Hanson Wen

New Blog Post! AI in Clinical Trials: Stop Talking About “Replacement” open.substack.com/pub/quixoticme…

Computer use is now in Claude Code. Claude can open your apps, click through your UI, and test what it built, right from the CLI. Now in research preview on Pro and Max plans.

The western blot is a flat circle

@bitforth What LLM image model can infer brain activation images to that level of detail? I’m highly skeptical

Implementé TRIBE v2 localmente en mi laptop, le pasé un video familiar, y luego le di a un LLM las imágenes de activación cerebral que generó, sin mucho contexto, y esto fue lo que dijo: "El patrón muestra actividad fuerte en áreas visuales y de reconocimiento, con poca participación de regiones frontales asociadas al pensamiento complejo. La inferencia es que el video muestra algo en movimiento, probablemente personas o caras, algo fácil de procesar y no muy informativo; está hecho para captar y sostener atención de forma pasiva, no para hacerte pensar." El clip en cuestión era un video de 15 segundos de mi hija manejando un scooter. Ahora, aquí es donde se pone interesante, mi esposa tiene una cuenta de TikTok que va creciendo. Voy a tomar 50 videos virales del nicho y los de ella, correrlos por TRIBE v2 y alinearlos con sus curvas de retención para ver qué cambia en el cerebro promedio unos segundos antes de que la gente haga scroll. De esta forma, en lugar de editar para retener atención basado en intuición y A/B testing, edita para mantener activación en las regiones correctas en los momentos correctos.

@petergostev I think the wild unclaimed results used caching which is submitted for another kind of non record track

Wild things are happening in OpenAI's Parameter Golf competition: best verified result is 1.09x the baseline, best claimed but unverified result is 42.7x better

The TurboQuant paper (ICLR 2026) contains serious issues in how it describes RaBitQ, including incorrect technical claims and misleading theory/experiment comparisons. We flagged these issues to the authors before submission. They acknowledged them, but chose not to fix them. The paper was later accepted and widely promoted by Google, reaching tens of millions of views. We’re speaking up now because once a misleading narrative spreads, it becomes much harder to correct. We’ve written a public comment on openreview (openreview.net/forum?id=tO3AS…). We would greatly appreciate your attention and help in sharing it.

@jpshipped @SimonTheWang friend.com would

@SimonTheWang in all fairness, no normal person is buying a domain for $1.65M

Making an offer they can refuse

@daviddorg This thing! x.com/AIatMeta/statu…

@_hansonw What’s this… Looking

Today I am scanning my brain before I spend a year without screens Doctors expect brain function to change, but whether the structure itself changes, nobody seems to know All the data will be public: baseline, 6-month, and 12-month scans included

Foundation model this foundational model that when will we get a foundation model for a foundation model

This is like letting agent swarms run overnight but so much cooler

@adibvafa I don’t think they collect the approval data right? It’s more of just a llm powered rule based decision engine based on what I’ve read

> Claude asks for permissions > Human approves or not > Claude collects data > Anthropic trains classifier on that data > Classifier decides on permissions now Soon Claude will learn to prompt itself too Then what is left for human?

They cloned a mouse and kept cloning. It went 58 generations before it just couldn't "This model predicts that in asexual lineages, deleterious mutations inevitably accumulate, ultimately producing mutational meltdown and extinction."

@rllm_project Yes bro pleaaase 🙏🙏Who should I reach out to lmao I just finalized a prototype for my swarm on GitHub

@_hansonw lol we can collab on something together

When I’m building my own agent swarm and I see this

The best prompt to quickly learn enough to have a conversation with any expert of any field: Look at everything I’ve studied so far—notes, questions, summaries, and all related material—and reconstruct the subject as a top-down concept tree. Show which concepts are fundamental, which are built on top of others, which are subsets or special cases of broader ideas, and what purpose each concept serves. Make clear which lower-level concepts exist to support higher-level ones, and which concepts are applications versus core foundations. Write a long, comprehensive Markdown report as a numbered hierarchical outline, traversed depth-first. Focus only on the essential conceptual structure. Ignore notation, logistics, and surface-level details.

Proteins can now talk. Introducing BioReason-Pro, the first reasoning model for protein function. A thread🧵

Closest thing we have to agi health rn. imagine using it to plan schedules around your sleep chronotype and purchase custom supplements / peptides etc

@r0b0t_sp1der @henloitsjoyce Damn I’ve never seen one ever

@henloitsjoyce They walk up and down the train continually. Should be only a few minutes to see one.

cant even ride the caltrain safely anymore :( a guy got into the seat next to me, i didnt take notice, then the stench hit me i realised he might have broke out of the hospital, he still had his hospital wristbands on. and he kept pouring powder from a tube into his hands and sniffing hard then sneezing it all over then he tipped over to my side and his hand grabbed onto my seat narrowly missing my thigh didnt know what button to press or who to call, everyone around just watched on and this man was clearly not okay what is going on

@AI_Andrew @petergyang it's equivalent to the halting problem.