Latest / Lenny’s AI Builders / OpenAI's new internal model is getting ridiculous
Transcript
- 0:00They've done it, OpenAI, they are sharing a solution to the Navier-Stokes Millennium
- 0:04Prize problem, one of the deepest problems at the frontier of mathematics.
- 0:08In this video, I will be showing you exactly what on earth that actually means, why this is
- 0:13a big deal, and whether this is the end of AI advancement or just the very beginning.
- 0:18Before we do that, I'm Lenny, I love making videos about AI, it is currently 1.54am in
- 0:24the morning.
- 0:25If you want more videos about AI from an Aussie dude with a Mo, subscribe.
- 0:29So back to the news at hand, if you have been living under a rock, GPT-6 Astra came
- 0:34out about a week ago now, and it is the best AI model in the world.
- 0:39What we're talking about now is a next-generation model significantly more capable than the
- 0:44best in the world right now, which is already smarter than most people on Earth.
- 0:48The problem at hand concerns whether the description of smooth three-dimensional fluid motion
- 0:54modeled by the Navier-Stokes equations can break down.
- 0:57It has remained unsolved for roughly 90 years.
- 1:01Here is the part that blew me away.
- 1:02This model represents a step function improvement on many benchmarks, and its training is ongoing.
- 1:07But here it is.
- 1:08Our internal model group arrived at the Navier-Stokes solution in 88 hours using around 10,000 coordinating
- 1:16AI agents.
- 1:17Here's a graph, performance of GPT-6 Astra and our internal model on a curated set of
- 1:22open math problems.
- 1:23We see two lines on a line graph.
- 1:25One is GPT-6 Astra, one is this unnamed internal model.
- 1:30GPT-6 Astra is basically significantly lower than the internal model, continuing on with
- 1:35the post.
- 1:36The solution, which is the thing that was just discovered, a spinning swirl of fluid that
- 1:42spirals inward and gets increasingly elongated like spaghetti.
- 1:46Then there's a link to the official paper titled finite time blow up for Navier-Stokes.
- 1:51And this is the vortex that was just referred to.
- 1:53I'll discuss exactly what all of this means in just a second, but we've just got a couple
- 1:57more things to take a look at here in the official blog post.
- 2:00Agents were subdivided into groups with the ability to communicate within the group.
- 2:04We encourage different groups of agents to explore a diversity of approaches.
- 2:08After some time, we cross-pollinated agent groups by using Codex to consolidate the
- 2:12most useful insights from each agent group.
- 2:15These follow-up prompts drew on the agent's own intermediate results, the group that found
- 2:19the solution to Navier-Stokes was guided in such a way.
- 2:21The agents arrived at their resolution on Saturday, September 5, about 88 hours after
- 2:26the first agents were launched.
- 2:28So that was done by the internal model and then lean formalization and verification of
- 2:33these results took an additional 17 hours via GPT-6 Astra, which is currently publicly
- 2:39available.
- 2:40Here's a cool part.
- 2:41Across all attempted problems, the agent sent 4.9 million messages and used about 300
- 2:47billion output tokens.
- 2:50And in the process of resolving the Navier-Stokes problem, the agent sent 2.7 million messages
- 2:56and used approximately 130 billion output tokens.
- 2:59So the way AI is billed is through tokens.
- 3:03GPT-6 Astra is currently the most expensive AI model offered by OpenAI.
- 3:10This is an internal model.
- 3:11Assuming this internal model was publicly available, which it isn't, and was priced at the same
- 3:16price as GPT-6 Astra.
- 3:19That means across all attempted problems, solving the Navier-Stokes Millennium Prize
- 3:24problem cost 15 million US dollars.
- 3:28And in the process of resolving the Navier-Stokes problem alone, the commercial cost was 6.5
- 3:35million US dollars.
- 3:36Our goal in releasing this result is to report on the substantial progress of AI models.
- 3:40We do not intend to claim the Millennium Prize for this result.
- 3:44Very interesting.
- 3:45Noble, I would say.
- 3:46This is not a culmination, but rather a snapshot in time of progress on AI development.
- 3:50We believe we are now in the next period of AI progress, and today's results provide
- 3:55further evidence of this.
- 3:56What is the significance of this?
- 3:57We're going to get to that in just a second.
