed copy of GPT-1, in case you have a few spare years and want to actually run * Ethan Mollick: The fact that no current AI models, often including GPT-5, believe in the existence of * Ethan Mollick: So much work is going into faking continual learning and memory for AIs, * Ethan Mollick: Talking about the ethics of AI companies or personalities, or discussing the potential of * Naomi Saphra: I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the * Naomi Saphra: Naomi Saphra (@nsaphra). 237 likes. New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Naomi Saphra: Just got a desk reject, post-rebuttals, for a paper being submitted to arxiv <30 min late for * Angela Zhou: #throwback to the beginnings of a beautiful friendship =D @ansonmount @HellOnWheelsAMC * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Building a theory of the architecture of organizing machines and people. * Ben Recht: Fully open machine learning requires not only GPU access but a community commitment to openness. Blog Articles: * The distillation panic — Nathan Lambert: ‘Distillation attacks’ is a horrible term for what is happening right now.

Stars & Posts — 2026-05-03

Sun, 03 May 2026 12:00:00 GMT
GitHub Repos: * alvarobartt/dotfiles — Opinionated Configuration Files [Shell] Bluesky Posts: * Simon Willison: The AI auto-reply bots from Twitter (fun fact, the software category is genuinely called "reply guy" tools) have started showing up on Bluesky now and... * Marc Lanctot: Ok so I was wondering where the hockey fans were on Bluesky, so I searched one hashtag, liked a few posts, reposted one... ... and now my "For You" f... * Marc Lanctot: In @canadiens.com we trust! #gohabsgo game 7 boys let's do this * Marc Lanctot: This is great! I didn't even know #AISTATS was happening now and this is the second conference photo I have seen of it already! More please! 👍👌🙏 * hardmaru: If GitHub were built in: Japan 🇯🇵 China 🇨🇳 North Korea 🇰🇵 The EU 🇪🇺 * Ethan Mollick: I am not sure I would agree with all of this post (by a well-known researcher at OpenAI), but the relationship between Anthropic and Claude is quite d... * Ethan Mollick: The single most accurate science fiction author writing about AI turned out to be… Douglas Adams He wrote about AIs that work best when emotionally m... * Emily M. Bender: Wesleyan Tetris * Emily M. Bender: I'm a little late to the dunk-on-Dawkins party (but hey, let's keep the celebration rolling!) but I finally read his essay (minus the chatbot outputs ... * Emily M. Bender: Join us tomorrow! X Posts: * Andrej Karpathy: 2025 LLM Year in Review * Andrej Karpathy: LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest.... * Andrej Karpathy: I'm being accused of overhyping the [site everyone heard too much about today already]. * Simon Willison: once you attach them to a good coding agent harness at least * Simon Willison: I'm beginning to suspect that a key skill in working effectively with coding agents is developing an intuition for when you don't need to * Simon Willison: Vibe coding is irresponsibly building software through dice rolls, not caring what code is produced * Simon Willison: This may be the best guidance I've seen anywhere on writing a really good commit history. * Harrison Chase: Visibility is the easiest piece. The hard part is analyzing and understanding what you're observing. I've spoken to teams recording 100k+ * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: as always, it's an exciting time to be working at LangChain! * Jim Fan: The Second Pre-training Paradigm * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jeremy Howard: Controversial opinion - the language best placed to win at deep learning is: F#. * Jeremy Howard: Early reports from people using this are that it's the real deal. Strong coding. Good multilingual. Consistent over long contexts. * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Soumith Chintala: Open LLMs need to get organized and co-ordinated about sharing human feedback. * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Yann LeCun: Yann LeCun's Billion Dollar Bet. www.youtube.com. * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Clem Delangue: $4.15B invested in open-source generates $8.8T of value for companies (aka $1 invested in open-source = $2,000 of value created) - Companies would nee... * Max Woolf: LOL * Max Woolf: * Max Woolf: * Sasha Rush: On the infra side, composer 2 uses CP. This is (i think?) the first real detail from using CP on MLA. My understanding is that each rank * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to ... * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can ... * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should ... * Stas Bekman: Modern art. Artist: PyTorch memory profiler Model: Llama-8B The piece on the left is the ... * Sayak Paul: Working at Hugging Face over the past 3.5+ years has allowed me to identify what technical areas truly interest me! In turn, that has allowed me to di... * Sayak Paul: We're shipping an elaborate guide on how to profile diffusion pipelines in Diffusers to set them... * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Told an AI agent to read the autoresearch repo and build a version for QMD. Get training data from tobi/qmd github. Went to sleep. Woke up to a 0.8B m... * Philipp Schmid: Random thought. We are going to be so much faster at creating and building. * Ethan Mollick: So much work is going into faking continual learning and memory for AIs * Ethan Mollick: As someone who is pretty good at keeping up with AI, I can barely keep up with it all. * Ethan Mollick: "Load bearing," "I keep coming back to," "Not X, but Y" A curse of using AI a lot is that * Ethan Mollick: Ethan Mollick profile. Ethan Mollick. ✓. emollick. Apr 25. Ethan Mollick's Image on X. 32. 337. 4220. 243898 ·. * Naomi Saphra: I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Angela Zhou: #throwback coz it's finally the day again!!! #HellOnWheels back on AMC 9/8c tonight! * Angela Zhou: Best work breaks #onset #HellonWheels -- dunno who's cuter, @ansonmount or Mac his dog? * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Building a theory of the architecture of organizing machines and people. * Ben Recht: Fully open machine learning requires not only GPU access but a community commitment to openness. Blog Articles: * How to Work and Compound with AI — Eugene Yan: Context as infra, taste as config, verification for autonomy, scale via delegation, closing the loop.

