expectedwrong hindsight

A Chinese Lab Just Nuked the Moat

DeepSeek R1 dropped four weeks ago and the vibes have not recovered.

2 min read 221 words #ai #deepseek #llm #compute #open-source
hindsight — still happening

The moat was nuked. The "haha" of watching scale-is-destiny people explain why efficiency proves their point is still playing out. Compute was not the moat. Careful thinking was.

Four weeks ago a Chinese lab released a reasoning model that matches o1, costs almost nothing to run, and is fully open. Nvidia lost six hundred billion dollars in market cap in a single day. The stock recovered. The implications did not.

The thing that's hard to sit with — and this is the "oh dang" part — is that the entire Western AI investment thesis was predicated on compute being the moat. You need the clusters. You need the chips. You need the billions. Turns out you need a team of researchers willing to think carefully about efficiency, which is a different resource entirely and one that doesn't show up on a balance sheet.

The "haha" part is watching the people who spent three years explaining why scale was destiny now explaining why this, actually, proves their point about scale — just in a different direction — if you think about it correctly — from a certain angle —

They trained it on synthetic data generated by a bigger model. The bigger model is American. So in some sense we're here.

We are genuinely at a moment where nobody knows what the next four weeks look like, let alone four years. That's not a complaint. It's just the thing that's true right now, in February 2025, before whatever happens next.