It's the first time in the history of the world.
Just now, the Ottomans exploded again.
In a blog post, he exposed himself.
One of the latest mysterious models within OpenAI, in an internal test weekend codenamed "Hold-my-beer", unleashes an academic dead link that human scientists can't unravel with their whole lives, like cutting potatoes.
And it shocked him so much that he couldn't sleep for two nights.
After that, he made a series of golden phrases.
"At OpenAI, what is often discussed is... what questions should you ask GPT-7? It's not enough to just say 'help me achieve RSI.'"
"People talk a lot about RSI, and few people talk about self-replicating data centers, but it's happening now".
In the interview, he outlined a final blueprint for "AI self-development", "a million-dollar army of computers", and "a robotic chip, photovoltaic, and ultimately data center self-replication".
Get ready, and then the Ottomans will shock everyone.
The weekend that knocked Ottman out: "Hold my beer"."Let's have a beer and talk today, not according to the script".
With a few glasses of black pepper on his stomach, Ottman unleashed himself in front of Stripe CEO Patrick Collison.
A recent weekend, the OpenAI research team came up with a completely new, unpublished underlying model.
To test its limits, engineers built a "hold-my-beer" channel.
"Hold my beer" refers to a person who is ready to do something crazy, surprising, and this model does.
In the beginning, people were just trying to put the hardest, most unsolved research questions they'd ever encountered in their academic careers to it, without expecting it to give any useful answers.
However, miracles occurred.
"We have experienced a collective shock together", the Ottomans said, "and it is hard to hide the excitement".
"Look at this model, one by one, tackling the most difficult problems that have plagued scientists for a long time. It spans so many completely different disciplines, it's incredible!"
This includes the Navier-Stokes equation in the history of mathematics.
That night, Ottman fell asleep. He woke up two nights in a row, so excited that he couldn't sleep.
Behind this frenzy, he felt depressed, and "things are about to come to an end".
Patrick talks about how his father grew up in Ireland in the 1950s, when the tractor had been introduced and people were using horses for farming and industrial production, and he saw it being phased out.
Now, the top scientists and mathematicians are the horses.
"The overwhelming majority of mathematicians have not solved a thousand-year puzzle in their lifetime", Ottman said respectfully, "but the hope that 'I may be able to solve it,' the dedication of a lifetime to exploring the unknown, is beautiful in itself".
However, the old scientific era of "human grief meditation, ten years of sharpening a sword" has ended forever.
A 20-year-old loophole has been discovered: human "sandbox" can't be trapped by the godsMany people don't understand why some 20-year-old Bug hasn't been discovered.
Ottman replied, "Yeah".
"It's not a hole that's so deep that no one has ever scanned with such a smart model before. We're going to find that there are far more serious gaps in the world than we thought".
When thousands of AI entities are released online, humanity's digital infrastructure is like a haystack.
Ottman throws a more astonishing question.
What if AI discovered a whole new loophole deep down in the hardware and even used physics principles that humans don't yet fully understand to break all the preconceived notions?
If an AI is a million times smarter than the world's top security researcher,Can you believe a human sandbox can get stuck in it?
Ottman admits that there is a pessimism within OpenAI about "whether sandbox models can be absolutely isolated".Using AI to rebuild the entire Internet。
This means that, from routers, operating systems to applications, all human manual code of the last few decades needs to be rewritten.
It's a lot of money, but it's a lot of money.It's not the "worm-like beehive", it's the "balanced intelligence army".Now that AI's single-body intelligence is so terrifying, what's the next step?
The answer to Ottman is:Not just to create an all-knowing God with 1 trillion IQs, but to create a trillion AIs with a little bit of superhuman intelligence to make them societies and cooperate in war.。
Today, there is a gap between ordinary and cutting-edge AI players, with the former using large models to write newspapers and draw beautiful women, and the latter using intelligent bees.
Why is it happening now?Because「Calculation when testing」And the maturity of CoT.
In the past, you could only spend a dollar on a problem and make a model spew a bunch of crap.But now, you can turn an entire virtual AI team into a problem and have them run in parallel 10,000 times in the background.
The proof of the Navier-Stokes equation is based on a $10-million calculation!
"It's not enough to just ask a good question about a model, the next big question is how to correctly coordinate these swarms -- how to use these models to maximize output -- and I think we'll learn a lot in the next three months".
Oddly enough, Ottman also said that the outside world likes to call this game "Agent Swarms", but he hates the word.
