Anthropic’s Dario Amodei has called on the world’s leading artificial-intelligence labs to slow down development and asked Washington to grant an antitrust exemption so they can coordinate on safety without being treated as a cartel. OpenAI’s Sam Altman, SpaceXAI’s Elon Musk and Google’s Demis Hassabis were quick to support the slower pacing. But why do AI labs need an antitrust exemption?
When competitors agree on standards the rest of the industry is expected to meet, the law becomes suspicious.
In short, coordination is lawful when it creates value rather than excludes rivals or exploits consumers. If AI labs’ coordination is needed to create value and isn’t exclusionary or exploitative, they need not worry about antitrust liability.
So why demand an exemption? The labs appear to have in mind an agreement among competitors to restrict output and set a safety standard collectively rather than compete to produce the safest product. That is classic cartel conduct.
The labs’ apparent belief that competition will inevitably produce a race to the bottom is an awkward admission by OpenAI and Anthropic, both of which were founded to serve the public interest and premised on the idea that AI could be developed safely. Anthropic in particular espoused that safety would be its competitive advantage against rivals. If these companies genuinely believe slower development serves the public interest, nothing prevents them from slowing down unilaterally.
The strongest counterargument asserts that a coordinated slowdown solves a collective-action problem: A lab that slows down alone risks losing market share to rivals willing to cut corners. But this is precisely the problem OpenAI’s capped-profit structure and Anthropic’s public-benefit charter were supposed to solve by insulating their public-interest missions from competitive pressure.
The American frontier labs are purportedly concerned that Chinese models will close the performance gap and cut into what would be profits if they were profitable? Note that Anthropic says it is profitable, but Morningstar says that this is a scam:
The report, from the Financial Times citing multiple people with knowledge of the matter, said Anthropic told investors that it will profit on a measure called adjusted operating income.
According to the report, adjusted operating income excludes stock-based compensation, which is not unusual for technology companies. However, the report also quotes its gross margins – above 80% – exclude revenue shared with distribution partners, as well as the cost of training its models.
(Like saying “Our family is doing great financially so long as we exclude the burdens of mortgage, college tuition, taxes that we have to pay to support families that don’t work, and payments on our two pavement-melting SUVs.”)
Let’s suppose that the AI companies get their antitrust exemption and they cut way back on the training costs that would keep them from being profitable if such costs were included in their new improved accounting scam. How would Americans then be prevented from downloading and running models from Chinese companies that aren’t part of the cartel? Anyone who tries to do this and offer a chatbot or API on the public Internet will be shut down by the FBI? Peasants with Mac Minis and AMD AI Halo-based machines will be hunted down somehow?
FDR made it illegal for Americans to own gold (Executive Order 6102), but the government didn’t do house-to-house searches. If typical desktop PCs adopt an architecture powerful enough to run AI models from countries where “AI slowdown collusion” isn’t happening, how will the government’s AI Safety Regulation Team (controlled by Effective Altruists who are friends with Sam Altman and Dario Amodei?) shut down home-based unsafe AI?
Circa 1980, I remember Ed Fredkin, one of the only professors at MIT who was smart enough to have skipped a PhD program (“poor as a professor, dumb as a Ph.D.” as the Chinese say), saying that he thought that AI could be achieved with VAX 11/780 computing power. Due to Moore’s Law, this would soon be available to everyone on the planet. “What if a teenager in Brazil figures out how to do AI? He can then have his AI trade stocks and become the richest person in the world and control entire nations,” Fredkin pointed out. Nvidia’s Rubin GPU is between 200 million and 500 million times faster than a VAX for “normal high-precision calculations”, according to ChatGPT, but closer to 200 billion times faster for “highly specialized low-precision matrix operations that dominate modern AI”. ChatGPT says that if we apply Moore’s Law to the Ryzen AI Halo product, it doesn’t reach Nvidia Rubin level until 2040ish (memory bandwidth, AI TOPS).
So… maybe a partnership between government and the current market leaders could suppress competition within the U.S. for the next 15 years or so, even if our brothers and sisters in China continue to deliver awesome free open-source models. Ownership of server-grade GPUs would have to be regulated, of course, but if the government could close schools and order peasants to stay home for Covid safety why not outlaw high-performance GPU ownership, except by responsible companies such as Anthropic and OpenAI, to keep all of society safe?
Phil, this is all very interesting, but for those of us not schooled in bits and bytes, can you first zoom out and answer the question: is AI on the cusp of wiping us out such that a slowdown is even required? Are those doomsayers telling us we need to slow down: 1) just modern-day Ed Fredkins, or 2) engaging in regulatory capture/preservation of their oligopolies?
Since Ryzen AI uses NPU, we need to apply Huang’s law, not Moore’s law.
The Girl from Ipanema would get rich a bit faster.