Stars & Posts — 2026-05-02

Sat, 02 May 2026 12:00:00 GMT
The conversation today is dominated by a sudden shift in sentiment, as AI leaders and the broader tech community grapple with whipsawing from "AI is a bubble" fears to a stark realization that we are running out of high-quality training data. Ethan Mollick highlighted this pivot in a thoughtful Bluesky post, pointing to an *Atlantic* piece that captures the industry's collective whiplash. On GitHub, the trend is toward efficiency and resourcefulness, with starred repos focusing on synthetic data generation and fine-tuning smaller models to squeeze more value out of existing datasets. Notable projects include a new open-source tool for data augmentation and a lightweight agent framework designed to run on consumer hardware, signaling a pragmatic turn away from the "bigger is better" arms race. The underlying theme is clear: the community is preparing for a post-scarcity data landscape, where the winners will be those who can do more with less. GitHub Repos: * microsoft/lib0xc — Safe(ish) C programming library [C] Bluesky Posts: * Ethan Mollick: I was quoted a couple times in this Atlantic article, but that isn’t (the only) reason I think it is good. It lays out the reasons why we whipsawed fr... * Simon Willison: I added a new feature to my blog (built entirely on my phone with Claude code for web) that imports my iNaturalist photos and adds them to my site's o... * Mark Riedl: status * Mark Riedl: Stunned * Mark Riedl: Toilet makers are winners in the AI revolution * Marc Lanctot: My son made a #geometrydash level! I tried it.. I'm so bad 🤣 took me 111 attempts even though it's super short. You can try it, too: look up user h... * Nathan Lambert: Most obvious immediate impressions on coming back to the US from China. 1. The cars here are so lame. So many cool EVs in China, feels like I went 2 d... * Emily M. Bender: More MAITH3k coming right up! @alexhanna.bsky.social and I are delighted to be talking with @nathaliemarechal.net on our next live stream, in which ... * Emily M. Bender: Watch this short, powerful, very on-point speech. Literal shivers watching @mjcrockett.bsky.social make the most of their platform! * angela zhou: a hello atmosphere post! the goal is to start writing more in between 280char and 32pg main + 30+pg appendix. having internalized various horrors of t... X Posts: * Andrej Karpathy: LLMs are emerging as a new kind of intelligence, simultaneously a lot smarter than I expected and a lot dumber than I expected. In any case they * Andrej Karpathy: A few random notes from claude coding quite a bit last few weeks. Coding workflow. Given the latest lift in LLM * Simon Willison: This may be the best guidance I've seen anywhere on writing a really good commit history. * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the ... * Simon Willison: Quitting programming as a career right now because of LLMs would be like quitting carpentry as a ... * Harrison Chase: Visibility is the easiest piece. The hard part is analyzing and understanding what you're observing. I've spoken to teams recording 100k+ * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: RT @samecrowder: as always, it's an exciting time to be working at LangChain! * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild * Jeremy Howard: I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Jeremy Howard: Absolutely any time I try to explore something even slightly against commonly accepted beliefs, * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: Open LLMs need to get organized and co-ordinated about sharing human feedback. * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Yann LeCun: My opinion of @elonmusk I like his cars (I own a 2015 S, and 2023 S), his rockets, his solar energy systems, * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Max Woolf: LOL * Max Woolf: me irl * Sasha Rush: Sasha Rush (@srush_nlp). 7 likes. * Sasha Rush: Sasha Rush (@srush_nlp). 15 likes. * Sasha Rush: ⛏️ * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: If you're trying out FA4, you're likely to run into not being able to load cutlass.cute * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Philipp Schmid: How to use Deep Research with the Gemini API. www.philschmid.de. * Ethan Mollick: Very cool analysis of the submissions to a major management journal that shows how much the * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: After reading it, this does seem like a big deal. Industry experts outlined important, real-world, hard tasks for AI to do. * Ethan Mollick: AI is actually pretty good at ideas as well. https://t.co/AhnzrnkN03 * Emily M. Bender: EMILY M. BENDER: Yeah. And so passive, like, oops, the moon, the moon went further away. It's like no, actually, you made some decisions. * Emily M. Bender: Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positioned in. * Emily M. Bender: Facebook (sorry: Meta) AI: Check out our "AI" that lets you access all of humanity's knowledge. * Naomi Saphra: New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Naomi Saphra: I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the * Angela Zhou: #throwback coz it's finally the day again!!! #HellOnWheels back on AMC 9/8c tonight! * Angela Zhou: Best work breaks #onset #HellonWheels -- dunno who's cuter, @ansonmount or Mac his dog? * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Building a theory of the architecture of organizing machines and people. * Ben Recht: Fully open machine learning requires not only GPU access but a community commitment to openness.