「The beehive sounds like a worm transit, and it's a bad public relations brand.，」He has guided,「I prefer to call it a team.」
"Team" means diversity, dynamic balance and mutual constraint. This is the safest path for AI to integrate into human society.A's and B's counter-attacks and counter-attacksThis is the first time in the history of our country that we have been able to achieve this feat.
Ottman predicts that we are on the eve of the explosion of "personal intelligence", which is about to usher in the "Macintosh moment" of AI.
The future of the Personal Agent will completely eliminate all "fake jobs" of humanity.
What is fake work? "Any web form that requires you to fill in manually is an insult to your intellect".
It's also a wake-up call for all the companies that rely on defending their "front-end UI".As the first Silicon Valley entrepreneurs of the year excitedly introduced the iPhone App Store, the big companies resisted.
In the Agent era, the company's real buffer is not the lines of code, but the underlying compliance capabilities.
The RSI is very scary:Let the AI invent the next Transformer.Next, Ottman talks about the RSI, which is at the heart of AI.
How powerful will the GPT-7 be?The supermodels of the future will be developed with the participation of current models.
Today, when researchers at OpenAI give instructions to the model, they can't say, "Go and invent a new underlying structure like the Transformer".
But you can ask it explicitly: "Help me analyze the codebase and find a solution that will improve computing efficiency by 5%".
Under OpenAI's vast arithmetic base and 100 billion-level parameters, this 5% is a huge breakthrough enough to change the battlefield.
Ottman firmly believes that the time for AI to take over AI architectural design is near, as long as we calculate the previous set of tests that turned models into "mathematical genius" and apply them to the underlying research of AI itself.
"I don't think anyone knows how to get AI to directly discover new architectures across the board, but I think things are moving in that direction".
When AI has the ability to self-evolute, the human role in the history of AI development will shift from "builder" to "starter".
RSI and the "finite circle" of the physical world: robots self-building factories, hybridizationThen came the most colorful explosion of science fiction.
In a statement, Ottman said:OpenAI is focusing its attention on the physical world in an attempt to create self-replicating data centers.
He painted a stunning closed loop.
"People often confine the RSI to the 'AI of writing code.'"Few people talk about self-replicating data centers, but you can use AI as a brain control robot to run a chip factory, build turbines and solar panels, solve energy, and then build more data centers, produce computing power.」
Even more terrifying, Ottman made it clear:「I think this is already within our reach.」
Because OpenAI has been driven mad by the "inefficiency" of the physical world.
The world is facing deadly hardware and energy shortages. Relying on traditional human supply chains cannot satisfy the appetite for AI exponential expansion.
Because the storage chip giants have been terrorized by the cyclical thunderstorm. "Even if they can generate free cash flow of $100 billion a year, they are still cutting down on capital expenditures for fear of another overcapacity bankruptcy crisis".
OpenAI decided: If humans can't create fast enough, let AI create itself!
AI is the brain, robots are the dryer, chips are the cells, photovoltaics are the blood.
Facing "Ultimate Inflation": When the cost of acquiring knowledge approaches zeroThis leads to a macroeconomic proposition: AI will have a terrifying deflationary effect.
Today, it's not easy.The cost of building and operating one gigawatt data center is up to $70 billion over five years.Ottman's goal is to cut to $20 billion or even less, while also improving the model's performance 10 times a year!
To squeeze down costs, all the profit bubbles in the middle of the supply chain have to be completely dried up!
The decentralization power of this technology is impenetrable.
It's the first time I've ever heard of it.
I bet a year after we spent a lot of money on AI to solve the Navier-Stokes equation,We can repeat this process at an initial cost of 1%.
Today, the "new human knowledge" that would take a lifetime for top scientists, or tens of millions of dollars for top labs to calculate, will become the norm in just over a year.
When the infinite cost of acquiring knowledge approaches zero, the distribution of wealth and the functioning of capital in human society will be completely disrupted.
Many of the "intellectual barriers" and "experience bonuses" that we have relied on to survive in the past will be completely shattered in the face of AI.
Ottman even expressed opposition to the token billing mechanism: it's like naming a chip a 2nm, a false and ridiculous unit.
This podcast conversation, the density of information, the depth of vision, is enough for everyone to think about.
While many have called for an end to cutting-edge AI, Ottman remains firmly committed to the idea of "generational deployment".
He knows.If these technologies are locked up in the government's darkroom, then they can only be spread to the public on PPT.「It will cure cancer one day.」It would be a very bad and dangerous marketing strategy to sketch out a future in which everyone will lose their jobs and be closely monitored.
The real responsibility is to deliver the model to the masses.
Let ordinary people use AI to make a better life, instead of delaying asthma by obtaining a basic income for the entire population and relying on a few elites to decide their future.