Intelligence.Log
2026-04-28
Extracted: 63 items. Sources: GitHub, Bluesky, X.
++ AI OVERVIEW ++
A fascinating tension is emerging in the AI world today between vintage and cutting-edge. The standout project is **talkie**, a "vintage language model" co-created by Alec Radford and trained exclusively on 260 billion tokens of pre-1931 English text—small enough to run on-device, sparking Ethan Mollick's thought experiment about a fully Downton Abbey-era Siri. Mollick is also probing whether such a model can independently "invent" later technologies like modern coding from first principles, raising deep questions about the nature of knowledge and progress. Meanwhile, Nathan Lambert reports feeling "the AGI at Zhipu AI," hinting that the frontier of state-of-the-art intelligence is advancing rapidly in China. The day's discussion thus bridges historical constraints and futuristic ambitions, with talkie offering a unique sandbox for testing how much modern capability is latent in older language.
grep TOPIC=
grep SOURCE=
sort --by=
GH
PyWry is a cross-platform app factory, rendering engine and UI toolkit for Python that produces native desktop, web, and notebook experiences from a single API.
“PyWry is a cross-platform app factory that lets you build native desktop, web, and notebook experiences from a single Python API. It leverages Tauri and WebView2 for rendering, and integrates with Jupyter, Plotly, and MCP servers, making it a versatile tool for creating rich interactive applications.”
GH
Practical, Colab-friendly notebooks for fine-tuning and running audio AI models
“Smol-audio provides practical, Colab-friendly notebooks for fine-tuning and running audio AI models, making advanced audio AI accessible without heavy infrastructure. It focuses on hands-on experimentation with state-of-the-art audio models.”
BSKY
Some notes on talkie, a new "vintage language model" from a team including Alec Radford (yes, that Alec Radford) "trained on 260B tokens of historical pre-1931 English text" simonwillison.net/2026/Apr/28/...
❤️ 20 Likes|[LLM][Fine-tuning]
BSKY
BSKY
The new LLM trained only on pre-1931 text is small enough that it can potentially run on device, so, with the right tools, you can get a fully vintage version of Siri, but from the era of Downton Abbey (also a small model). Here, I asked for it to arrange for sushi delivery in Philadelphia. Hmmm...
❤️ 54 Likes|[LLM][Deployment]
BSKY
Here is an AI trained just using text from 1931 or earlier, which leads to a lot of interesting experiments: can the model independently develop later inventions? Can it learn to code from examples alone? You can talk to the model here: talkie-lm.com/chat Details here: talkie-lm.com/introducing-...
❤️ 95 Likes|[LLM]
BSKY
I would very much like to see the 2,000 lb stellar sea lion at San Francisco Pier 39, who I believe has now been named "Chonkers" Does anyone know if he keeps a regular schedule?
❤️ 89 Likes|
BSKY
BSKY
BSKY
A big problem with all AI at work punditry right now is that it all rests on data from the pre-agentic era (which is basically just now ending) and we have very little information about what has been happening since the Claude Code moment. So everything now requires some caveat.
❤️ 54 Likes|[Agent]
BSKY
This is an actual line that was added to the official system prompt for Codex for GPT-5.5 by OpenAI. Usually the system prompt is as minimal as possible, so I assume it would otherwise mention goblins a lot. AIs are weird.
❤️ 1248 Likes|[LLM]
BSKY
Reading Anthropic's recent nonsense about "emotion vectors" and was struck by this remark. Is there really such a taboo? Because we see anthropomorphizing language about "AI" systems *all the time*.
❤️ 105 Likes|[Safety]
BSKY
This is written by and for linguists, but I suspect there is useful information here no matter what your field, if your field touches on people.
❤️ 11 Likes|[LLM]
BSKY
Really proud to have been a part of this paper, with Rob, Martin, Alicia, Alex, Anna and @kirbyconrod.bsky.social Check it out for how to conceptualize "race" and "ethnicity" in linguistic research. Spoiler alert: it's not something that can be addressed with word choice at the very end.
❤️ 18 Likes|[Evaluation][Safety]
BSKY
BSKY
I got a call from the “assistant” of someone I have an existing business relationship with. it was a robot with fake office conversation and keyboard typing sounds in the background. Genuinely think it should be illegal to not immediately disclose it’s automated, especially with deceptive realism.
❤️ 37 Likes|[Safety]
BSKY
To my Madison people: I’ll be talking about The Irrational Decision at 12:30 tomorrow at the Wisconsin Institute for Discovery. Would be great to see you there. silo.wisc.edu/talk/2026-04...
❤️ 7 Likes|[Agent]
X
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 confused :) I'm definitely a bit sus'd to run OpenClaw specifically - giving my private data/keys to 400K lines of vibe coded
[Agent][Tooling]
“DeepSeek Summary: Karpathy bought a Mac mini to experiment with 'claws' (likely a typo for 'Claude' or an agent framework) and expresses concern about running open-source code with private data.”
