Reid Hoffman is pushing back on the growing call for a broad slowdown in frontier AI research.
At Politico's Decoded Summit in Washington, the LinkedIn cofounder said fears that AI will destroy everything are too remote and controllable to justify stopping. He said the debate often sounds like 'Look at me, I'm Cassandra, I'm saying it's the end of time.'
Hoffman said he believes continued frontier AI work can help address biosecurity and cybersecurity risks, while also supporting economic prosperity. 'I don't think that slowing down is important,' he said, adding that this is a reason 'to keep going.'
He also questioned why leaders at Anthropic and OpenAI are now calling for a slowdown while their companies are among the field leaders. 'You're in the lead, you can slow down anytime you like,' Hoffman said.
The discussion comes as frontier AI labs face a sharper debate over recursive self-improvement, the process by which AI can learn to build itself and eventually its successors. Some in the industry worry that this could create exponential growth that model makers may not be able to fully control.
Jacob Coxon, a former Anthropic researcher, wrote that Anthropic 'largely initiated' the race to recursive self-improvement. Over the weekend, Anthropic CEO Dario Amodei called for frontier AI labs 'to pace' their developments and use the time to better align AI models. Elon Musk and Sam Altman later said they agreed with Amodei's broad aims.
In an essay published on Wednesday, Hoffman said Amodei and Altman's calls were steps 'in the right direction.' Mark Zuckerberg also argued that companies can slow down individually without an industry-wide pact, and said Meta delayed the release of Muse by months to strengthen safety and security.
For founders, the split matters because AI is no longer just a product story. It is a capital allocation question. If frontier labs can keep shipping while tightening safety, the competitive edge stays with the firms willing to move fast and manage risk instead of handing the pace-setting role to regulators or consensus-driven pauses.
The math is simple: delay can protect users, but it can also protect incumbents. The winners will be the teams that pair speed with disciplined controls, not the ones that confuse caution with strategy.
