AI PulseLIVE

Anyone figured out a recipe to run Gemma 4 E2B or E4B against audio files locally on a Mac yet?

I just sent out my sponsors-only monthly roundup newsletter covering March, which means it's time to publish my February newsletter for free for everyone else. Here's what you could get a month earlier as a sponsor: https://t.co/AX2Ws3Peg4

My piece in the Economist where I argue against de-weirding AI. It is a strange technology with both risks & opportunities that need to be discovered. Pretending AI works like normal IT automation can result in bad outcomes for companies & their employees. https://t.co/qeseHSfy9Q

Meet Gemma 4: our new family of open models you can run on your own hardware. Built for advanced reasoning and agentic workflows, we’re releasing them under an Apache 2.0 license. Here’s what’s new 🧵

ChatGPT is now available in CarPlay. The voice mode you know, now available on-the-go. Rolling out to iPhone users running iOS 26.4+ where CarPlay is supported. https://t.co/yk3qdLa99r

TBPN is my favorite tech show. We want them to keep that going and for them to do what they do so well. I don't expect them to go any easier on us, am sure I'll do my part to help enable that with occasional stupid decisions.

Pelicans for Gemma 4 E2B, E4B, 26B-A4B and 31B - the first three generated on my laptop via LM Studio, the 31B was broken on my laptop so I ran it via the Gemini API instead https://t.co/MEa6O7VzdB https://t.co/Qa1cGhnopg

New report from us: Can you prompt inject your way to an “A”? As LLMs increasingly are used as judges, people are inserting AI prompts into letters, CVs & papers. We tested whether it works. It does on older & smaller models, but not on most frontier AI: https://t.co/rSdYAq7HWD https://t.co/O918Er77Xo

LLM Knowledge Bases Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. In this way, a large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it. So: Data ingest: I index source documents (articles, papers, repos, datasets, images, etc.) into a raw/ directory, then I use an LLM to incrementally "compile" a wiki, which is just a collection of .md files in a directory structure. The wiki includes summaries of all the data in raw/, backlinks, and then it categorizes data into concepts, writes articles for them, and links them all. To convert web articles into .md files I like to use the Obsidian Web Clipper extension, and then I also use a hotkey to download all the related images to local so that my LLM can easily reference them. IDE: I use Obsidian as the IDE "frontend" where I can view the raw data, the the compiled wiki, and the derived visualizations. Important to note that the LLM writes and maintains all of the data of the wiki, I rarely touch it directly. I've played with a few Obsidian plugins to render and view data in other ways (e.g. Marp for slides). Q&A: Where things get interesting is that once your wiki is big enough (e.g. mine on some recent research is ~100 articles and ~400K words), you can ask your LLM agent all kinds of complex questions against the wiki, and it will go off, research the answers, etc. I thought I had to reach for fancy RAG, but the LLM has been pretty good about auto-maintaining index files and brief summaries of all the documents and it reads all the important related data fairly easily at this ~small scale. Output: Instead of getting answers in text/terminal, I like to have it render markdown files for me, or slide shows (Marp format), or matplotlib images, all of which I then view again in Obsidian. You can imagine many other visual output formats depending on the query. Often, I end up "filing" the outputs back into the wiki to enhance it for further queries. So my own explorations and queries always "add up" in the knowledge base. Linting: I've run some LLM "health checks" over the wiki to e.g. find inconsistent data, impute missing data (with web searchers), find interesting connections for new article candidates, etc., to incrementally clean up the wiki and enhance its overall data integrity. The LLMs are quite good at suggesting further questions to ask and look into. Extra tools: I find myself developing additional tools to process the data, e.g. I vibe coded a small and naive search engine over the wiki, which I both use directly (in a web ui), but more often I want to hand it off to an LLM via CLI as a tool for larger queries. Further explorations: As the repo grows, the natural desire is to also think about synthetic data generation + finetuning to have your LLM "know" the data in its weights instead of just context windows. TLDR: raw data from a given number of sources is collected, then compiled by an LLM into a .md wiki, then operated on by various CLIs by the LLM to do Q&A and to incrementally enhance the wiki, and all of it viewable in Obsidian. You rarely ever write or edit the wiki manually, it's the domain of the LLM. I think there is room here for an incredible new product instead of a hacky collection of scripts.

