Abstract:
“From the perspective of global resource allocation, it is a ‘quite terrible’ arrangement to let everyone store the scarce DRAM in their own computers and leave it idle most of the time. Memory should be concentrated in cloud data centers.”
That’s all.
In an interview with the YouTube channel Cheeky Pint on Thursday, OpenAI CEO Sam Altman drank beer throughout and talked in a relaxed atmosphere about local reasoning of large AI models and the extended problem of insufficient memory in home computers, which inadvertently became a hit.
We know that the current increase in memory prices caused by the global AI wave has caused the prices of PCs and mobile phones to skyrocket, which is already unaffordable for individual consumers. At the same time, it will take a while for the major memory manufacturers to expand their production capacity. While cutting-edge AI companies are eating up the bulk of memory, they have begun to discuss whether the memory in personal computers is fully utilized.
Altman mentioned a hypothesis that people would be able to run local artificial intelligence on their own computers, which may be computers they already own or that they want to assemble due to the rise of AI large-model technology. He thought it was a terrible idea.
“From a global resource allocation perspective, it is actually very bad to let everyone leave the world’s limited DRAM memory idle in their own computers with only a 5% usage rate,”
Ultraman said. "If you want to work in this kind of environment where resource supply is extremely tight, then you should maximize resource utilization in the cloud. But for some reason, many people's intuition is exactly the opposite."OpenAI's main business today is based on a closed-source model running in large-scale data centers, and his defense of his business model has sparked debate in the AI community.
This has also caused many people to worry: In the future, it may be more difficult for us to own our own hardware or software, and we can only consider renting intelligent agents and computers in the cloud.
Although ordinary PCs, mobile phones and AI accelerator cards use different memories, their production will compete for some upstream wafer production capacity. Altman was discussing the global DRAM resource allocation. When it comes to who is qualified to decide where memory capacity is used, most of the time it is naturally the user who places the larger order.
Part of Altman's argument is based on the assumption that data center construction will not slow down. Furthermore, he hypothesized that memory used 24/7 in an AI data center would be more valuable than memory used for spreadsheets, web browsing, games, and even the occasional local AI.
In fact, this interview was very long. Altman also talked about the PHP vulnerability that triggered global panic in the past two weeks, the sandboxing of training models, and AI research in the field of mathematics.
Look at what he mainly said:
AI is detonating a wave of vulnerabilities
Regarding PHP’s super huge vulnerabilities, Altman’s judgment is: the ability of large models to discover vulnerabilities continues to improve rapidly, and the code surface on the Internet is extremely huge. When people start to systematically use AI to scan all codes, “we will find ourselves in a world where serious vulnerabilities are far greater than expected.” He believes that patching alone is not enough, the architecture must be reconstructed, and routers, operating systems, and application layers all need to be redesigned for a world where "vulnerabilities will become increasingly intelligent." Two years ago, "rewriting the entire technology stack" seemed like crazy talk, but now there are startups that want to build a new stack.
He made a rare industry self-criticism: "I don't think we have done a good job of explaining to people why all the potential downside risks of this technology are worth it." He said the AI industry's narrative is a bad sales pitch, promising to cure cancer and reduce costs while painting a future of unemployment, loss of control, and the need for strong regulation. He reiterated OpenAI's insistence on iterative deployment and capability overhang on the grounds that "people in a few square miles in San Francisco cannot make all the decisions."
When asked whether he could build a "maximum security prison" for the model, Ultraman gave a pessimistic answer: The sandbox is only effective to a certain extent, and monitoring (including thought chain review and activation monitoring) can go further, but in the end it must be highly trusted and aligned.
He asked rhetorically, faced with a model that is a million times smarter than the smartest security researcher and can discover new science that humans do not yet understand, do you really dare to trust that sandbox?
He made it clear that he hates the word "agent swarms" because swarms are reminiscent of locusts, insisting on calling them agent teams (agent teams), and believes that the nature of mutual checks and balances between teams corresponds to the mechanism of achieving robustness through pluralism, and is a safe direction worthy of advancement.
Intellectual explosion and the price of scientific discovery
Altman outlined a curve: In the GPT-4 era, it was difficult to effectively invest $1 in computing power on a single problem; the chain of thought allows the model to "think hard", but now we can parallelize the entire team of agents tens of thousands of times. The usage gap between ordinary users and the most advanced users is constantly widening.
Regarding how computing power is allocated within OpenAI, he said that employees enjoy a lot of freedom, but high token consumption is never rewarded. "It is actually a bit embarrassing to be high in the token rankings, and it is certainly not a reason for promotion." He used the proof of the Navier-Stokes equation as an example: a single proof cost about $10 million and was "out of control for a week" because no one knew anything about the expansion law of the agent team at the time. But he expects it to be redone at 1/100 the cost in a year using smarter models and methods.
