Ex-OpenAI employee writes AI essay: War with China, resources and robots
The race for AGI has begun, but according to Leopold Aschenbrenner, the race for the necessary resources and power is even more important.
(Image: Shutterstock/Alexander Supertramp)
Leopold Aschenbrenner looks a bit like the villain of a Bravo photo love story from the 90s, blonde quiff, broad smile. But the scientist, who originally comes from Germany, is above all a high-flyer: according to his LinkedIn profile, Aschenbrenner graduated from Columbia University in the USA at 19, top of his class. After a few jobs in research and investment, Aschenbrenner joined OpenAI in 2023 in the Superalignment department. This is the team that deals with the security of superintelligence. One and a half years and a few squabbles at the AI company later, Aschenbrenner was made redundant.
Shortly afterward, the economist published a 165-page essay that is currently attracting a lot of attention. Aschenbrenner is apparently one of those people who can be described as AI ultras. He assumes, for example, that in just one or two years, AI will be smarter than most college graduates. By the end of the decade, he says, AI will be smarter than all of us, including Aschenbrenner himself. "We will have a superintelligence, in the true meaning of that word," he writes, for example. It also sounds a little dubious when he writes about "The Project", which he capitalizes. This project could either lead to a race with the Chinese Communist Party or to "total war".
Reasons for Aschenbrenner's dismissal from OpenAI
This concern is also said to have been one of the reasons why Aschenbrenner was kicked out of OpenAI. In a podcast, he said that there had been a security incident at OpenAI, after which he compiled a memo that he claims to have shared with some members of the board. The memo was about the possible theft of "important algorithmic secrets from foreign actors", writes Business Insider. The HR department described worrying about espionage by the Chinese Communist Party as "racist and unconstructive" and warned him. In addition, according to Aschenbrenner, he was accused of passing on confidential information. This allegedly involved a brainstorming document that he shared with scientists to get their opinion.
Sharing is probably also the exciting thing about the essay. Aschenbrenner is in contact with various tech entrepreneurs, employees and investors in San Francisco and Silicon Valley. In his essay, you can read about how some of these people see the future, what they talk about - and what mindset prevails among them.
While some scientists assume that it will be a long time before AI is truly intelligent and reliable, Aschenbrenner believes that the next breakthrough is just around the corner. This is although he himself writes that "the magic of deep learning is that it simply works". Other experts call this magic a "black box" because no one is quite sure how and why it works so well. And that's why it's also clear to many that progress will not necessarily be linear. Meta's head of AI, Naila Murray, for example, has said that she sees clear limits to large language models and that scaling does not reflect the intelligence that we humans have. And even if AI continues to make great progress, there are limiting factors: energy and data.
Where does the energy for AI come from?
Aschenbrenner is also clear when it comes to the resources that AI needs. He believes that by the end of the decade, individual training clusters will consume more than 20 percent of all the electricity generated in the USA. Nvidia's soaring stock market and sales are just the beginning, he says. "Trillions of dollars of investment will produce hundreds of millions of GPUs per year", the essay says. "Where can I find 10GW? (Power for the $100B+, trending 2028 cluster) is a popular topic of conversation in SF," says Aschenbrenner. He calculates the demand and investments in detail. Aschenbrenner says, "Surprisingly, even 100GW clusters are easy to implement. Aschenbrenner is relying on natural gas, saying that only around 1200 new boreholes would be needed. For comparison: According to EnBW, the total net output of all generation plants in Germany is around 252.8 gigawatts (GW).
Read also
Bosch smart e-bike system: Using AI to combat range anxiety
Report: Apple Intelligence not for Apple's speakers
Monopoly: EU has further questions for Microsoft, OpenAI, Google and Samsung
Bill Gates misrepresents the power consumption of AI data centres
HPE: AI stack with models, servers, network, data management and admin console
Aschenbrenner's calculations also include the fact that OpenAI is continuously increasing in value. While the company was worth one billion US dollars in August 2023, it was already worth two billion US dollars in February 2024. Aschenbrenner therefore expects a valuation of ten billion US dollars by the end of 2024 or early 2025. And he writes: without there being a new "next generation model".
Bill Gates also recently commented on the subject of energy. His calculations are also typical of Silicon Valley. Gates believes that what AI will consume, it will recoup. Namely, AI algorithms will lead to savings through better calculations and progress. OpenAI's CEO Sam Altman is already investing in a start-up that relies on fusion power plants. There are said to be talks with OpenAI about contracts to purchase the energy that does not yet exist. Helion has even already signed contracts with Microsoft. The power plant in the state of Washington is expected to be up and running by 2028.
According to his own website, Aschenbrenner himself has now founded an investment company that, of course, specializes in AI. He may therefore also benefit if the hype continues and expectations remain high.
Incidentally, in Aschenbrenner's vision, the data that AI needs comes from robots. Built by robots, in factories created by robots.
(emw)