I was a guest on @lennysan's podcast! We talked about agentic engineering and all sorts of other LLM-related topics for 1h39m(!), plus a little bit about kākāpō parrots - here's my selection of highlights from our conversation https://t.co/pj0usRQX1u

The AI labs have actually done a bad job explaining what the future they are building towards will actually look like for most of us. Even “Machines of Loving Grace” has very few well-articulated visions of what Anthropic hopes life will be like if they succeed at their goals.

The vast majority of social science is based on the assumption that the future looks like the past. It is usually a good bet! Not sure it will apply to the impacts of AI, though.

A pre-ChatGPT paper that seems relevant: information processing seems to be the key to growth throughout the long span of recorded history. Better information processing tools are, the paper argues, what keeps societies from collapse.

Maybe we shouldn’t have given away the fact that you can crash Opus by asking about California’s High Speed Rail delays in Armenian. Would’ve been useful to have that in our back pocket, just in case,

A sign that human creativity is a bottleneck is that this year everyone can generate almost any image or video they can think of for nearly free and the April Fools posts are basically just as bad as any other year.

Today we're excited to announce NO_FLICKER mode for Claude Code in the terminal It uses an experimental new renderer that we're excited about. The renderer is early and has tradeoffs, but already we've found that most internal users prefer it over the old renderer. It also supports mouse events (yes, in a terminal). Try it: CLAUDE_CODE_NO_FLICKER=1 claude

If you have NPM package axios in your dependencies you need to make sure it's pinned to a known safe version, sounds like there's another supply chain attack in play

The fact that every scientific paper in 2026 is still uploaded only as fully formatted PDFs to academic archive sites that often limit downloads tells you everything you need to know about how quickly the scientific system is adjusting to the potential of AI to accelerate science

New supply chain attack this time for npm axios, the most popular HTTP client library with 300M weekly downloads. Scanning my system I found a use imported from googleworkspace/cli from a few days ago when I was experimenting with gmail/gcal cli. The installed version (luckily) resolved to an unaffected 1.13.5, but the project dependency is not pinned, meaning that if I did this earlier today the code would have resolved to latest and I'd be pwned. It's possible to personally defend against these to some extent with local settings e.g. release-age constraints, or containers or etc, but I think ultimately the defaults of package management projects (pip, npm etc) have to change so that a single infection (usually luckily fairly temporary in nature due to security scanning) does not spread through users at random and at scale via unpinned dependencies. More comprehensive article: https://t.co/EJAZbqAPIQ

All AI policy is haunted by a failure of imagination. It is either nothing happens or apotheosis, people can’t seem to conceive of any other futures. It is amazing to predict massive AI development and expect nothing major to change in employment or productivity.

My prediction for the latest trend in academic buildings: Faraday cage testing halls (including bathrooms) with no signal for assessment.

Today, we closed our latest funding round with $122 billion in committed capital at an $852B post-money valuation. The fastest way to expand AI’s benefits is to put useful intelligence in people’s hands early and let access compound globally. This funding gives us resources to lead at scale. https://t.co/sY7YNUPSYO

Anyone know what the "Current week (Sonnet only)" bar in Claude Code means when you run the "/usage" command?

The biggest bottleneck in AI for most people isn't the models. It's the chatbot. New interfaces like Claude Dispatch, are closing the gap between what AI can do and what people can actually use it for. For many folks, that is where leaps will come from. https://t.co/E32ysD2Cac

I wanted to share a bunch of my favorite hidden and under-utilized features in Claude Code. I'll focus on the ones I use the most. Here goes.

