The avalanche that hit math is coming for AI development.

he finally got here 🥹 welcome brother. into the rsi we shall sail.

literally no one: sf guy: i've updated my priors

One of us! One of us! One of us!

If the labs spent as much compute as Navier Stokes on a prompt like "invent a machine learning technique with the same magnitude of impact as the transformer", what would the result be? cc @tszzl

what is the equivalent of a lean proof in AI research? some benchmark? how does this exactly produce super intelligence? this all seems like hopeful thinking hidden in jargon

A large part of the post is detailing the verifiable proxies for AI progress.

We're getting RSI before GTA6

Bro updated his weights

> But as we’ve seen in math, models generalize far beyond what they were directly trained on. How do I reconcile these two?

DWARKESH. JESUS FUCKING CHRIST. You interviewed Noam Brown about this EXACT problem. He told you the 10,000-agent swarm contributed less than 10% to OpenAI's Navier-Stokes result. He admitted they couldn't establish its scaling efficiency. He explained that recursive self-improvement runs into serial experiments, GPU capacity, training time, and delayed feedback. You heard all that and returned with a 44-minute documentary announcing THE AVALANCHE. A mathematical proof has a verifiable terminal condition. Improving an AI model requires a completely different experimental loop. Hypothesis. Implementation. Training. Independent evaluation. Regression detection. Resource accounting. Promotion or rollback. Repeat. We've already implemented bounded versions of that workflow across heterogeneous hardware, with deterministic acceptance gates, retained evidence, and measurable execution costs. That's research automation. Calling it recursive self-improvement requires demonstrating that successive model generations actually improve the research process itself. Show the generation-over-generation capability gains. Compute-normalized improvement. Experiment throughput. Failed candidates. Human interventions. Independent validation. Where does the improvement compound? Where does it saturate? What happens when the experiments cost more than the hypotheses are worth? These are the actual engineering questions. You have direct access to researchers who understand them. Your own guests have explained the constraints. And somehow the output is another exquisitely produced video about the future arriving. You've built a world-class observation deck overlooking an engineering project and mistaken your ability to describe the construction for progress on the building. Forty-four minutes of avalanche forecasting. SHOW US THE FUCKING SNOWPLOW.

Three swarms of civilizations attacked poor Jensen.

why this feels like I'm watching AI talking first 20 seconds lol

You’ve basically made a long form video showcasing how inept you are.. congrats

That analogy to math is sharp, and quietly alarming for AI research.

me running from the avalanche with my old benchmarks

Welcome

math had checkable proofs, so the hard part for ai development is the grader.

It already did lol like 2 times and third is coming too

@Dominic2306 What the hell do I invest in?!

Your footnote 3 is where I'd push. By your own Amdahl numbers, automate 90% of research at 50× and you get ~8.5×, so the 10% nobody can check sets the pace. In math that part came free, since a proof either checks or it doesn't. In ML, choosing which loss to climb was the 2018 breakthrough itself. How big do you think that unchecked share is?

If AI starts improving AI like it cracked math, the bottleneck is compute, not researchers. Who owns the GPUs then?

Ziotarded grifter

if ai starts speeding up its own research the next few years get weird fast

Outside of energy and atoms what are the next/biggest bottlenecks?

Incredible essay, well done

the avalanche is retroactive now, claude is out here fact checking gödel

Such a gifted technical communicator, dwarkesh. Great job

Choosing the next useful experiment could matter even more than writing its code. That’s the part of this feedback loop that makes my head spin

Looking forward to watching this. I think just seeing you start off by acknowledging that you are updating is great. I hope it encourages more people to at least be open to it