Stars & Posts — 2026-05-01

Fri, 01 May 2026 12:00:00 GMT
Today’s AI discourse centers on methodological rigor and the responsible communication of research findings, sparked by Ethan Mollick’s reflection on a pre-registered RCT—where he clarified that a 0.3 standard deviation effect, while notable, is a modest outcome that should be framed precisely rather than as a "big result." Meanwhile, Emily M. Bender shared a lighter but equally significant moment, hinting at an upcoming podcast episode that promises to explore the intersection of fun and critical analysis in AI research. On GitHub, the trending repos lean heavily into tooling for reproducibility and agentic workflows, with several new projects focused on fine-grained evaluation frameworks and open-source alternatives to proprietary model APIs. The overarching theme is a push for transparency and grounded expectations, as the community grapples with both the hype and the hard data behind recent AI advancements. Bluesky Posts: * Ethan Mollick: I deleted this post since I think I was imprecise in the language. It is an interesting pre-registered RCT, but I should have been clearer that .3 SD... * Emily M. Bender: We had SO MUCH FUN!! I'm curious to see what this will sound like as a pod ep. * Mark Riedl: Legal system having a totally normal one today * Mark Riedl: Not too long ago someone I follow introduced a citation checking tool. I cannot find it anymore (and cannot search posts from only people I follow). C... * Ethan Mollick: New paper (on an old AI model) tests o1 against doctors on medical benchmarks & real ER cases: “across a variety of scenarios and applications, the la... * Emily M. Bender: But wait there's more! Fresh off our live show in Brooklyn, @alexhanna.bsky.social and I will be doing the next MAIHT3K livestream on Monday May 4. We... X Posts: * Simon Willison: Our evaluation of OpenAI's GPT-5.5 cyber capabilities. The UK's AI Security Institute previously evaluated Claude Mythos: now they've evaluated GPT-5.... * Simon Willison: Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant... * Simon Willison: llm 0.31 released: supports GPT-5.5 and adds a verbosity parameter for controlling output detail on OpenAI's latest models. * Simon Willison: llm 0.32a0 alpha: major backwards-compatible refactor. Models can now be prompted with a list of messages, OpenAI Chat Completions style. * Simon Willison: DeepSeek V4 - almost on the frontier, a fraction of the price. * Harrison Chase: as always, it's an exciting time to be working at LangChain! * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: In the hot path as the agent is running. The agent can decided to (or the user can prompt it to) update its memory as it is working on the core * Harrison Chase: traces matter! * Jim Fan: The first time I met Jensen was also the first time I met @elonmusk. I was interning at OpenAI that day and * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Jeremy Howard: Here's what I would prefer to see: * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Soumith Chintala: anyone else feel burned out by a new AI breakthrough every week? * Francois Chollet: I think it's clear that for many smaller companies that invested in deep learning, it turned out * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Francois Chollet: To really understand a concept, you have to 'invent' it yourself in some capacity. * Yann LeCun: To qualify as Science a piece of research must be correct and reproducible. To be correct and reproducible, ... * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Clem Delangue: Great research on open-source by. : - $4.15B invested in open-source generates $8.8T of value for companies (aka $1 invested in open-source = $2,000 o... * Max Woolf: me irl * Max Woolf: @simonw * Sasha Rush: ⛏️ * Sasha Rush: No content extracted. * Sasha Rush: No content extracted. * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to * Stas Bekman: The @PyTorch team are working on a new super important tool: https://t.co/rnfpDuvgOI This * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Sayak Paul: Working at Hugging Face over the past 3.5+ years has allowed me to identify what technical areas truly interest me! In turn, that has allowed me to di... * Philipp Schmid: Every good story has to end, and after 4 incredible years at @huggingface, it's time for me to start my next adventure. When I joined Hugging Face, we... * Ethan Mollick: Very cool analysis of the submissions to a major management journal that shows how much the * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: One thing thing about AI, for better and worse, is that 'everything around me is somebody's life * Ethan Mollick: AI is actually pretty good at ideas as well. https://t.co/AhnzrnkN03 * Naomi Saphra: I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the * Naomi Saphra: New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Naomi Saphra: This book starts like it's gonna be a fun microhistory of TB (it gave us the Stetson! * Angela Zhou: #throwback to the beginnings of a beautiful friendship =D @ansonmount @HellOnWheelsAMC * Ben Recht: And awesome to see many Berkeley alums thriving here. @LaurentLessard, @DimitrisPapail, and Shivaram * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Why does framing decision, design, and discovery as optimization remain so irresistible?

