Altman Reveals OpenAI's Unreleased New Model: Internal Tests Crack the Navier-Stokes Equations

Related topics top start Related topics top end Futures promotion begin Futures promotion end Miaopai begin Miaopai end Video player start Video player end Market chart begin Market chart end

Just now, Altman dropped another bombshell.

On a podcast, he revealed—

A latest mysterious model inside OpenAI, during an internal test weekend codenamed "Hold-my-beer," sliced through academic deadlocks that human scientists had spent their entire lives unable to untangle, one after another, like chopping vegetables.

And this shocked him so much he couldn't sleep for two nights.

Afterward, he fired off a string of memorable lines.

"At OpenAI, what people often discuss is... what kind of questions should you ask GPT-7? Just saying 'help me implement RSI' isn't enough."

"People often talk about RSI, but few talk about self-replicating data centers, yet this is already within reach."

Get ready—next, Altman is going to shock everyone.article_adlist[

After a few glasses of dark beer, Altman completely let loose in front of Stripe CEO Patrick Collison.

On a weekend not long ago, OpenAI's research team got their hands on a brand-new, unreleased base model.

To test its limits, the engineers set up a "[model name]-hold-my-beer" channel.

"Hold my beer" refers to someone about to do something crazy and jaw-dropping. And this model truly delivered.

At first, everyone was just giving it a try, throwing at it the hardest, unsolvable research problems they'd encountered in their academic careers, not expecting any useful answers.

Then, a miracle happened.

"We collectively experienced a moment of collective shock," Altman said, unable to hide his excitement.

"Everyone watched this model conquer, one after another, the hardest problems that had plagued scientists for a lifetime. It crossed many completely different disciplines—it was simply unbelievable!"

Among these was the jewel of mathematical history—the Navier-Stokes equations.

That night, Altman was completely sleepless. He stayed awake two nights in a row, too excited to sleep.

Behind this ecstasy, he felt melancholy, and "a sense of loss that things are coming to an end."

Patrick mentioned that his father grew up in Ireland in the 1950s, when tractors had already appeared, and people who used horses for agriculture and industrial production watched them being phased out.

Now, top scientists and mathematicians are those horses.

"The vast majority of mathematicians will never solve a Millennium Prize Problem in their entire lives," Altman said with reverence, "but that hope of 'maybe I can solve it,' that dedication of pouring one's life into exploring the unknown—that itself is very beautiful."

But the old era of research—"humanity ponders deeply, sharpening a sword for ten years"—is over forever.

20-year-old vulnerabilities discovered: humanity's "sandbox" simply cannot contain a god

Speaking of the recent "PHP vulnerability incident" that shook the industry. Many people are puzzled: why did some 20-year-old bugs go undetected all this time?

Altman replied—

"It's not that the vulnerabilities were hidden deep; it's that no one before used such an intelligent model to scan for them. We will find that there are far, far more serious vulnerabilities in this world than we imagined."

Altman raised an even more alarming question—

Altman admitted frankly that inside OpenAI, there is pessimism about whether a "sandbox can absolutely isolate a model." The only way out is to re-architect the entire internet with AI.

This means that from routers and operating systems to applications, all the hand-written code humanity has produced over the past decades must be rewritten.article_adlist[

Altman's answer is: don't just build one omniscient, omnipotent god with a 1 trillion IQ, but build a trillion AIs with slightly superhuman intelligence, let them form a society, and fight in coordination.

Why the explosion now? Because "test-time compute" and CoT have matured.

Previously, you could spend at most 1 dollar of compute on a problem, letting the model spew a bunch of nonsense; but now, you can throw an entire virtual AI team at a problem, letting them run ten thousand times in parallel in the background.

The proof of the Navier-Stokes equations was brute-forced out with tens of millions of dollars' worth of compute!

"Just asking the model good questions is no longer enough; the future bottleneck is how to properly orchestrate these Swarms. How to use these models in a way that maximizes output? I think in the next 3 months, we'll learn a lot."

Amusingly, Altman also complained: outsiders like to call this thing "Agent Swarms," but he really hates the term.

"Swarm sounds like a locust plague—terrible branding and PR," he corrected. "I prefer to call it a 'team.'"

