Sacha Raffaud

New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.

He's training a two layer CNN on MNIST. Set dropout to 0.95. Good, now set the learning rate to 10^-8 and remove all of the 7's from the train set.

is it just me or has claude fable 5 gotten noticeably worse?

In the quest toward automated synthesis, there's a massive gap between "this reaction looks good in paper" and actually running it. Turns out, someone has to select the conditions: temperature, ligands, additives... the right conditions turn an impossible reaction into real product! Today we’re announcing Palladium, our agentic solution for reaction condition and HTE plate design. Instead of starting with blind optimization, Palladium leverages precedent experiments and substrate-specific reasoning to design a strong first plate. Our beta users in academia and pharma have found conditions as high as 98%, and typically above 90% yield, on the first experiment, often with no follow-up experiments needed. This is how at b12 Labs (YC S25) we accelerate real chemistry. Read more b12-labs.com/blog/palladium… If you’re a chemist trying to accelerate difficult reactions, we’d love to talk.

@b12Labs @FutureLabs_Live @ZJoncev @drecmb Super cool! Wish I could join :)

The @b12Labs team is joining @FutureLabs_Live next week in Basel! Meet our founders @ZJoncev and @drecmb at Booth S9. At @b12Labs (YC S25) we believe the future of Synthetic Chemistry is Agentic, and we are building that future. If you're around, let's meet and discuss chemistry and AI: we'll be excited to share our latest breakthroughs in retrosynthesis, reaction optimization, and integration with robotic labs! We’d love to connect with partners and chemists who are ready to embrace the agentic future of the lab. Let’s talk about how we can optimize your reactions and automate your workflows. #FutureLabs #AgenticChemistry #AI4Science #Chemistry #Synthesis

hello world 👀 🤖⚗️

New Anthropic research: Natural Language Autoencoders. Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read. Here, we train Claude to translate its activations into human-readable text.

When my brilliant new attention method violates causal masking

@PropheticAI New ad space unlocked

my claude explore agent that hasn’t responded in 47m

We are pleased to share that we have successfully demonstrated the ability to increase lucidity in dreams. This critical milestone opens the door for our devices to ship soon.

French gov moving faster than most companies we now have a skill to do taxes, what a time to be alive

@samsinai @Dyno_Tx @AnthropicAI how long did it run for?

We @Dyno_Tx gave Claude Mythos Preview our take home interview challenge in collaboration with @AnthropicAI. It performed on par with the best humans we’ve seen since 2019, many of whom went on to found and lead at top AIxBio companies. What does it mean for the future? Read more 👇

@JamesTThorn List every book on the bookshelf 😀 good recall?

It still feels like magic that you can point Gemini to a bookshelf and prompt for recommendations g.co/gemini/share/0…

@Yuchenj_UW 10,000B to train

Anthropic’s new model, Capybara: “Compared to Claude Opus 4.6, Capybara achieves dramatically higher scores in software coding, academic reasoning, and cybersecurity.” According to Dario's previous interview, it might be a 10T-parameter model that cost $10 billion to train.

2026 may be the year AI starts to truly reason about biology. AlphaFold helped close the sequence → structure gap. The next frontier is sequence → functions. Today, together with @genophoria and the team at @arcinstitute , we’re releasing BioReason-Pro — the first multimodal reasoning model for protein function prediction.

This is wild. theaustralian.com.au/business/techn…

1/ Transformers don't count like computers. We assume they have hidden "registers" to track variables. We were wrong. New research by @AnthropicAI reverse-engineered Claude 3.5 Haiku and found it works with 6D helical manifolds. It's geometry, not math. 🧵

are models using the leak though?

Here is a game: ask an AI to generate an image. Ask a fresh model to describe it. Use that description to generate a new image. Repeat and see what happens. The results are strange and unsettling. To start: "An apple sits on a table." 1/N