Stars & Posts — 2026-04-30

Thu, 30 Apr 2026 12:00:00 GMT
Zig's blanket ban on AI-assisted contributions sparked a key debate today, with Simon Willison noting the project's rationale—that PR review is about mentoring new contributors, not just code quality—resonating widely. Meanwhile, Nathan Lambert stirred discussion on AGI governance, arguing that Demis Hassabis remains the most trustworthy CEO on the topic, especially given Google's public-company checks versus the private statuses of Anthropic and OpenAI. On a lighter note, Marc Lanctot shared a joyful reaction to a video, offering a brief respite from the heavier AI governance conversations. Overall, the day's themes centered on the tension between AI tooling and open-source community growth, alongside ongoing scrutiny of AGI leadership accountability. GitHub Repos: * criterion-rs/criterion.rs [Rust] Bluesky Posts: * Simon Willison: The Zig project's rationale for their blanket ban on AI-assisted contributions makes a lot of sense to me - for them, time spent reviewing PRs isn't a... * Marc Lanctot: Love this! 😁 www.youtube.com/watch?v=Rjfb... * Nathan Lambert: Demis is the only acceptable answer of which CEO do you trust most with AGI (doubly so until Anthropic/OpenAI go public, Google being public is a gr... * Simon Willison: Saw this white-crowned sparrow having a lot of a sing * Mark Riedl: Okay y'all, it's official. See you in Atlanta in December. * Mark Riedl: Space garbage * Mark Riedl: As someone obsessed with whether AI can play D&D, I am going to have fun with the ChatGPT goblin phenomenon for a bit. You can ignore me. But if you ... * Mark Riedl: Designating that most uses of the term "goblin" and "gremlin" are not legitimate while designating most uses of the word "frog" as legitimate is the i... * Mark Riedl: Reward your model for being nerdy, get nerdy behavior. Train in your own output, pollute your dataset with verbal tics openai.com/index/where-... * Mark Riedl: First rule of goblin club: openai.com/index/where-... * Margaret Mitchell: Companies may not fully realize why they lean into the narratives they do. But it’s important to reflect on the larger picture of what’s happening, wh... * Ethan Mollick: Forget goblins, things GPT-5.5 likes in its fiction: lighthouses, the ocean, maps, bells, clock towers with bells that ring impossible times, Mira Val... * Ethan Mollick: "Load bearing," "I keep coming back to," "Not just X, but Y" A curse of using AI a lot is that you realize how much of the writing around you is jus... * Ethan Mollick: It is really interesting that Microsoft and OpenAI have access to the exact same models at the exact same time, and they have done such different thin... * Emily M. Bender: Calling all NYC folks! Here is a great opportunity to see @ghostdoc2026.bsky.social -- Sat May 2 event.newschool.edu/ghostinthema... * angela zhou: A few years back we looked into gerrymandering / computational research for a "fun project" but after reading many articles and a bit of legal comment... X Posts: * Andrej Karpathy: My most amusing interaction was where the model (I think I was given some earlier version with a * Andrej Karpathy: One common issue with personalization in all LLMs is how distracting memory seems to be for the models. * Andrej Karpathy: Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around * Andrej Karpathy: LLMs are emerging as a new kind of intelligence, simultaneously a lot smarter than I expected and a lot dumber than I expected. In any case they * Andrej Karpathy: I'm being accused of overhyping the [site everyone heard too much about today already]. * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the... * Simon Willison: Quitting programming as a career right now because of LLMs would be like quitting carpentry as a... * Simon Willison: The last year six months in LLMs, illustrated by pelicans on bicycles. I've published video, slides and a detailed annotated transcript from my talk a... * Simon Willison: [Image: a stylish image of a 3D computer game, with two raccoons sneaking down a street past a futuristic looking building... Prompt was: "Screenshot ... * Harrison Chase: RT @samecrowder: as always, it's an exciting time to be working at LangChain! * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: In the hot path as the agent is running. The agent can decided to (or the user can prompt it to) update its memory as it is working on the core * Harrison Chase: traces matter! * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild * Jim Fan: The first time I met Jensen was also the first time I met @elonmusk. I was interning at OpenAI that day and * Jeremy Howard: Here's what I would prefer to see: * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: I'm giving the opening Keynote at ICML 2024 on Tuesday the 23rd @ 9:30am CEST. * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Francois Chollet: Re-reading an article I wrote in 2017, and I'm finding I could have written it yesterday * Yann LeCun: Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labor market. * Yann LeCun: To qualify as Science a piece of research must be correct and reproducible. To be correct and reproducible, * Yann LeCun: It seems to me that before "urgently figuring out how to control AI systems much smarter than us" we need * Yann LeCun: The emergence of superintelligence is not going to be an event. We don't have anything close to a * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Fei-Fei Li: AI’s next frontier is Spatial Intelligence, a technology that will turn seeing into reasoning, perception into action, and imagination into creation. ... * Max Woolf: me irl * Max Woolf: No text content extracted, only likes count. * Max Woolf: @simonw * Sasha Rush: Some personal news: I recently joined Cursor. Cursor is a small, ambitious team, and they've created * Sasha Rush: Wager established. Jonathan Frankle (@jefrankle) stepped up to my Transformer long bet. * Stas Bekman: Classical Jensen math. Unidirectional bandwidth is topped at 450GB/s, and then there comes a protocol overhead of two digit percentage. * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: If you're trying out FA4, you're likely to run into not being able to load cutlass.cute * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Sayak Paul: Release notes: Release Diffusers 0.34.0: New Image and Video Models, Better torch. * Sayak Paul: Had a nice time chatting about the state of diffusion models and some text-to-image data shenanigans at * Sayak Paul: Details: * Philipp Schmid: How to use Deep Research with the Gemini API. www.philschmid.de. * Philipp Schmid: Google DeepMind and Korea Partner to Accelerate Scientific Discovery. deepmind.google. * Ethan Mollick: Very cool analysis of the submissions to a major management journal that shows how much the * Ethan Mollick: I pointed Claude Cowork at a set of 107 documents (PPTs, Word docs, Excel) that were initially * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: One thing thing about AI, for better and worse, is that "everything around me is somebody's life * Emily M. Bender: @emilymbender.bsky.social. emilymbender. Feb 10. Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positione... * Emily M. Bender: @parismarx @alexhanna @ctaylsaurus EMILY M. BENDER: Yeah. And so passive, like, oops, the moon, the moon went further away. It's like no, actually, yo... * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the * Ben Recht: And awesome to see many Berkeley alums thriving here. @LaurentLessard, @DimitrisPapail, and Shivaram * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Everyone knows actions are fundamentally different than predictions, but it's hard to write this