X
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 cardio recently so I decided to do a more srs, regimented experiment to try to lower my Resting Heart Rate from 50 -> 45, over https://t.co/EDULdIpWmE
[Agent][Tooling]
“DeepSeek Summary: Karpathy envisions a future of highly personalized software, using his own health experiment as an example.”
X
By training LLMs against automatically verifiable rewards across a number of environments (e.g. think math/code puzzles), the LLMs spontaneously develop strategies that look like 'reasoning' to humans - they learn to break down problem solving into intermediate calculations and they learn a number of probl
[LLM][Fine-tuning][Evaluation]
“DeepSeek Summary: Karpathy explains how training LLMs with verifiable rewards leads to emergent reasoning behaviors.”
X
It's interesting how "better at code" has become the defining goal of almost every AI lab over the
[LLM][Evaluation]
“DeepSeek Summary: Simon Willison observes that improving code generation has become the primary objective for AI labs.”
X
I came up with a somewhat foolish new benchmark for testing image generation models, to exercise the new ChatGPT Images 2.0:
[Multi-modal][Evaluation]
“DeepSeek Summary: Simon Willison created a benchmark for testing image generation models, specifically for ChatGPT Images 2.0.”
X
No tweet text available (profile/engagement page).
“DeepSeek Summary: No substantive content from this result.”
X
RT @samecrowder: as always, it's an exciting time to be working at LangChain!
[Tooling]
“DeepSeek Summary: Retweet expressing excitement about working at LangChain.”
X
TL;DR: More and more agents need a workspace: a computer where they can run code, install packages, and access files. Sandboxes provide this
[Agent][Infra]
“DeepSeek Summary: Agents require sandboxed environments for code execution and file access.”
X
traces matter!
[Evaluation][Tooling]
“DeepSeek Summary: Emphasizes the importance of tracing in AI systems.”
X
Resource constraints are a beautiful thing. Survival instinct in a cut-throat AI competitive land
[Agent][Deployment]
“DeepSeek Summary: Resource constraints foster innovation and survival instinct in competitive AI landscape.”
X
I've been a bit quiet on X recently. The past year has been a transformational experience.
[Agent]
“DeepSeek Summary: Jim Fan acknowledges a period of silence and hints at significant personal or professional change.”
X
It gives me a lot of comfort knowing that we are the last generation without advanced robots everywhere.
[Multi-modal][Safety]
“DeepSeek Summary: Reflects on the imminent ubiquity of advanced robotics and its societal impact.”
X
Everyone's freaking out about vibe coding. In the holiday spirit, allow me to share my anxiety on the wild
[Agent][LLM]
“DeepSeek Summary: Jim Fan comments on the hype around 'vibe coding' and shares his own concerns.”
X
I replicated this result, that Grok focuses nearly entirely on finding out what Elon thinks in
[Safety][Evaluation]
“DeepSeek Summary: Jeremy replicated a finding that Grok AI is heavily biased toward Elon Musk's opinions.”
X
Absolutely any time I try to explore something even slightly against commonly accepted beliefs,
[Safety]
“DeepSeek Summary: Jeremy notes that exploring ideas against common beliefs is met with resistance.”
X
reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins
[LLM]
“DeepSeek Summary: Soumith recommends reading 'AI News' as a high-leverage use of time.”
X
I think it's clear that for many smaller companies that invested in deep learning, it turned out
[Evaluation]
“DeepSeek Summary: Chollet comments on the outcome of deep learning investments for smaller companies.”
X
Folks who work in AI or software engineering feel like the world is changing exponential fast.
[LLM]
“DeepSeek Summary: Chollet observes that AI and software engineers perceive rapid exponential change.”
X
Unveiling our new startup Advanced Machine Intelligence (AMI Labs). We just completed our seed round: $1.03B / 890M€, one the largest seeds ever, probably the largest for a European company. We're hiring! [the background image is the Veil Nebula - a picture I took from my
[Agent][Infra]
“DeepSeek Summary: Yann LeCun announces his new startup AMI Labs with a record seed round of $1.03B.”
X
Very excited to share @theworldlabs 's latest research work RTFM!! It's a real-time, ...
[Multi-modal][Agent]
“DeepSeek Summary: Fei-Fei Li announces RTFM, a real-time research work from The World Labs.”
X
congrats to OpenAI on winning the Turing Test
[LLM]
“DeepSeek Summary: Sarcastic or genuine congratulations to OpenAI for Turing Test achievement.”
X
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
[Infra][LLM]
“DeepSeek Summary: Sasha Rush discusses infrastructure details about composer 2 using CP (context parallelism) on MLA (Multi-Head Latent Attention), noting it's the first real detail from using CP on MLA.”