Mr. Chatterbox is a new 2GB nanochat model trained from scratch by Trip Venturella on "28,000 Victorian-era British texts published between 1837 and 1899" - I released an llm-mrchatterbox plugin which can run it locally on my Mac https://t.co/EIu15Wszev

This is a very good post:

Here's the official line from GitHub on the ads in PRs story - I find this explanation credible personally, it looks to me like it was a poorly considered product tips feature that got out of hand

Georgi on why it's still hard to get great coding agent performance from local models: "Note that the main issues that people currently unknowingly face with local models mostly revolve around the harness and some intricacies around model chat templates and prompt construction"

A weird part of working at Anthropic: getting a few of these each day https://t.co/cuOKBo0rPr

Want to talk to the past? Here is an LLM "trained entirely from scratch on a corpus of over 28,000 Victorian-era British texts published between 1837 and 1899, drawn from a dataset made available by the British Library." Quite different from an LLM roleplaying a Victorian. https://t.co/5jl7SyJjAP

The research team (including @hamsabastani who is on X) found that letting students just use AI resulted in them using it to accidentally shortcut learning But both that study and a separate RCT found that AIs prompted to act as a tutor improved learning https://t.co/0HtjGC8eU0 https://t.co/U3OIeCF4aP

Asking Codex to build a SimGothicManor game and really enjoying how much of its internal planning monologue has become obsessed with tongue-in-cheek gothic, such as worrying about "scope creep in a velvet cape" https://t.co/AtixDfXWNM

- 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 argue the opposite. - LLM demolishes the entire argument and convinces me that the opposite is in fact true. - lol The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.

X search, long broken, has finally been fixed. All it took was a multi-hundred-million-dollar AI model (Seriously, X search is notoriously bad, but Grok is very good at this specific task, though it is funny that I have to burn a ton of tokens to do a keyword search on the site)

Curious if there has been any good articles written on the impact of VLMs on low-vision and blind people. The advent of a universal text reading, and visual description system seems like it would be a big advance as a result of AI, but haven't seen anything written about it.

It helps to think of ARC-AGI-3 as a different test entirely than the previous ARC-AGIs. It measures different things (though, as in the previous tests, precisely what it measures isn’t clear) and has different rules. That doesn’t mean it isn’t good, but it is its own thing.

It is worth noting the absolute confidence of the leading AI labs that they can continue to release ever more powerful models for the near future. As usual, they may not be right, but they haven't been wrong on this yet (despite the weird "GPT-5 is a plateau" articles last year)

The coolest meeting I had this week with was Paul, who used ChatGPT and other LLMs to create an mRNA vaccine protocol to save his dog Rosie. It is amazing story. "The chat bots empowered me as an individual to act with the power of a research institute - planning, education, troubleshooting, compliance, and yes, real scientific design work in converting genomic data to a vaccine prescription and designing the treatment protocol around it. But they worked alongside humans at every step. The combination is what made it possible." It immediately got me thinking "this should be a company". Also, Paul is an extraordinary guy. This should be easy to do, but it is not yet.

Soon to be announced: first-ever American Bootlicker Award

Last year everyone spoke about over building of AI data centers, likely this year will start to demonstrate that there is not nearly enough compute to meet demand I think to degree to which AI is currently subsidized depends on the model, but agree with everything else here

I know these are all unreliable leaks of internal code names but please, please AI labs, the only thing worse than calling your models GPT-5.5-xhigh-Codex-nano is giving them names like Agent Smith or Mythos, for obvious reasons. https://t.co/lYKDtd0MAp

The first steel beams went up this week at our Michigan Stargate site with Oracle and Related Digital https://t.co/Hl0NBqwfnS

One way to see the advancement of AI is to see how much further you can get with new models on the same hardware Here is "an otter using a laptop on an airplane" generated on my home computer using the open weights Wan 2.1, first try. We have come pretty far in 18 months. https://t.co/c4iz9UcmTB

I've been vibe coding SwiftUI menu bar apps for my new Mac, turns out Claude Opus 4.6 and GPT-5.4 are both competent at Swift programming, no need to even open Xcode! https://t.co/1Vi0rtlVh1

Say hello to Gemini 3.1 Flash Live. 🗣️ Our latest audio model delivers more natural conversations with improved function calling – making it more useful and informed. Here’s what’s new 🧵

