Altman on memory allocation: stuffing memory into personal computers is too wasteful; it should be concentrated in AI data centers

"From the perspective of global resource allocation, having everyone hoard scarce DRAM in their own computers, where it sits idle most of the time, is a 'quite terrible' arrangement; memory should be concentrated in cloud data centers."

That pretty much says it all.

This Thursday, in an interview with the YouTube channel Cheeky Pint, OpenAI CEO Sam Altman drank beer throughout and chatted in a relaxed atmosphere about local inference of large AI models, and the resulting problem of insufficient memory in home computers—and he let slip a memorable remark without meaning to.

As we know, the memory price surge triggered by the current global AI wave has already sent PC and smartphone prices soaring, leaving individual consumers unable to bear it. At the same time, capacity expansion by the major memory manufacturers will take a while. Just as frontier AI companies are taking the lion's share of memory, they have already begun discussing whether the memory in personal computers is being fully utilized.

Altman raised a hypothetical: that people could run local AI on their own computers—computers they may already own, or ones they want to build because of the rise of large AI model technology. He thinks this is a bad idea.

"From a global resource allocation perspective, having everyone leave the world's limited DRAM sitting idle in their own computers with only 5% utilization is actually quite terrible," Altman said. "If you want to operate in an environment where this resource supply is extremely tight, then in the cloud you should maximize resource utilization. But somehow, many people's intuition is exactly the opposite."

Today, OpenAI's main business is based on closed-source models running in large-scale data centers, and his defense of his company's business model has sparked debate across the AI community.

It has also raised concerns among many: in the future it may become harder for us to own our own hardware or software, and we may only be able to consider renting cloud-based agents and computers.

Although ordinary PCs, phones and AI accelerator cards use different types of memory, their production competes for some upstream wafer capacity. Altman was discussing the allocation of global DRAM resources, and when it comes to who gets to decide where memory capacity is used, most of the time it is naturally the customer placing the larger order.

In fact, this interview was long, and Altman also talked about the PHP vulnerability that set off global panic these past two weeks, sandboxing of trained models, and AI research in mathematics.

Let's see what he mainly said:

Regarding the enormous PHP vulnerability, Altman's judgment was: the ability of large models to find vulnerabilities continues to improve rapidly, and the code surface on the internet is extremely vast; when people begin systematically using AI to scan all code, "we will find ourselves in a world where serious vulnerabilities far exceed expectations." He believes patching alone is not enough and that the architecture must be rebuilt—routers, operating systems and the application layer all need to be redesigned for a world where "vulnerabilities will get smarter and smarter." Two years ago, "rewriting the entire tech stack" still sounded crazy; now startups aiming to build an entirely new stack have emerged.

He rarely offered industry self-criticism: "I don't think we've done a good job explaining to people why all the potential downside risks of this technology are worth it." He said the AI industry's narrative is bad salesmanship—promising to cure cancer and cut costs on one hand, while painting a future of job losses, loss of control and the need for strong regulation on the other. He reiterated that OpenAI insists on iterative deployment and guarding against capability overhang, on the grounds that "a few square miles of people in San Francisco" must not be allowed to make all the decisions.

When asked whether a "maximum-security prison" could be built for models, Altman gave a pessimistic answer: sandboxes are only effective to a certain extent, and monitoring (including chain-of-thought review and activation monitoring) can go a step further, but ultimately there must be a high degree of trust in alignment. He asked in return: facing a model a million times smarter than the smartest security researcher, one capable of discovering new science that humans do not yet understand, would you really dare to trust that sandbox?

He made clear that he hates the term "agent swarms," because swarm evokes locusts, and insists on calling them agent teams; he believes that the property of mutual checks and balances among teams corresponds precisely to the mechanism of achieving robustness through pluralism, and is a safety direction worth pursuing.

Intelligence explosion and the cost of scientific discovery

Altman sketched a curve: in the GPT-4 era, it was hard to effectively spend even $1 of compute on a single problem; chain of thought let models "think hard," and now we can run an entire agent team in parallel tens of thousands of times. The usage gap between ordinary users and the most advanced users is constantly widening.

Regarding how compute is allocated internally at OpenAI, he said employees enjoy a great deal of freedom, but high token consumption is never rewarded: "A high ranking on the token leaderboard is actually a bit embarrassing—it's certainly not a reason for promotion." He used the proof of the Navier-Stokes equations as a case study: a single proof cost about $10 million, and during that period things "went off the rails for a week," because at the time no one knew anything about the scaling laws for agent teams. But he expects that a year from now, with smarter models and methods, it could be redone at 1/100 the cost.

