Ever wondered why presenting more facts can sometimes *worsen* disagreements, even among rational people? 🤔 It turns out, Bayesian reasoning has some surprising answers - no cognitive biases needed! Let's explore this fascinating paradox quickly ☺️

I think acquiring compute is a great strategy atm

Tinker's new StableReinforce objective is Reinforce + Score Centering. Here's how I reverse-engineered it 🧵

🥐 Hi everyone! FlashAttention's backward pass has the same zero-sum bug @jon_barron just fixed in softmax. With FP4 stochastic rounding it's wrong on every single draw. We fixed it in our ICLR'27 submission, CroissAttn and unlock e2e FP4 training.

The non-disparagement clauses in Google's UK employment contracts are indeed incredibly broad. This is from the Confidentiality and Proprietary Information Agreement I signed in 2024 before I started. It forbids any criticism forever (regardless if it's true or not)

A first (short & compute-constrained) research project after leaving Google DeepMind: an LSTM-style tied transformer with one shared core KV cache The main question is how to train for the cache histories that the model produces recursively at inference

Really cool (and surprising) that there are still so many ways in which we can improve "basic" operations in ML 👏👏👏

Kimi should have called it 3.5 Pro 🥶

I think people should be more outraged about this It seems Google saw the Right to Warn letter about AI and the congressional investigation into OpenAI's practices and simply shrugged it off without changes Why should anyone have any trust and confidence in Google in the future? The stakes for AGI and ASI are too high and the question is: how bad was the chilling effect, and what else is there that we don't know about?

Based on this article, OpenAI has made false accusations about me. The three researchers fired from OpenAI were told I facilitated architecture leaks to The Information. This is false. A journalist from The Information did reach out to me shortly before that article came out. I did not comment at all beyond asking the journalist to make it clear that a blog post I wrote which was referenced in the article was preexisting and public. The article also mentions an internal OpenAI monitorability proposal. I have heard that this memo contained highly sensitive architecture information, and that OpenAI accused me of accessing and reading it. This accusation is also false. As reported in the article, I viewed a version of the proposal that did not have sensitive IP and was approved for sharing with me. I believe OpenAI now knows this accusation was based on a conflation of these two documents, though I do not know if they have clearly retracted that accusation. I do not and have never solicited proprietary information, and the three employees who were recently fired from OpenAI never leaked IP to me. As the article mentions, I did have Signal chats with them that had disappearing messages on; they’re professional contacts that I’ve had on Signal for a while, and use of Signal is common in the tech industry. During the Hugging Face investigation, I was careful to avoid sending Signal messages with confidential information, though I did sometimes discuss high level things on Signal calls. As far as I understood, this was normal.

Introducing Reinforcement Learning from Collective Rewards (RLCR). RLCR is a recipe that trains language models as agent organisations, which share a message board and workspace. We use it to train Avalon-M1-Mini (120B) and study its behaviour. We find a number of interesting results, including: 🛠️ Learnt organisational hierarchies and divisions of labour 🕵️ Increased collective verification, through hires of verifiers and testers 👥 The rise of collective thinking, where agents think more about other agents and how to help them We believe RLCR can provide a laboratory for studying collective intelligence and multi-agent safety. Blog post and technical report 👇

AI mode and chatbots falling victim to SEO and disinformation is a huge (to be expected) issue Disformation becomes laundered through LLMs as the models might be too trusting with external information We trust the models answers bc they are usually quite useful already and end up being manipulated

Disappearing messages aren't smoke; lots of people use them with their friends and the AI safety community is well-connected in social circles. Did they also check their Snapchat? If they fired three of their best AI safety researchers for this, OpenAI is just dumb The real risk is the chilling effect this is intended to have Also more good safety researchers might decide to just leave OpenAI instead of getting smeared in public

New on our Frontier Red Team blog: using Claude to formally verify software at large scale. Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time). With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks. Read our research in full here: anthropic.com/research/forma…

I hope Google will pay up and make good somehow. It sounds like they moved too fast without knowing local laws. Not sure how that can happen. Not great when people are already against datacenters. This is a gift for anyone who doesn't want more compute sovereignty in Europe 🥺

Concerning because of the possible lack of transparency. Especially after OpenAI fired three AI safety researchers, which might have a strong chilling effect on speech. Auditors should be free to write about their findings

"On the recently proposed o(n log n) integer multiplication algorithm" by David Harvey web.maths.unsw.edu.au/~david…

The US government under Trump has become so incredibly corrupt(ed). All Trump enablers who willfully broke the law hoping for a general pardon in the future need to face real consequences and be banned from public office. They're clearly not fit for it The US has turned into a banana republic

A little additional context, as this is getting attention: (1) These clauses are pretty common in tech/corporate contexts, to the extent that HR for big firms like Google can be really taken by surprise if people make a fuss about them. (2) They're buried in very long contracts - several people I spoke to only discovered they were under in perpetuity non-disparagement after I encouraged them to go back to their contracts and go keyword searching. (3) Freedom to express concerns is critical in a fast-moving, ethically charged area like AI risk. DM leadership is aware of this. It has been a core tenet of much of the governance material produced this decade (e.g. see arxiv.org/abs/2004.07213). (4) The autonomous weapons/surveillance thing is particularly galling in a DeepMind context. Hassabis and co profited a LOT over the years from stating principles against these uses clearly. It gained them academic and civil society support*, and staff who have contributed a lot to their success joined and stayed in part because of those principles. As the BBC article says, it's this old guard who are most upset about what is happening. Hassabis is of course welcome to change his principles (last year he said 'the world is becoming a more dangerous place') but to silence those who still hold to DeepMind's original principles, or punish those who speak up (and both these have been happening) is IMO abhorrent. (5) It's unclear how strictly enforced this all is. For current employees, there is clearly *robust* comms oversight and message control. For ex-employees, it seems much laxer. But the uncertainty is enough for it to have a chilling effect. If you were just out from your role at GDM and looking for your next role, would you really want the risk of dealing with Google's giant legal team, or finding yourself industry-blackballed in some way? (6) OpenAI and Anthropic both had clauses like these, and removed them in the face of public and employee pressure. It'll be a harder task with GDM, as it's a much bigger and more opaque organisation. But it's worth making our feelings known - the stakes are too high.

Thanks for saying this and I hope you can put external pressure on Google to be better and not have such clauses in the future The public might have been misled in how open Google is when it comes to allowing valid criticism The fact there are such chilling effects (and potentially others) should give everyone goosebumps The industry simply cannot be allowed to regulate itself as long as it follows such practices

Important, and I'm glad this is finally public. Google DeepMind researchers are under in perpetuity non-disparagement clauses forbidding them from ever speaking critically of the company. I'd already confirmed this with multiple people. Yes, some ex-employees have spoken out nonetheless. But it has also discouraged others from doing so. And yes, it most certainly has a chilling effect with regards to discussions on safety and risk. Others have written compellingly about how important the ability to speak freely and critically is, especially on topics like safety and risk. And what a testament to the research employee community's coordination and hard-won bargaining power it has been to get there at other companies. GDM burn that commons by having this. And everyone who signs burns that commons too. Let's get it changed.