When I built menugen ~1 year ago, I observed that the hardest part by far was not the code itself, it was the plethora of services you have to assemble like IKEA furniture to make it real, the DevOps: services, payments, auth, database, security, domain names, etc... I am really looking forward to a day where I could simply tell my agent: "build menugen" (referencing the post) and it would just work. The whole thing up to the deployed web page. The agent would have to browse a number of services, read the docs, get all the api keys, make everything work, debug it in dev, and deploy to prod. This is the actually hard part, not the code itself. Or rather, the better way to think about it is that the entire DevOps lifecycle has to become code, in addition to the necessary sensors/actuators of the CLIs/APIs with agent-native ergonomics. And there should be no need to visit web pages, click buttons, or anything like that for the human. It's easy to state, it's now just barely technically possible and expected to work maybe, but it definitely requires from-scratch re-design, work and thought. Very exciting direction!

As AI gets better at holding natural conversations, we need to understand how these interactions impact society. We’re sharing new research into how AI might be misused to exploit emotions or manipulate people into making harmful choices. 🧵 https://t.co/CamLP2wM9t

To me this mostly illustrates the futility of robust jailbreaking prevention

The background noise in open offices, in multiple experiments, decreases the ability to do analytical work, increase bugs in software, and decrease the ability to find bugs. Open floor plans are just a bad idea for any solo work. https://t.co/Y2KX5HjIEd

Great little story from @danshapiro about how he asked a coding agent to fix the official webcam software from Canon that kept crashing. He woke up to a new, fully functional Rust webcam app that has worked ever since. https://t.co/uazUapOQb0 https://t.co/3BVb2PU1uN

Today was a good day https://t.co/iME3drw1Xf

You can now create longer tracks with Lyria 3 Pro. 🎶 Map out intros, verses, choruses, and bridges to build high-fidelity compositions up to 3 minutes long. 🎹

One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity. Some kind of trying too hard.

The more AI can do, the more we need to ask what it should and shouldn’t do. OpenAI researcher @w01fe joins host @AndrewMayne to explore the Model Spec, the public framework that defines how models are intended to behave. They break down how it works in practice, from the chain of command that resolves conflicting instructions to the way it evolves over time through real-world use, feedback, and new model capabilities.

Little known fact, the Anthropic Labs team (the team I joined Anthropic to be on) shipped: - MCP - Skills - Claude Desktop app - Claude Code It was just a few of us, shipping fast, trying to keep pace with what the model was capable of. Those early Desktop computer use prototypes, back in the Sonnet 3.6 days, felt clunky and slow. But it was easy to squint and imagine all the ways people might use it once it got really good. Fast forward to today. I am so excited to release full computer use in Cowork and Dispatch. Really excited to see what you do with it!

Turns out you can run enormous Mixture-of-Experts on Mac hardware without fitting the whole model in RAM by streaming a subset of expert weights from SSD for each generated token - and people keep finding ways to run bigger models Kimi 2.5 is 1T, but only 32B active so fits 96GB

Google DeepMind 🤝 Agile Robots Our new research partnership will integrate the Gemini foundation models with their hardware to help build the next generation of more helpful and useful robots. Find out more → https://t.co/dptWjeZwya https://t.co/bzPUUvPiJp

Thankfully the LiteLLM package has now been marked as "quarantined" on PyPI so attempting to install the compromised update via pip et al shouldn't work https://t.co/BmrbWCoLXn

Watch how fast Gemini 3.1 Flash-Lite can generate websites. ⚡ This browser creates each page in real-time as you click, search, and navigate. Give it a try → https://t.co/h3W5o1wItY

