Just another day at Anthropic. From https://nothinghuman.substack.com/p/some-memories-of-early-anthropic.

Anthropic's IPO filing is notable for its modesty.

Your chances of becoming a billionaire just by telling your AI to go make money keep getting better.

The money Anthropic’s IPO will raise (~$100B) is a drop in the bucket of their spending commitments ($~1T).

This is what the AI future will really be like. Spread the word. https://contraptions.venkateshrao.com/p/cognitive-mercantilism

As the price of intelligence collapses, demand will skyrocket.

A total addressable market that's far larger than the current one is a double-edged sword: it gives you room to grow, but also means the ra...

The problem with Nick Cammarata's interpretability results is that they're all subjective: I can look at a neuron and imagine it's represent...

Anthropic loses a lot of money, but makes it up on volume.

Anthropic will soon have the most shorted stock in history.

Physicists now have AI envy.

1980s: Lean manufacturing 2020s: Lean mathematics

Mathematicians are a case study in how to not handle the AI transition. https://twitter.com/hesamation/status/2114153199414157752

Sam Altman’s strategy for winning the AI race is to corner the market in compute.

TL;DR: An attention head is a fast-weight MLP.

Some AI jobs apocalypse.

If intelligence is just one algorithm, then when we’ve discovered it all that’s left to do is make more of it. But we haven’t discovered it ...

When the technique X you’ve been pushing doesn’t work and you don’t want to lose face, you do something else and call it X. E.g., X = RL.

Reinforcement learning doesn’t solve any problem that can’t be solved by better generalization.

Yet more evidence that reinforcement learning is basically useless. https://twitter.com/svlevine/status/2112762529701405127