He also rarely showed emotion: the situation at the time was truly extraordinary—on the weekend when a certain new model launched internally, people threw all kinds of the toughest problems at it, and it managed to solve them one by one. People had never expected the model to be able to do these things. Together everyone witnessed many of the hardest research problems in people's academic careers being solved one after another by this model, spanning multiple different fields—the scene was simply astonishing.

"For two nights I barely slept at all, because I was so excited I simply couldn't sleep. A lot of the results we later published were actually those problems solved at that time."

Altman felt immensely excited and happy at the time, and at the same time an unexpected sorrow, melancholy, and a sense of "an era is coming to an end" welled up. It was not just nostalgia, but also a kind of empathy for human scientists' "hear the Way in the morning, and die content in the evening." He insists that humanity deserves access to basic knowledge, but the industry owes the practitioners in these fields a solution.

Although it is still too early to talk about superintelligence, Altman believes that a trillion slightly-superhuman AIs are better than one trillion-IQ AI:

"I would much rather intuitively have a trillion AIs of slightly superhuman intelligence join society than one AI of one trillion IQ join society."

His argument is: society's strong robustness comes from a large number of members checking and balancing and cooperating with one another; a "single-brain" world lacks this kind of resilience. But he stresses that this will not happen automatically—it must be actively pushed for.

On the issue of the global memory shortage, Altman laid out OpenAI's plans: in-house chips (improving intelligence performance per watt), self-built data centers, robotics projects, and an even bolder concept, "self-replicating data centers," using AI to control robots to operate fabs, build turbines and solar panels, and construct data centers. He said "this is already within our grasp."

Asked why "the market regulating itself" has been so slow to materialize, he said current progress is terrible, and explained the behavior logic of memory manufacturers: this industry has long been a hard business of investing $20 billion in capex to earn $1.9 billion, and historically giants have repeatedly come close to bankruptcy from overbuilding plants. Today each company can earn about $100 billion in free cash flow a year, but the painful memories make them extremely conservative about expanding production. He thus asserted that "the invisible hand is failing right now."

On the real level of recursive self-improvement (RSI), Altman said models can now complete tasks like "find a 5% compute efficiency improvement," but no one yet knows how to get AI to discover a new architecture on the level of the transformer—what's missing is smarter models. He even predicted that society may need to agree for a period of time to limit the IQ of a single model: "If future progress is insufficient, that will stem from this decision, not from AI's ability to do research."

He laid out his deflationary faith: the current 5-year total cost of ownership per GW of leading AI compute is about $70 billion, and it would be even better if that could drop to $20 billion, while models must also improve performance 10x every year. The path to achieving this is making energy cheaper and compressing the layers of margin stacked between "sand and output tokens."

Regarding enterprises' anxiety about token budgets, he believes the more common problem today is not using AI too much, but not using it enough.

How should personal agents be priced?

Regarding the recently hot personal agent products (such as Muse), Altman's current judgment is that "the Macintosh moment has not yet arrived." People like these products, but they haven't reached the point of "I can't imagine life without it"; the winner's form will resemble today's products, but "it'll just work"—reliable, able to handle more complex tasks. As for the cat-and-mouse game of computer use, he believes it will ultimately converge to the extent that "you have the right to have AI operate your own computer on your behalf."

On how agents will be commercialized, Altman believes enterprises cling to the AI pop-ups and HTML inside their own products, fearing the loss of customer relationships, but consumers want to have AI do everything for them inside their main chat app. He believes the real moat comes from brand, product selection, and fulfillment networks, not HTML. He volunteered that he chooses Amazon Pay 10 times out of 10, even at a higher price, and praised Amazon's fulfillment network, while admitting that the partner platforms within ChatGPT are not yet good enough. He also called all web forms "fake work," and compared the present to the App Store era back then.

It is precisely for this reason that when designing pricing, OpenAI almost learned from AWS by quoting per task and completely abandoning token-based billing, but this met strong opposition from customers: "I need to know what I'm paying for"—even if they pay more, they want tokens as an understandable unit.

Reference content:

This article comes from the WeChat public account "Machine Heart" (ID: almosthuman2014), author: Ze Nan, published by 36Kr with authorization.