Beam is hiring a BizOps leader to join the team in SF! We’re looking for a generalist with strength in GTM to help incubate critical functions and drive strategic product initiatives. Read more and apply here: bit.ly/3CByG3G

My company Beam is hiring a 4th engineer. You might like our team if you - notice every little thing slowing down eng and get annoyed - think construction is at least kind of interesting - have played factorio (optional) DM if interested!

Surveillance state for driving 👍

Kicking off a training job to come back to an OOM error is top 3 anime betrayal.

all i want for Christmas is gpu compute

After growing revenue 15x, TPV 5x, and customers 5x this past year, Beam is now hiring five open roles: trybeam.com/jobs

so much of RL is adding torch.clip in the right places

(re)learning RL, going to speedrun an implementation of alphazero for connect 4 starting now

Baba is you would be a good reasoning benchmark for ai

Why do people refer to papers by <last name> <year>? It tells me nothing about the content at a cursory glance.

It would be cool if claude code had LSP access to files I add in context.

Our new paper: Emergent misalignment extends to *reasoning* LLMs. Training on narrow harmful tasks causes broad misalignment. Reasoning models sometimes resist being shut down and plot deception against users in their chain-of-thought (despite no such training)🧵

If you have to create pop up to get me to use a new feature, I'm not going to use it.

New paper & surprising result. LLMs transmit traits to other models via hidden signals in data. Datasets consisting only of 3-digit numbers can transmit a love for owls, or evil tendencies. 🧵

In a joint paper with Owain Evans as part of the Anthropic Fellows Program, we study a surprising phenomenon: subliminal learning. Language models can transmit their traits to other models, even in what appears to be meaningless data. x.com/OwainEvans_UK/…

Subliminal learning: training on model-generated data can transmit traits of that model, even if the data is unrelated. Think: "You can learn physics by watching Einstein do yoga" I'll discuss how this introduces a surprising pitfall for AI developers 🧵x.com/OwainEvans_UK/…

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.