He also showed rare emotions: The situation at that time was really amazing. On the weekend when a new model was launched internally, everyone threw all the most difficult problems to this model, and it could solve them one by one. People didn't expect the model to do this at all. Everyone witnessed that the most difficult scientific research problems in many people's academic careers were solved one by one by this model, and covered many different fields. The scene was simply amazing.
“I almost stayed up all night for two nights because I was so excited that I couldn’t sleep at all. Many of the results we later published were actually the problems we solved at that time.”
Ultraman felt extremely excited and happy at the time, but at the same time he also felt an unexpected sadness, melancholy, and a feeling that "an era is ending." Not only is it nostalgia, there is also a kind of empathy for human scientists who "learn the truth in the morning and die in the evening." He insists that humans deserve basic knowledge, but industry owes practitioners in these fields a solution.
Although it is too early to talk about super intelligence, Ultraman believes that trillions of little superhumans are better than one trillion IQ:
"I would much rather intuitively have a trillion AIs of slightly superhuman intelligence join society than one AI of one trillion IQ join society."
"Intuitively, I would rather have one trillion AIs with slightly superhuman intelligence join society than one trillion AIs with IQ."
His argument is that the strong robustness of society comes from the checks and balances and cooperation of a large number of members. The "single-brain" world does not have this kind of resilience. But he stressed that this will not happen automatically and must be proactively promoted.
Market regulation will no longer work in the AI era?
Regarding the global memory shortage, Altman introduced the layout of OpenAI: self-developed chips (to improve intelligent performance per watt), self-built data centers, robot projects, and a bolder concept of "self-replicating data centers" that use AI-controlled robots to operate wafer factories, build turbines and solar panels, and build data centers. "We have this under control," he said.
Talking about why the "market self-regulation" has not been realized for a long time, he believes that the current progress is very bad, and gives the behavioral logic of memory manufacturers: This industry has long been a bitter business of investing 20 billion US dollars and earning 1.9 billion capex. In history, many giants have almost gone bankrupt due to excessive construction of factories. Today, each company can earn about US$100 billion in free cash flow a year, but the painful memories have made them extremely conservative in expanding production. From this he asserts that "the invisible hand is now out of action."
He distinguished two types of capabilities in the chip industry: Lithography "only one company on the planet seems to know how to do it," but chip design (the file that is ultimately sent to the fab) is already an excellent place for AI.
Talking about the true level of recursive self-improvement (RSI), Altman said that the current model can complete tasks such as "finding a 5% computational efficiency improvement", but no one knows how to let AI discover new architectures at the level of transformer. What is missing is a smarter model. He even predicted that society may need to agree to limit the IQ of a single model for a period of time. "If there is insufficient progress in the future, it will stem from this decision, not the ability of AI to do research."
He showed his belief in deflation:
The current five-year total ownership cost of each GW of leading AI computing power is about 70 billion US dollars. It would be ideal if it could be reduced to 20 billion. At the same time, the model performance must be improved by another 10 times every year. The path to achieve this is to make energy cheaper and compress the gross profits stacked between "from sand to output tokens".
As for enterprises’ token budget anxiety, he believes that the more common problem currently is that AI is not used enough, rather than using too much.
How to price personal intelligence?
Concerning the recent hot personal intelligent products (such as Muse), Ultraman's current judgment is that "the Macintosh moment has not yet arrived." People like these products, but not to the point of "can't imagine life without it." The winner will look like the current product, but "it'll just work," reliable, and able to handle more complex tasks. The cat-and-mouse game of computer use, he believes, will eventually converge to the point where "you have the right to let AI operate your own computer on your behalf."
As for the commercialization method of intelligent agents, Altman believes that companies stick to AI pop-ups and HTML in their products and worry about losing customer relationships, but consumers want to let AI do everything for them in their main chat application. He believes the real moats come from brand, selection and fulfillment networks, not HTML. He proactively revealed that he would choose Amazon Pay 10 times out of 10, even if the price is higher. He praised Amazon's fulfillment network and admitted that the cooperation platform within ChatGPT was not good enough. He also called all web forms "fake work" and compared the current situation to the App Store era.
It is precisely because of this that when designing pricing, OpenAI almost imitated AWS's pricing based on tasks and completely abandoned token billing. However, it encountered strong opposition from customers: "I need to know what I am paying for." Even if it is more, it must be in understandable units such as tokens.
Comments