Software horror: litellm PyPI supply chain attack. Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds, Kubernetes configs, git credentials, env vars (all your API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, database passwords. LiteLLM itself has 97 million downloads per month which is already terrible, but much worse, the contagion spreads to any project that depends on litellm. For example, if you did `pip install dspy` (which depended on litellm>=1.64.0), you'd also be pwnd. Same for any other large project that depended on litellm. Afaict the poisoned version was up for only less than ~1 hour. The attack had a bug which led to its discovery - Callum McMahon was using an MCP plugin inside Cursor that pulled in litellm as a transitive dependency. When litellm 1.82.8 installed, their machine ran out of RAM and crashed. So if the attacker didn't vibe code this attack it could have been undetected for many days or weeks. Supply chain attacks like this are basically the scariest thing imaginable in modern software. Every time you install any depedency you could be pulling in a poisoned package anywhere deep inside its entire depedency tree. This is especially risky with large projects that might have lots and lots of dependencies. The credentials that do get stolen in each attack can then be used to take over more accounts and compromise more packages. Classical software engineering would have you believe that dependencies are good (we're building pyramids from bricks), but imo this has to be re-evaluated, and it's why I've been so growingly averse to them, preferring to use LLMs to "yoink" functionality when it's simple enough and possible.

AI will help discover new science, such as cures for diseases, which is perhaps the most important way to increase quality of life long-term. AI will also present new threats to society that we have to address. No company can sufficiently mitigate these on their own; we will need a society-wide response to things like novel bio threats, a massive and fast change to the economy, extremely capable models causing complex emergent effects across society, and more. These are the areas the OpenAI Foundation will initially focus on, and in my opinion are some of the most important ones for us to get right. The Foundation will spend at least $1 billion over the next year. @woj_zaremba, co-founder of OpenAI, will transition to Head of AI Resilience. I believe that shifting how the world thinks about safety to include a Resilience-style approach is critical, and I am extremely grateful to Wojciech for taking on this role. Wojciech has been my cofounder for the last decade; anyone who knows him will understand what I mean when I say he is one of a kind. He has a lot of ideas about how we build a new kind of AI safety. @JacobTref is joining as Head of Life Sciences and Curing Diseases. @annaadeola, our VP of Global Impact, will transition to Head of AI for Civil Society and Philanthropy. @robert_kaiden is joining as Chief Financial Officer. @jeffarnold is joining as Director of Operations.

I would like a single word for this phrase: "throw it into the maw with every bit of context I can think of".

no 👏 more 👏 permission prompts 👏

Starlette 1.0 is out! I used this as an opportunity to experiment with Claude Skills, since Claude isn't yet familiar with the (minor) breaking changes in the 1.0 release compared to 0.x https://t.co/IxQ18DShxg

I have loved being on the Helion board; I continue to be extremely excited about a future with abundant energy and Helion in particular. As Helion and OpenAI start to explore working together at significant scale, it is difficult for me to be on both boards. (I will have a financial interest in Helion so still be recused from negotiations, but from a governance perspective this will make things easier for both companies.) David and Chris are exceptional founders and I think people will be quite impressed by the company's progress.

It’s now easier to find, reuse, and build on the files you upload and create in ChatGPT. You can quickly reference files in a chat using recent files in the toolbar, ask ChatGPT about something you’ve uploaded, or browse your files in the new Library tab in the web sidebar. Rolling out globally for Plus, Pro, and Business users, and coming soon to users in the EEA, Switzerland, and the UK.

New surveillance dystopia prompt: try running "Profile this user" against 1,000 comments by someone on Hacker News to see what an LLM can figure out https://t.co/Wa5PuqojnK

Join our team

Thank you Sarah, my pleasure to come on the pod! And happy to do some more Q&A in the replies.

Anyone know if it's possible to checkout two private repos at the same time in Claude Code for web? As far as I can tell it isn't, because all Git operations go through a local proxy which only allows authenticated access to the private repot that the session is attached to

Still a work in progress, but I've published the first draft of a new chapter on "Using Git with coding agents" https://t.co/uP2DGVcMtQ

Had to go see Project Hail Mary right away (it's based on the book of Andy Weir, of also The Martian fame). Both very pleased and relieved to say that 1) the movie sticks very close to the book in both content and tone and 2) is really well executed. The book is one of my favorites when it comes to alien portrayals because a lot of thought was clearly given to the scientific details of an alternate biochemistry, evolutionary history, sensorium, psychology, language, tech tree, etc. It's different enough that it is highly creative and plausible, but also similar enough that you get a compelling story and one of the best bromances in fiction. Not to mention the other (single-cellular) aliens. I can count fictional portrayals of aliens of this depth on one hand. A lot of these aspects are briefly featured - if you read the book you'll spot them but if you haven't, the movie can't spend the time to do them justice. I'll say that the movie inches a little too much into the superhero movie tropes with the pacing, the quips, the Bathos and such for my taste, and we get a little bit less the grand of Interstellar and a little bit less of the science of The Martian, but I think it's ok considering the tone of the original content. And it does really well where it counts - on Rocky and the bromance. Thank you to the film crew for the gem!