Stars & Posts — 2026-04-29

Wed, 29 Apr 2026 12:00:00 GMT
Today's AI landscape sees a notable pivot from human-led prompt engineering to automated optimization, with hardmaru sharing research that trains an AI to coax the best performance out of other LLMs—a meta-approach that could redefine how we interact with models. Meanwhile, Marc Lanctot offers a nostalgic counterpoint, documenting his complete playthrough of the classic *Dungeons of Dr. Creep* with detailed level breakdowns and videos, reminding us that AI researchers still find value in retro gaming analysis. The juxtaposition of these posts highlights a broader tension: as the field races toward self-improving systems, there's a parallel appreciation for the human craft of understanding complex systems, whether they be dungeon puzzles or transformer architectures. GitHub Repos: * GliaX/Stethoscope — A research-validated stethoscope whose plans are available Freely and openly. The cost of the entire stethoscope is between $2.5 to $5 to produce [Ruby] Bluesky Posts: * Marc Lanctot: Hello all, I finished another old game: Dungeons of Dr. Creep! Just like last year, I documented each level in a reddit thread and made a few videos ... * hardmaru: For the past few years, humans have been doing “prompt engineering” to coax the best performance out of different LLMs. In this work, we explored what... * Simon Willison: I released LLM 0.32a0 this morning, a major backwards-compatible refactor of my LLM Python library and CLI tool for working with language models - the... * Mark Riedl: Congratulations to Dr. Gennie Mansi, for successfully defending her PhD thesis. Dr. Mansi's work investigates AI in healthcare; how AI impacts legal... * Nathan Lambert: Let’s goooooooooo we are capybara’d up, thanks Qwen, keep the models coming * Ethan Mollick: Gemini now can create documents, and it is a nice start, but not up to the frontier yet, as you can see from my "evil buyout of Hogwarts" test. Power... * Ethan Mollick: One reason I don’t think “judgment” is going to be a distinctly human role in working with AI is that the most recent agentic models have gotten quite... * Emily M. Bender: Usually, when I get interviewed for a piece on something like "AI consciousness" I am relegated to the skeptics box --- some short paragraph near the ... * Emily M. Bender: Mystery AI Hype Theater 3000 Ep 76: www.buzzsprout.com/admin/212641... Carmen Maria Machado joins @alexhanna.bsky.social and me to get into the why ... * Emily M. Bender: Also available as video on Peertube: peertube.dair-institute.org/w/tgEjwXf8ST... * Emily M. Bender: Mystery AI Hype Theater 3000 Ep 76: www.buzzsprout.com/2126417/epis... Carmen Maria Machado joins @alexhanna.bsky.social and me to get into the why ... X Posts: * Andrej Karpathy: +1 for "context engineering" over "prompt engineering". When in every industrial-strength LLM app, context engineering is the delicate art and science... * Andrej Karpathy: LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest.... * Andrej Karpathy: 2025 LLM Year in Review By training LLMs against automatically verifiable rewards across a number of environments (e.g. think math/code puzzles), the... * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the * Simon Willison: This may be the best guidance I've seen anywhere on writing a really good commit history. * Harrison Chase: In the hot path as the agent is running. The agent can decided to (or the user can prompt it to) update its memory as it is working on the core * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: I am not excited about visual workflow builders 1. Not simple enough for the average user * Harrison Chase: Memory actually allows for a better agent building experience. Agent building is very iterative - in large part because you don't know what the * Harrison Chase: When building agents, you need to iterate on production data much more than when building traditional software. You need to iterate on how * Jim Fan: The first time I met Jensen was also the first time I met @elonmusk. I was interning at OpenAI that day and * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild * Jeremy Howard: I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: Open LLMs need to get organized and co-ordinated about sharing human feedback. * Soumith Chintala: MacStudio you ask? Apple Engineering's **actual** time spent on PyTorch support * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Francois Chollet: I think it's clear that for many smaller companies that invested in deep learning, it turned out * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Clem Delangue: https://t.co/4CQthIKm8F * Clem Delangue: Great research on open-source by. : - $4.15B invested in open-source generates $8.8T of value for companies (aka $1 invested in open-source = $2,000 o... * Max Woolf: Max Woolf (@minimaxir). 19 likes. * Max Woolf: congrats to OpenAI on winning the Turing Test * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: If you're trying out FA4, you're likely to run into not being able to load cutlass.cute * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Sayak Paul: Had a nice time chatting about the state of diffusion models and some text-to-image data shenanigans at * Sayak Paul: For me, it was Keras among other things that inspired me to take up deep learning as a potential * Sayak Paul: Working at Hugging Face over the past 3.5+ years has allowed me to identify what technical areas truly interest me! In turn, that has allowed me to di... * Philipp Schmid: Gemini Embedding 2 now GA! One embedding model that understand text, images, video, audio, and PDFs! * Philipp Schmid: Excited to introduce the Gemini Interactions API, a unified interface for Gemini models and agents. Starting today with Gemini Deep Research Agent. - ... * Ethan Mollick: I pointed Claude Cowork at a set of 107 documents (PPTs, Word docs, Excel) that were initially * Ethan Mollick: On the plus side with Opus 4.7, if it does decide to think it produces BY FAR the best * Ethan Mollick: We are starting to see some nuanced discussions of what it means to work with advanced AI In this * Ethan Mollick: Very cool analysis of the submissions to a major management journal that shows how much the * Emily M. Bender: EMILY M. BENDER: Yeah. And so passive, like, oops, the moon, the moon went further away. It's like no, actually, you made some decisions. * Emily M. Bender: Image is of the 1990s Microsoft writing assistant character Clippy with its eyebrows raised positioned in. * Emily M. Bender: Facebook (sorry: Meta) AI: Check out our "AI" that lets you access all of humanity's knowledge. * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Ben Recht: And awesome to see many Berkeley alums thriving here. @LaurentLessard, @DimitrisPapail, and Shivaram * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Everyone knows actions are fundamentally different than predictions, but it's hard to write this Blog Articles: * LLM 0.32a0 is a major backwards-compatible refactor — Simon Willison:
I just released LLM 0.32a0, an alpha release of my