X
If you were holding off to try @MSFTDeepSpeed ZeRO++ it looks like deepspeed@master should
[Infra][Deployment]
“DeepSeek Summary: Stas notes that DeepSpeed ZeRO++ is now available on master branch, encouraging users to try it.”
X
Hear, hear, I'm excited to introduce a new performance metric: Maximum Achievable Matmul
[Evaluation][Infra]
“DeepSeek Summary: Stas introduces a new metric called Maximum Achievable Matmul for evaluating performance.”
X
If you're trying out FA4, you're likely to run into not being able to load cutlass.cute
[Infra][Tooling]
“DeepSeek Summary: Stas warns about a common issue with FA4 involving cutlass.cute loading.”
X
Thanks to an awesome contribution from @omarnomad The Machine Learning Engineering Open book now can
[LLM][Tooling]
“DeepSeek Summary: Stas thanks a contributor for enhancing the Machine Learning Engineering Open Book.”
X
Had a nice time chatting about the state of diffusion models and some text-to-image data shenanigans at
[Multi-modal]
“DeepSeek Summary: Discussed diffusion models and text-to-image data issues.”
X
X
For me, it was Keras among other things that inspired me to take up deep learning as a potential
[Fine-tuning]
“DeepSeek Summary: Keras inspired his deep learning journey.”
X
How to run a local coding agent with Gemma 4 and Pi
[Agent][Deployment]
“DeepSeek Summary: Philipp Schmid shares a guide on running a local coding agent using Gemma 4 and Pi.”
X
Guide: ReAct agent from scratch with Gemini 2.5 and LangGraph | Gemini API | Google AI for Developers
[Agent][LLM]
“DeepSeek Summary: Philipp Schmid published a guide on building a ReAct agent from scratch using Gemini 2.5 and LangGraph.”
X
Excited to introduce the Gemini Interactions API, a unified interface for Gemini models and agents. Starting today with Gemini Deep Research Agent. - Unifies access to models and agents via a single RESTful endpoint. - Access Gemini Deep Research agent via API. Just read this new research paper from Google AI called
[LLM][Infra]
“DeepSeek Summary: Announces the Gemini Interactions API, providing a unified RESTful interface for Gemini models and agents, starting with the Deep Research Agent.”
X
Very cool analysis of the submissions to a major management journal that shows how much the
[Evaluation]
“DeepSeek Summary: Analysis of management journal submissions reveals significant patterns.”
X
I pointed Claude Cowork at a set of 107 documents (PPTs, Word docs, Excel) that were initially
[Agent][Tooling]
“DeepSeek Summary: Testing Claude Cowork with a large set of documents to see how it handles multi-format data.”
X
We are starting to see some nuanced discussions of what it means to work with advanced AI In this
[LLM]
“DeepSeek Summary: Nuanced discussions emerging about working with advanced AI.”
X
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
[Deployment]
“DeepSeek Summary: Business school students are hired for their human skills that complement AI.”
X
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.
[Safety]
“DeepSeek Summary: Bender critiques passive language in AI discourse, emphasizing human agency in decision-making.”
X
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 sentience. Joining us are AI researchers Alex Hanna and Emily M. Bender
[Evaluation]
“DeepSeek Summary: Bender announces participation in a discussion about the Turing Test and AI sentience.”
X
For those playing along at home, here's a "AI is sentient!" argument bingo card.
[Safety]
“DeepSeek Summary: Bender shares a bingo card for common arguments claiming AI sentience.”
X
what a perfect space for scientific discourse! I'll start off with a few images of myself
[Safety]
“DeepSeek Summary: Sarcastic comment about using images for scientific discourse.”
X
Life update: I'm starting as faculty at Boston University in 2026! BU ...
[LLM]
“DeepSeek Summary: Announcement of new faculty position at Boston University.”
X
I work on understanding and improving training for NLP models, with a focus on studying how structures and mechanistic behaviors emerge over the
[LLM][Fine-tuning]
“DeepSeek Summary: Describes her research focus on NLP model training and mechanistic behaviors.”
X
In honor of the 39th AI Winter, I'm going to spend the week disentangling the culture and code of
[Evaluation]
“DeepSeek Summary: Ben Recht humorously acknowledges an 'AI Winter' and plans to analyze the culture and code behind it.”
X
And awesome to see many Berkeley alums thriving here. @LaurentLessard, @DimitrisPapail, and Shivaram
“DeepSeek Summary: Ben Recht celebrates the success of UC Berkeley alumni in the field.”
X
For the first time in almost a decade, I'm teaching a class on learning and control.
[Evaluation]
“DeepSeek Summary: Ben Recht announces teaching a course on learning and control after a long hiatus.”
-- END OF LOG --
[STATS] 63 items · Filter applied
Powered by Horizon + DeepSeek