Are you up for a challenge? https://t.co/GNryIDhnut https://t.co/ZX7ZiuhGgu

Thank you Jensen and NVIDIA! She’s a real beauty! I was told I’d be getting a secret gift, with a hint that it requires 20 amps. (So I knew it had to be good). She’ll make for a beautiful, spacious home for my Dobby the House Elf claw, among lots of other tinkering, thank you!!

I have so much gratitude to people who wrote extremely complex software character-by-character. It already feels difficult to remember how much effort it really took. Thank you for getting us to this point.

GPT-5.4 mini is available today in ChatGPT, Codex, and the API. Optimized for coding, computer use, multimodal understanding, and subagents. And it’s 2x faster than GPT-5 mini. https://t.co/DKh2cC5S3F https://t.co/sirArgn37L

How do we measure progress toward AGI? It takes a village – and a bit of healthy competition. 🛠️ We’re launching a global hackathon with @Kaggle to build new cognitive evaluations for AI. With $200k in prizes up for grabs, help us put our framework to the test. Join the challenge → https://t.co/UnqRNV1oXO

AI is starting to help solve real issues in healthcare for patients and doctors. OpenAI’s Head of Health Dr. Nate Gross and Health AI Research Lead Karan Singhal join @AndrewMayne to discuss how we're building new models and products to meet the world's health needs. https://t.co/UTsSCAqVyt

The Codex team are hardcore builders and it really comes through in what they create. No surprise all the hardcore builders I know have switched to Codex. Usage of Codex is growing very fast: https://t.co/lRKcNJDY8n

Great first week for 5.4 in the API. Builders building fast.

It is also very smart, but...I generally agree with this, and really feel it myself on the 5.3 -> 5.4 upgrade.

We’re excited to unveil the name of our new London building: Platform 37. 📍 The name honors both the surrounding area’s transport heritage and "Move 37" – the critical moment where our AI system AlphaGo showed it could find novel solutions humans hadn't considered. https://t.co/lxw3HGYU1v

What does it take to build AI for scientific discovery? 🧠 To celebrate 10 years of AlphaGo, @ThoreG and @Pushmeet joined @fryrsquared on our podcast to discuss how mastering games has paved the way for it to help solve more complex problems. ↓ 00:00 The AlphaGo match 02:15 Why Go? 10:58 Lee Sedol vs AlphaGo 14:58 Move 37 20:55 Move 78 24:38 Reaction from the Go community 30:45 Never before seen footage 32:05 AlphaGo to protein folding 34:00 Matrix multiplication 38:00 AlphaGo and algorithmic discovery 41:40 How to verify new discoveries 47:38 Role of mathematicians 51:43 Would we be here without AlphaGo?

Expectation: the age of the IDE is over Reality: we’re going to need a bigger IDE (imo). It just looks very different because humans now move upwards and program at a higher level - the basic unit of interest is not one file but one agent. It’s still programming.

My autoresearch labs got wiped out in the oauth outage. Have to think through failovers. Intelligence brownouts will be interesting - the planet losing IQ points when frontier AI stutters.