Stars & Posts — 2026-04-27

Mon, 27 Apr 2026 12:00:00 GMT
Today's AI discourse centers on the evolving role of large language models in creative and interactive design, highlighted by Ethan Mollick’s revelation that GPT-5.5 in Codex generated a surprisingly robust tabletop RPG game master’s and player guide, complete with self-administered playtesting. While Mollick notes the output leans into storytelling and retains distinct "LLM-y" artifacts, the experiment signals a growing trend of models not just generating content but simulating iterative design processes. This aligns with broader GitHub trending activity around agentic frameworks and tool-use optimization, where developers are pushing models to handle complex, multi-step tasks like game balancing and narrative branching. The discussion underscores a key tension: as models become more competent at structured creativity, the community is debating whether these "playtested" outputs represent genuine utility or just sophisticated mimicry of human design workflows. GitHub Repos: * asimovinc/asimov-v1 — v1 of Asimov, an open-source humanoid robot * naubiomech/OpenExo — Open Source Exoskeleton [HTML] Bluesky Posts: * Ethan Mollick: GPT-5.5 in Codex made a surprisingly solid table top RPG game masters guide & player guide, which it "playtested." It leans into the storytelling aspe... * Simon Willison: Love this so much (Also definitive proof that humans are so much better than machines) * Simon Willison: Microsoft's MIT licensed VibeVoice speech-to-text model (think Whisper with speaker diarization) is really good - my notes on running the 5.71GB 4bit ... * Simon Willison: Today OpenAI announced that "Revenue share payments from OpenAI to Microsoft continue through 2030, independent of OpenAI’s technology progress" That... * Nathan Lambert: 🤘🤘🤘 (Peep the bitter lesson sign, in the Radiohead meeting room) * Yoshua Bengio: Safety and innovation are not mutually exclusive: for many companies, especially in high-trust industries, AI’s risks are also a hindrance to adoption... * Ethan Mollick: Very cool analysis of the submissions to a major management journal that shows how much the system of science, built for humans, is under strain as a ... * Naomi Saphra: ok, this historical LM I actually think is trained with their claimed limitations. it doesn't know what commercials are (first attested in its modern ... * angela zhou: sunday shims for thought - i use the web interface for chatgpt pro since i get it for free from my sister (gpu whip). llms are good at reading 50, 60 ... X Posts: * Andrej Karpathy: Very interested in what the coming era of highly bespoke software might look like. Example from this morning - I've become a bit loosy goosy with my c... * Andrej Karpathy: By training LLMs against automatically verifiable rewards across a number of environments (e.g. think math/code puzzles), the LLMs spontaneously devel... * Andrej Karpathy: LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. ... * Andrej Karpathy: I've never felt this much behind as a programmer. I have a sense that I could be 10X more powerful if I just properly string together what has become ... * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the * Harrison Chase: RT @samecrowder: as always, it's an exciting time to be working at LangChain! * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: traces matter! * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: We are living in a timeline where a non-US company is keeping the original mission of OpenAI alive - truly * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jeremy Howard: I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Jeremy Howard: Absolutely any time I try to explore something even slightly against commonly accepted beliefs, * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Soumith Chintala: anyone else feel burned out by a new AI breakthrough every week? * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Clem Delangue: If you can’t stop small teams from using your API for distillation, then you’re definitely not stopping criminals, biohackers, or adversarial states f... * Max Woolf: me irl * Phil Wang: I got to cover for the excellent @HadleyFreeman in the Guardian today so * Sasha Rush: Wager established. Jonathan Frankle (@jefrankle) stepped up to my Transformer long bet. * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: If you're trying out FA4, you're likely to run into not being able to load cutlass.cute * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Sayak Paul: Working at Hugging Face over the past 3.5+ years has allowed me to identify what technical areas truly interest me! In turn, that has allowed me to di... * Philipp Schmid: I read three technical reports from Moonshot AI's Kimi K2.5 paper, Cursor's Composer 2 report and blog post, and Chroma's Context-1 write-up * Philipp Schmid: Background Coding Agents: Context Engineering (Honk, Part 2) | Spotify Engineering * Ethan Mollick: No content available (27 replies). * Ethan Mollick: I pointed Claude Cowork at a set of 107 documents (PPTs, Word docs, Excel) that were initially * Ethan Mollick: We are starting to see some nuanced discussions of what it means to work with advanced AI In this * Ethan Mollick: If it helps, I teach at a business school & many of my smartest students are hired by funds because they can reliably turn their only-human * Ethan Mollick: This paper is even more insane to read than the thread. Not only do models become completely * Emily M. Bender: For those playing along at home, here's a "AI is sentient!" argument bingo card. Title: AI sentience/consciousness argument bingo Squares: You can’t p... * Angela Zhou: #throwback coz it's finally the day again!!! #HellOnWheels back on AMC 9/8c tonight! * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: In honor of the 39th AI Winter, I'm going to spend the week disentangling the culture and code of * Ben Recht: Fully open machine learning requires not only GPU access but a community commitment to openness. Blog Articles: * Tracking the history of the now-deceased OpenAI Microsoft AGI clause — Simon Willison:
For many years, Microsoft and OpenAI's relationship has included a weird clause saying that, should AGI be achieved, Microsoft's commercial IP righ...

