Jeffrey Ladish says there is a simple way to slow AI progress. The chips already exist, and companies already decide how to use them. He wants the government to ask the companies to change that decision.
Ladish is executive director of Palisade Research and previously built security infrastructure at Anthropic. He made the proposal in a conversation with Steven Bartlett on The Diary of a CEO, in an episode published October 8, 2026.
The brake pedal
"We have a brake pedal we could implement," Ladish said. He called it "fairly simple."
He started with the hardware. AI companies run huge data centers full of GPUs, the specialized chips used both to train AI models and to run them. Every time someone uses ChatGPT or another AI product, the request runs in those data centers. Ladish said the leading companies, including Anthropic and OpenAI, divide their computing power between two jobs. The first is training, which means building the next, more powerful model, with AI agents helping to design it. The second is inference, which means running existing models to serve customers. He described the current split as "50-50."
His proposal is to dial that "way towards serving customers" and use far less computing power to train the next model. Bartlett asked whether the government could ask the companies to do that. Ladish said that was the proposal: the government should tell them "this is going too fast" and ask them to focus on serving the models they already have.
A related plan
Ladish said the idea is something "Daniel" has been working on since appearing on the podcast. He gave no surname. A closely related proposal appears in AI 2040: Plan A, a policy recommendation whose authors include Daniel Kokotajlo, along with Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland and Ryan Greenblatt. Whether Ladish meant Kokotajlo is not confirmed.
Plan A's authors estimate that major AI companies spend roughly half of their 2026 computing budgets on research and development, including training and large experiments. That is close to the split Ladish described. They propose limiting that share to slow the growth of AI capabilities while services built on existing models keep running. They also discuss "inference-only verification": checking that permitted computing power serves existing models rather than conducting new frontier training.
The authors explore the proposal through a detailed future scenario. It is a policy scenario, not their forecast. Its dates, including a transition to superintelligence in 2040, are assumptions used to examine how the policy might play out.
Why he is more hopeful than before
Ladish's overall outlook is bleak. Placing himself on Bartlett's range of futures, he said that "on the trajectory we're on right now, I think human extinction is very likely," though not inevitable. But he said his view had been moving in a better direction. He was more optimistic about avoiding extinction than he had been a month earlier, and more optimistic a month earlier than a year before.
His reason was rising awareness. People do not care much about which tools they have, he said, but "people care about their kids being able to grow up and go to school." He added: "I believe in people." In his view, once people see AI as a threat to their families, they will not stand for it. The problem, he said, is that the technology is so new and moving so fast that people have not yet seen it.
Calling Congress
Ladish said many people feel small against multibillion-dollar corporations and geopolitics. He disagreed, citing his own meetings. He said he has talked with members of Congress, including Bernie Sanders, and senators "on both the left and the right," and that they are starting to realize this could threaten public safety.
Bartlett asked what simple thing listeners could do. Ladish pointed to callcongress.ai, which he said some of his friends made to walk people through calling their representatives. He knows it can seem "a little cheesy," he said, but insisted it works: lawmakers "have to get reelected," and some are becoming concerned themselves. If they see a signal from constituents that this is a very important issue, he said, "I think Congress can act."
CallCongress.ai helps users find their representatives from an address or ZIP code, choose policy requests and prepare an editable calling script. It also offers reminders to call again. The requests it suggests are the site's own advocacy positions: international agreements to stop or slow frontier AI development, enabling companies to coordinate on safety, and transparency about AI security incidents.
"We provide the incentives"
Bartlett took the argument further. Power is pushing AI leaders in one direction, he said, but staying in office is also a strong incentive. He expected AI to be one of the most important subjects on the ballot in the 2028 election, with voters wanting their jobs preserved, safety and a future for their children. Donald Trump cannot legally be reelected, Bartlett said, but if he could get a third term he would have to change his position to win in 2028.
Ladish answered: "Incentives aren't just a thing that happen out there. Like, we are part of the incentives. We provide the incentives."
"For now," Bartlett added. Ladish repeated it: "For now."