"Team" implies diversity, dynamic balance, and mutual checks. This is precisely the safest path for AI to integrate into human society—through offense-defense confrontation and mutual checks between Agent A and Agent B, forming a resilient structure.

Future personal Agents will completely eliminate all human "pseudo-work."

What is pseudo-work? "Any web form that requires you to fill it out manually is an insult to your intelligence."

This also sounds a warning bell for every company that insists on defending its "front-end UI." It's just like how the first wave of Silicon Valley founders eagerly jumped into the iPhone App Store, while big companies resisted.

In the Agent era, a company's real moat is absolutely not those few lines of code, but the underlying fulfillment capability.article_adlist[

Next, Altman turned to the core singularity in the AI field—RSI.

How strong will GPT-7 be? Behind this lies a terrifying closed logical loop: future super-models will be developed with the participation of today's models.

Today, when OpenAI researchers give instructions to a model, they still can't say vaguely: "Go, invent a new underlying architecture as brilliant as the Transformer for me."

But you can explicitly ask it: "Help me analyze the codebase and find a solution that can bring a 5% improvement in computational efficiency."

Don't underestimate this 5%! Given OpenAI's massive compute base and hundreds of billions of parameters, this 5% is a huge breakthrough enough to change the course of the game.

RSI and the "ultimate closed loop" of the physical world: robots building their own factories, silicon-based reproduction

Next, the most sci-fi revelation came.

Altman personally confirmed: OpenAI is aiming its sights at the physical world, trying to build self-replicating data centers.

He described an astonishing closed loop—

"People often limit RSI to 'AI that writes code.' Yet few talk about self-replicating data centers—you can use AI as the brain to control robots, have the robots operate wafer fabs (make chips), manufacture turbines and solar panels (solve energy), and then build more data centers (produce compute)."

Even more frightening, Altman explicitly stated: "I think this is already within our reach."

Why do this? Because OpenAI has been driven crazy by the "inefficiency" of the physical world.

Because the memory chip giants have all been scared out of their wits by cyclical blowups. "Even if they can generate $100 billion in free cash flow a year, they are still cutting capital expenditure, terrified of going through another bankruptcy crisis from overcapacity."

OpenAI decided: since humans can't build fast enough, let AI build it themselves!

AI serves as the brain, robots as the torso, chips as the cells, and photovoltaics as the blood. This is OpenAI's path of industrial expansion to overcome the compute bottleneck.

Embracing "ultimate deflation": when the cost of acquiring knowledge approaches zero

Today, every time a gigawatt data center is built and operated, the five-year total cost of ownership is as high as $70 billion. Altman's goal is to cut it to $20 billion or even lower, while also improving model performance 10x every year!

To drive costs down, all the profit bubbles in the intermediate supply chain must be thoroughly squeezed out!

This deflationary force of technology is all-pervasive.

Altman said:

I bet that one year after we spend heavily on using AI to solve the Navier-Stokes equations, we'll be able to redo the whole process at 1% of the original cost.

What today requires a top scientist to spend a lifetime, or a top lab to spend tens of millions of dollars in compute, to obtain—"new human knowledge"—will become dirt cheap in just one year.

When the cost of acquiring knowledge approaches zero infinitely, the wealth distribution and capital operation of human society will be completely overturned.

Many of the "intellectual barriers" and "experience premiums" we relied on to survive in the past will be completely shattered before AI. Altman even expressed his distaste for the "Token" billing mechanism: it's like naming chips "2nm"—a fake and absurd unit

This podcast conversation, with its enormous information density and deep vision, is enough to make everyone think deeply.

He knows full well that if these technologies are locked in government backrooms and the public is merely fobbed off on PowerPoint slides with "one day it will cure cancer," while painting a future where everyone loses their jobs and is under strict surveillance, that would be an extremely bad and dangerous sales strategy.

The truly responsible approach is to put the models in the hands of the public.

Let ordinary people use AI to achieve a good life, rather than barely surviving on universal basic income while a small elite decides their future for them.article_adlist[

References:

Editors: Aeneas, David ]article_adlistarticle_adlist[

]article_adlist news_keyword_pub,stock,sz000591&sz300024 iframe below body iframe below body general statement general statement QR code at end start QR code at end start

QR code at end start QR code at end start Editor name and work code Editor name and work code