🕳️🐇Into the Rabbit Hull – Part I (Part II tomorrow) An interpretability deep dive into DINOv2, one of vision’s most important foundation models. And today is Part I, buckle up, we're exploring some of its most charming features.

In this new research with @apolloaievals, we found behaviors consistent with scheming in controlled tests across frontier models, including OpenAI o3 and o4-mini, Gemini-2.5-pro, and Claude Opus-4. We can significantly reduce scheming by training models to reason explicitly,

I’m thrilled to share that I’ve joined @OpenAI as a research resident, focusing on safety research and AI for health. After nearly 10 years studying the brain, I’m excited for this next chapter building AGI to accelerate medical progress and scientific research! #feeltheagi

We have data on the environmental impact per AI prompt: Gemini: 0.00024 kWh & 0.26 mL water ChatGPT: 0.0003 kWh & 0.38 mL ...the same energy as one Google search in 2008 & 6 drops of water. Seems to be improving, too: Google reports a 33x drop in energy use per prompt in a year.

New Anthropic research: Persona vectors. Language models sometimes go haywire and slip into weird and unsettling personas. Why? In a new paper, we find “persona vectors"—neural activity patterns controlling traits like evil, sycophancy, or hallucination.

📣 Excited to share our real-world study of an LLM clinical copilot, a collab between @OpenAI and @PendaHealth. Across 39,849 live patient visits, clinicians with AI had a 16% relative reduction in diagnostic errors and a 13% reduction in treatment errors vs. those without. 🧵

I’ve always loved how Japanese food looks so so good in anime movies. So I tried recreating some of the Indian street food in Ghibli style — and I’m absolutely loving the results 🫶

In the physical world, almost all information is transmitted through traveling waves -- why should it be any different in your neural network? Super excited to share recent work with the brilliant @mozesjacobs: "Traveling Waves Integrate Spatial Information Through Time" 1/14

New brain-computer interface preprint led by PhD student Tyler Singer-Clark! This video of a man with paralysis accurately controlling a cursor looks like something you've seen since ~2017. BUT! This is driven by multielectrode arrays in ventral (speech) motor cortex‼️ 1/

New result: The first iPSC-derived corneal stem cells (iCEPS) reverse vision loss The novelty here is that the cells were derived from iPSCs, which did not come from the patients (allogenic). Summary: A Japanese team reported yesterday that their iPSCs-derived corneal stem

Super pumped to announce that our new work “Face cells encode objects parts more than facial configuration of illusory faces” with @kasper_vinken, @akjags and Marge Livingstone is now out in @NatureComms!

Preprint: It's been suggested that Aphants may be able to visualise, but lack insight, as they can do Mental Rotation (MR) tasks. Instead, we show MR tasks are a weak measure of the propensity to visualise. https://t.co/ZMWeimlaEp

Is there a name for the hypothesis that task-trained neural networks and biological brains develop similar representations when they perform well, despite differences in architecture and learning methods? I.e. that optimization (whether backprop or evolution) leads to a similar

The *action* space is what matters. Successful actions can tell you, compositionally, in situ insights about the system to higher and higher order. Build your models around actions - not measurements.

Sounds like they’re going straight for cortical stimulation… very curious how well this will work, given everything we know about the complexities of the cortex.

Why do varied DNN designs yield equally good models of human vision? Our preprint with @michaelfbonner shows that diverse DNNs represent images with a shared set of latent dimensions, and these shared dimensions turn out to also be the most brain-aligned. https://t.co/vtOOYHQb47

New study @CellReports🥳 Using recordings & #deepnet driven image synthesis we find an intriguing difference btw🐭&🐒 visual cortex/V1 🐭: complex, non-Gabor like receptive fields 🐒: Gabor-like RFs Team effort🙌@jiakunfu @AToliasLab @sinzlab & more👇 https://t.co/RSwCIea4nQ

Our new study is out today in the New England Journal of Medicine! We demonstrate a speech neuroprosthesis that decodes the attempted speech of a man with ALS into text with 97.5% accuracy, enabling him to communicate with his family, friends, and colleagues in his own home. 1/9

"People who regularly eat processed red meat, like hot dogs, bacon, sausage, salami and bologna, have a greater risk of developing dementia later in life."

yearly reminder everything looks exponential from the middle of a sigmoid