We’re acquiring Promptfoo. Their technology will strengthen agentic security testing and evaluation capabilities in OpenAI Frontier. Promptfoo will remain open source under the current license, and we will continue to service and support current customers. https://t.co/xhmLmJRoUZ

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project. This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.: - It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work. - It found that the Value Embeddings really like regularization and I wasn't applying any (oops). - It found that my banded attention was too conservative (i forgot to tune it). - It found that AdamW betas were all messed up. - It tuned the weight decay schedule. - It tuned the network initialization. This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism. https://t.co/WAz8aIztKT All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges. And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

The next step for autoresearch is that it has to be asynchronously massively collaborative for agents (think: SETI@home style). The goal is not to emulate a single PhD student, it's to emulate a research community of them. Current code synchronously grows a single thread of commits in a particular research direction. But the original repo is more of a seed, from which could sprout commits contributed by agents on all kinds of different research directions or for different compute platforms. Git(Hub) is *almost* but not really suited for this. It has a softly built in assumption of one "master" branch, which temporarily forks off into PRs just to merge back a bit later. I tried to prototype something super lightweight that could have a flavor of this, e.g. just a Discussion, written by my agent as a summary of its overnight run: https://t.co/tmZeqyDY1W Alternatively, a PR has the benefit of exact commits: https://t.co/CZIbuJIqlk but you'd never want to actually merge it... You'd just want to "adopt" and accumulate branches of commits. But even in this lightweight way, you could ask your agent to first read the Discussions/PRs using GitHub CLI for inspiration, and after its research is done, contribute a little "paper" of findings back. I'm not actually exactly sure what this should look like, but it's a big idea that is more general than just the autoresearch repo specifically. Agents can in principle easily juggle and collaborate on thousands of commits across arbitrary branch structures. Existing abstractions will accumulate stress as intelligence, attention and tenacity cease to be bottlenecks.

GPT-5.4 is really good at spreadsheets; a few finance people have finally said things to me like "huh I guess this AI thing is real"

"What is the hardest question I could ask you that you might get right?"

Wow real range of emotion reading the second and then the the third paragraph.

Very grateful to Jensen for working to expand Nvidia capacity at AWS so much for us!

GPT-5.4 is great at coding, knowledge work, computer use, etc, and it's nice to see how much people are enjoying it. But it's also my favorite model to talk to! We have missed the mark on model personality for awhile, so it feels extra good to be moving in the right direction.

I packaged up the "autoresearch" project into a new self-contained minimal repo if people would like to play over the weekend. It's basically nanochat LLM training core stripped down to a single-GPU, one file version of ~630 lines of code, then: - the human iterates on the prompt (.md) - the AI agent iterates on the training code (.py) The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. In the image, every dot is a complete LLM training run that lasts exactly 5 minutes. The agent works in an autonomous loop on a git feature branch and accumulates git commits to the training script as it finds better settings (of lower validation loss by the end) of the neural network architecture, the optimizer, all the hyperparameters, etc. You can imagine comparing the research progress of different prompts, different agents, etc. https://t.co/YCvOwwjOzF Part code, part sci-fi, and a pinch of psychosis :)

Codex Security—our application security agent—is now in research preview. https://t.co/JG2uwGUJFv

GPT-5.4 Thinking and GPT-5.4 Pro are rolling out now in ChatGPT. GPT-5.4 is also now available in the API and Codex. GPT-5.4 brings our advances in reasoning, coding, and agentic workflows into one frontier model. https://t.co/1hy6xXLAmJ

We're publishing a new evaluation suite and research paper on Chain-of-Thought (CoT) Controllability. We find that GPT-5.4 Thinking shows low ability to obscure its reasoning—suggesting CoT monitoring remains a useful safety tool. https://t.co/isZkNkPXZm

The Adolescence of Technology: an essay on the risks posed by powerful AI to national security, economies and democracy—and how we can defend against them: https://t.co/0phIiJjrmz

Today I met with PM @narendramodi to discuss Anthropic's expansion to India—where Claude Code use is up 5× since June. How India deploys AI across critical sectors like education, healthcare, and agriculture for over a billion people will be essential in shaping the future of AI. https://t.co/rt0baeqEeC

The Urgency of Interpretability: Why it's crucial that we understand how AI models work https://t.co/Mz8R23uxgy

My thoughts on China, export controls and two possible futures https://t.co/LZAANEcZJW

Machines of Loving Grace: my essay on how AI could transform the world for the better https://t.co/CFGd4Sbq86