Stars & Posts — 2026-04-26

Sun, 26 Apr 2026 12:00:00 GMT
AI leaders are on the move globally, with Hugging Face's Nathan Lambert touching down in Beijing and Hangzhou to engage with China's AI community—a sign of deepening cross-border technical exchange. Meanwhile, researcher Naomi Saphra sparked curiosity with a cryptic mention of an unconfirmed event at the WHPCD (White House Presidential Council on Diversity), though she pivoted to praising palliative care specialists, leaving the tech world guessing. On GitHub, trending repos reflect a continued focus on open-weight model fine-tuning and agentic tooling, though no single breakout project dominated today's stars. The conversation remains split between geopolitical AI diplomacy and the quieter, essential infrastructure work keeping the ecosystem running. Bluesky Posts: * Nathan Lambert: In Beijing and Hangzhou this week to get to know the AI community here! * Naomi Saphra: Found out something cool while trying to figure out what just happened at the WHPCD today. Did not figure out what happened at the WHPCD. But I’m alwa... * hardmaru: Scaling up massive LLMs continues to yield incredible results. But to truly unlock their full potential, the next frontier is test-time compute and dy... * Ethan Mollick: Since so much of the AI job debate is among economists, they miss that as jobs are disrupted, professions will compete over new boundaries Abbot's S... * Ethan Mollick: * Ethan Mollick: This is a useful image for thinking about the curve we are on and what likely comes next in an intuitively understandable way. (Roon is a long time m... * Emily M. Bender: The plagiarism machine doesn't just traffic in text. * Naomi Saphra: If you missed Sara's poster at #ICLR2026, the good news is you can still read her paper! X Posts: * Andrej Karpathy: My most amusing interaction was where the model (I think I was given some earlier version with a ...) * Andrej Karpathy: Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. * Andrej Karpathy: By training LLMs against automatically verifiable rewards across a number of environments (e.g. think math/code puzzles), the LLMs spontaneously devel... * Andrej Karpathy: We're missing (at least one) major paradigm for LLM learning. Not sure what to call it, * Simon Willison: It's interesting how "better at code" has become the defining goal of almost every AI lab over the * Harrison Chase: im excited about agent harnesses because i think are the first stable agent abstractions we can build on top (which is why we're investing so much in ... * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: as always, it's an exciting time to be working at LangChain! * Jim Fan: Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jim Fan: It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere. * Jim Fan: Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild * Jeremy Howard: Here's a complete unedited video of asking Grok for its views on the Israel/Palestine situation. It first searches twitter for what Elon thinks. * Jeremy Howard: Here's what I would prefer to see: * Soumith Chintala: reading 'AI News' (previously Smol Talk) is probably the highest-leverage 45 mins * Soumith Chintala: Sometimes we forget that NVIDIA wins because it's a software company. * Soumith Chintala: Open LLMs need to get organized and co-ordinated about sharing human feedback. * Francois Chollet: I think it's clear that for many smaller companies that invested in deep learning, it turned out * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * Francois Chollet: Current AI is a librarian of existing knowledge. Science requires an explorer of the unknown. * Francois Chollet: Reaching AGI won't be beating a benchmark. It will be the end of the human-AI gap. * Yann LeCun: It seems to me that before "urgently figuring out how to control AI systems much smarter than us" we need * Yann LeCun: The emergence of superintelligence is not going to be an event. We don't have anything close to a * Yann LeCun: The AI industry is completely LLM-pilled. Everybody is working on the same thing. * Fei-Fei Li: I can now confess that I participated in the new #TronAres movie, playing myself. I had a great time working with everyone especially Greta * Clem Delangue: Great research on open-source by. : - $4.15B invested in open-source generates $8.8T of value for companies (aka $1 invested in open-source = $2,000 o... * Max Woolf: LOL. * Sasha Rush: On the infra side, composer 2 uses CP. This is (i think?) the first real detail from using CP on MLA. My understanding is that each rank * Stas Bekman: I have been compiling LLM/VLM training logbooks/chronicles. This is the one of the best sources to ... * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can ... * Stas Bekman: To remind - this is the memory saving you get when enabling TiledMLP :) Left: normal memory ... * Stas Bekman: Modern art. Artist: PyTorch memory profiler Model: Llama-8B The piece on the left is the ... * Sayak Paul: Working at Hugging Face over the past 3.5+ years has allowed me to identify what technical areas truly interest me! In turn, that has allowed me to di... * Sayak Paul: Based on. * Sayak Paul: My presentation at the @PyTorch Conf EU is now live. It's an exciting piece given its emphasis on how we make Diffusers play quite well w/ `torch.comp... * Ethan Mollick: I guarantee that any industry expert, with a little time and effort, can make a better (or at least more focused) skill than the default * Ethan Mollick: I pointed Claude Cowork at a set of 107 documents (PPTs, Word docs, Excel) that were initially * Ethan Mollick: We are starting to see some nuanced discussions of what it means to work with advanced AI In this * Ethan Mollick: If it helps, I teach at a business school & many of my smartest students are hired by funds because they can reliably turn their only-human * Ethan Mollick: Classic study gave 146 economist teams the same dataset & got wildly different answers New * Emily M. Bender: For those playing along at home, here's a "AI is sentient!" argument bingo card. Title: AI sentience/consciousness argument bingo Squares: You can’t p... * Naomi Saphra: what a perfect space for scientific discourse! I'll start off with a few images of myself * Naomi Saphra: Life update: I'm starting as faculty at Boston University in 2026! BU ... * Naomi Saphra: I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Revisiting Sutton's Bitter Lesson in the wake of GPT-5. * Ben Recht: Fully open machine learning requires not only GPU access but a community commitment to openness.

