Demis Hassabis

People usually think that in order to get neurons in a neural network to be selective to a stimulus (e.g. to recognize a particular face), the network has be trained. There's a wild result by Pehlevan and Sompolinsky 2014 (reproduced by me here) that in a *randomly connected* balanced recurrent network, you can actually get stimulus-selective neurons *without any training at all*. Still not sure of the implications.

4

1 Comment

Jotted down some end of year biopharma reflections while waiting to board a flight. Likely no surprises. 2025 was tough, but there's a hint of currents shifting for the better 🤞 1. BIG gulf between what people think scientists are clambering for, when it comes to AI tools, and what scientists are *actually* asking about, using regularly, and willing to pay for. "LLM for this, agent for that" really misses the deeper-rooted challenges and bottlenecks. 2. Lots of killing of discovery, shifting emphasis to later stage assets. Even saw this at orgs where the discovery engine is working (producing clinic stage assets that continue to get good clinical readouts). I understand this reactivity to markets and investor sentiment, but it’s nevertheless sad to see strong scientific engines be shut off, teams laid off, and novel discovery stopped as a result. TBD what longer term effects will be over next several years. 3. It often takes the experience of having gone through a cycle to realize what problems you want to avoid at all costs. Our most well-aligned and motivated champions have been people who directly experienced the painful alternatives ('no action' or 'build-it-yourself'). Conversely, those who haven't had to grapple with the problems before often maintain a "we can just do everything ourselves" stance. 4. A lot of work is being outsourced. When managing these CRO relationships, complexity can balloon quickly. Our partners have increasingly turned to Kaleidoscope.bio to drastically streamline this pain. 5. There is painful disconnect between how much time people waste on preventable stuff, and how much time/budget/awareness leadership will provide to address this. I encourage leaders to empower their team to solve problems that will help them move faster, even if they as a senior leader may not deal with the day-to-day (and thus may not feel it directly). 6. Seems to be an increasing number of scientific PMs spearheading operations (we at Kaleidoscope like this).

58

7 Comments

H/T Yann LeCun Learning Latent Action World Models In The Wild (FAIR at Meta, INRIA & NYU, January 2026) Paper: https://lnkd.in/erBFcHfW Abstract: "Agents capable of reasoning and planning in the real world require the ability of predicting the consequences of their actions. While world models possess this capability, they most often require action labels, that can be complex to obtain at scale. This motivates the learning of latent action models, that can learn an action space from videos alone. Our work addresses the problem of learning latent actions world models on in-the-wild videos, expanding the scope of existing works that focus on simple robotics simulations, video games, or manipulation data. While this allows us to capture richer actions, it also introduces challenges stemming from the video diversity, such as environmental noise, or the lack of a common embodiment across videos. To address some of the challenges, we discuss properties that actions should follow as well as relevant architectural choices and evaluations. We find that continuous, but constrained, latent actions are able to capture the complexity of actions from in-the-wild videos, something that the common vector quantization does not. We for example find that changes in the environment coming from agents, such as humans entering the room, can be transferred across videos. This highlights the capability of learning actions that are specific to in-the-wild videos. In the absence of a common embodiment across videos, we are mainly able to learn latent actions that become localized in space, relative to the camera. Nonetheless, we are able to train a controller that maps known actions to latent ones, allowing us to use latent actions as a universal interface and solve planning tasks with our world model with similar performance as action-conditioned baselines. Our analyses and experiments provide a step towards scaling latent action models to the real world."

11

I'm incredibly proud to announce a productive NeurIPS for our Post-training AI Research team. We had two papers on test-time scaling accepted into the Workshop on Efficient Reasoning, contributing to a fantastic ~6 acceptances for the wider Applied AI Research @ Cerebras. Both of our papers are now available on ArXiv:  - The Conductor and the Engine: A Path Towards Co-Designed Reasoning. Link - https://lnkd.in/g5Dnnz6C. - Calibrated Reasoning: An Explanatory Verifier for Dynamic and Efficient Problem-Solving. Link - https://lnkd.in/gH2jTGeP. In the spirit of collaboration, we are also open-sourcing our updated CePO flow as part of our OptiLLM library on GitHub. This is the same methodology that achieved top scores on the Artificial Analysis Leaderboard. Link - https://lnkd.in/gd6anwFN Congratulations to the team on this incredible milestone: Anisha Garg, David Bick, Engin Tekin, Michael Wang, Pawel Filipczuk, Amaan Dhada, Yash More, Nishit N. and rest of the Post-training Team! Looking Ahead A key enabler for these results and our upcoming work in Coding Agents has been using Reinforcement Learning (RL) to provide LLMs with new, specialized expertise, which makes them highly amenable to test-time compute scaling. This brings me to our next exciting step... 🚀 Announcing Limited Early Access to our RL Service! 🚀 Is your team excited by the potential of powerful LLM models — closed source like GPT-5/Claude/Gemini or Open-source like Qwen3, GPT-OSS — but frustrated when they fail that "last mile" on your specific, critical tasks? We are opening up a limited early access program to help you solve this. Our RL service is designed to transform general models into world-class experts for your unique domain. If you're interested in building an AI system that is an expert at solving your tasks, reach out to me to see if you're a fit for the program. 📧 Email: ganesh.venkatesh@cerebras.net When you reach out, please include: - Subject: "RL Service Early Access" - Body: A brief description of your application and the "last-mile" challenges you're facing. Looking forward to post-training your problems away!

122

1 Comment

DINOv3 by Meta Open Source is the first vision model trained on zero labeled data that outperforms supervised models. I have been tracking this research since DINO v1 (May 2021). In fact some research, I was part of at NYU about Self-supervised state estimations were basically derived from Dino v2. What this means practically: - No more paying for annotation - Train on any image data you have - One model for detection, segmentation, depth estimation NASA JPL is already using it for Mars robots. World Resources Institute cut tree measurement error from 4.1m to 1.2m. For any company sitting on image data they couldn't afford to label, the game just changed. #ComputerVision #AI #MachineLearning #Meta Source: https://lnkd.in/eVWDky5q

6

1 Comment

Our latest technical report shows AI-generated de novo antibodies with *drug-like* properties. Excited to release this... AI is moving so fast. It feels like the models improve in real-time while working at Chai Discovery. These kind of breakthroughs used to take years… now we get each in months 🚀

47

6 Comments