- 3:59But what the end of that blog post is talking about is just basically saying AI is getting
- 4:04really fucking smart, really fucking quickly.
- 4:07Just to put this into perspective, GPT-5 came out 13 months ago.
- 4:11The chart that we're looking at now, think of it as a good rule of thumb for intelligence.
- 4:17Higher is better.
- 4:18One year ago at its peak was 23 out of 100.
- 4:2213 months later, it's now at 53 out of 100.
- 4:26And most of those gains came in the last three months.
- 4:30And that's only going to continue to accelerate and to get faster.
- 4:33Like we're just at the beginning of what people call the exponential, where things are going
- 4:37to get crazy really quickly.
- 4:39Anyway, to better understand this, who better to help us, or what better to help us, than
- 4:44GPT-6 Astra itself.
- 4:46Explain to me in super simple terms what the Navier-Stokes Millennium Prize problem is.
- 4:52It's a challenge about the equations we use to describe how fluids move, like air or water.
- 4:57The big question is, do these equations always behave nicely, or can they sometimes blow
- 5:03up and produce impossibly wild results?
- 5:05In simple terms, it's asking whether smooth fluid motion stays smooth or can break down.
- 5:11Explain it even simpler.
- 5:12Sure.
- 5:13It's asking if you start with water flowing smoothly, will it always keep making sense?
- 5:18Or could the math suddenly go crazy and stop working?
- 5:21Simpler again.
- 5:22It's asking, will the math for flowing water ever break?
- 5:26What's the significance of solving it?
- 5:27Give me an example, put this into perspective.
- 5:30Think of a flight simulator.
- 5:31It uses maths to predict how air moves around a plane.
- 5:34This problem asks, can that math hit a point where it stops giving a sensible answer?
- 5:40Even if everything starts smoothly, OpenAI says it found a carefully constructed case
- 5:44where that happens.
- 5:46If that holds up, it settles a question mathematicians have wrestled with for about 90
- 5:50years, and reveals a fundamental limit and a hugely important tool.
- 5:54It's not about real water going infinitely fast, and it doesn't mean planes or weather
- 5:59forecasts don't work.
- 6:01The AI angle is also big, helping discover new maths, not just answering old questions.
- 6:06What would you say are the five most interesting or important things to know?
- 6:10One, they say they also solved a separate major maths problem on the way involving the
- 6:16Euler equations, which helped point them in the right direction.
- 6:19Two, in their example, the water starts completely still, and a smooth external force gets it
- 6:25moving, so it's not starting broken.
- 6:27Three, the AI was still being trained while this was happening.
- 6:31They upgraded the agents midway through the project.
- 6:34Four, another team had a related result, but not the same result.
- 6:39Those details matter.
- 6:40Five, OpenAI says it won't claim the million dollar prize.
- 6:44They frame this as showing how far the AI has come.
- 6:47Cool, so I didn't actually know that it was a million dollar prize, but yeah, OpenAI
- 6:51probably doesn't need any more money.
- 6:53So I heard whispers that Claude and Anthropic had solved the Navier Stokes Millennium prize
- 6:58problem on X or Twitter.
- 7:00Wouldn't be surprised if OpenAI got wind of this.
- 7:03They thought, fuck you, Claude, deployed 10,000 agents at the cost of around $15 million,
- 7:09and then just solved the thing in a few days.
- 7:12OpenAI is so back.
- 7:13Imagine AI teams exploring scientific ideas and helping finding something genuinely new.
- 7:18I think that is the future.
- 7:19The headline from this is they deployed 10,000 super intelligent AI agents to solve this
- 7:25thing.
- 7:26They just pointed compute at it, spent $15 million and then just broke frontier math.
- 7:33If that doesn't tell you that we have just entered the AGI era, I don't know what will.
- 7:37Let me know in the comments below.
- 7:39Do you think this is the future of scientific discovery, deploying ginormous swarms of super
- 7:44intelligent AI agents at some unknown problem and getting them to just work 24 seven around
- 7:51the clock, thinking about multiple different solutions, like Dr. Strange looks at 13,000
- 7:56different possibilities.
- 7:58That's what these AI agents are doing to solve new problems.
- 8:02Do you think that is the future?
- 8:04Let me know down below.
- 8:05If you want more videos about AI from an Aussie dude with a Mo, subscribe, peace.