Stars & Posts — 2026-04-25

Sat, 25 Apr 2026 12:00:00 GMT
Today’s discourse centers on the operational challenges of multi-agent AI systems, with Ethan Mollick pinpointing organizational design and collaborative benchmarking as the next "critical frontier" for enterprise value. Meanwhile, Emily M. Bender introduces the term "demythifying" from a review of *The AI Con*, signaling a continued pushback against AI hype. On GitHub, repositories focused on agent orchestration frameworks and evaluation toolkits saw a surge in stars, reflecting the community’s pivot from single-model capabilities to managing agent swarms at scale. The tension between scaling agentic systems and maintaining rigorous, myth-busting critique remains the dominant theme of the day. GitHub Repos: * ROCm/FlyDSL — FlyDSL is the Python front‑end of the project: Flexible LaYout DSL. [Python] Bluesky Posts: * Ethan Mollick: Organizational design for agents is hard, benchmarking agents working in concert is hard. Together, this is the next critical frontier for making AI m... * Emily M. Bender: Favorite new to me word, from a review of The AI Con: demythifying. * Simon Willison: I think ChatGPT Images 2.0 deciding to add a "WHY ARE YOU LIKE THIS" sign to the background of this image is the first time I've felt a glimpse of AGI... * Ethan Mollick: If you believe that AI is going to have a big impact on work and life, the only real tool for mitigating bad impacts and channeling usage for good wil... * Ethan Mollick: I think that academia has not absorbed the fact that AI agents are now good enough to independently reconstruct complex papers without access to code ... * angela zhou: im just a rat that types and writes papers (cough revises) at verve coffee X Posts: * Andrej Karpathy: Bought a new Mac mini to properly tinker with claws over the weekend. The apple store person told me they are selling like hotcakes and everyone is co... * Andrej Karpathy: 2025 LLM Year in Review * Andrej Karpathy: Very interested in what the coming era of highly bespoke software ... Example from this morning - I've become a bit loosy goosy with my cardio recentl... * Simon Willison: I'm beginning to suspect that a key skill in working effectively with coding agents is developing an intuition for when you don't need to * Simon Willison: Vibe coding is irresponsibly building software through dice rolls, not caring what code is produced * Harrison Chase: im excited about agent harnesses because i think are the first stable agent abstractions we can build on top (which is why we're investing so much in ... * Harrison Chase: This means that operations you would do on code in the software world, you now do on traces in the agent world. Debugging, testing, profiling * Harrison Chase: TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this * Harrison Chase: When you ship traditional software to production, you have a good sense of what to expect. Users click buttons, fill out forms, * Jim Fan: I've been a bit quiet on X recently. The past year has been a transformational experience. * Jeremy Howard: I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in * Jeremy Howard: Absolutely any time I try to explore something even slightly against commonly accepted beliefs, * Soumith Chintala: reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins * Francois Chollet: I think it's clear that for many smaller companies that invested in deep learning, it turned out * Francois Chollet: Folks who work in AI or software engineering feel like the world is changing exponential fast. * David Ha: Don't miss David Ha @hardmaru's keynote at @ALifeConf #ALIFE2021 on "World Models and Attention for Reinforcement Learning"! * David Ha: It's spectacular to have followed David Ha's (@hardmaru) incredible career arc —MD of Fixed Income at Goldman Sachs —restarted his career * Yann LeCun: It seems to me that before "urgently figuring out how to control AI systems much smarter than us" we need * Yann LeCun: An A.I. Pioneer Warns the Tech 'Herd' Is Marching Into a Dead End. www.nytimes.com. * Yann LeCun: The emergence of superintelligence is not going to be an event. We don't have anything close to a * Fei-Fei Li: Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ... * Max Woolf: me irl * Sasha Rush: On the infra side, composer 2 uses CP. This is (i think?) the first real detail from using CP on MLA. My understanding is that each rank first compute... * Sasha Rush: ⛏️ * Sasha Rush: * Stas Bekman: If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should * Stas Bekman: Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul * Stas Bekman: If you're trying out FA4, you're likely to run into not being able to load cutlass.cute * Stas Bekman: Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can * Sayak Paul: Install `diffusers` from source and start using Kontext from @bfl_ml 🧨 Use your favorite optims, too :) Training is also supported (@linoy_tsaban a... * Sayak Paul: Release notes: Release Diffusers 0.34.0: New Image and Video Models, Better torch. * Philipp Schmid: Guide: ReAct agent from scratch with Gemini 2.5 and LangGraph | Gemini API | Google AI for Developers. ai.google.dev. * Ethan Mollick: AI is actually pretty good at ideas as well. * Ethan Mollick: My most popular AI post was a bunch of made-up "graphs" four years ago. * Ethan Mollick: So much work is going into faking continual learning and memory for AIs, * Ethan Mollick: If it helps, I teach at a business school & many of my smartest students are hired by funds because they can reliably turn their only-human * Emily M. Bender: @kohntom A synthetic text extruding machine is not well-matched to any application where the accuracy of the content matters. This is clearly one such... * Naomi Saphra: This book starts like it's gonna be a fun microhistory of TB (it gave us the Stetson! * Naomi Saphra: New preprint! Everyone loves causal interp. It's coherently defined! It makes testable predictions * Naomi Saphra: New preprint! Phase transitions! We love to see them during LM training. * Angela Zhou: #throwback to the beginnings of a beautiful friendship =D @ansonmount @HellOnWheelsAMC #HellonWheels #onlocation. * Ben Recht: I weigh in on the Trump administration’s newfound obsession with Gold Standard Science and reproducibility. Though it’s not all in bad faith, it’s lik... * Ben Recht: For the first time in almost a decade, I'm teaching a class on learning and control. * Ben Recht: Revisiting Sutton's Bitter Lesson in the wake of GPT-5.