What Was Discussed at Google’s White House Meeting About A.I.
The technology giant met with administration officials last week to address a growing concern in Washington: insufficient computing power for artificial intelligence.
The Trump administration’s compute fears
The C.E.O. of Alphabet, Sundar Pichai, was at the White House on Thursday for a series of high-stakes meetings with senior Trump administration officials.
The official agenda was cybersecurity threats. But DealBook has learned that the subtext was more specific: worries about artificial intelligence “capacity” — and the government’s struggle to secure enough A.I. processing power to maintain its own defenses.
It’s the latest fallout from Anthropic’s Claude Mythos Preview model, access to which is restricted to several dozen corporate and government users. Anthropic has described the limited release as a safety-first measure, but it has also spooked the Beltway.
Recent tests suggest that Mythos can identify and exploit vulnerabilities in critical software infrastructure at a speed traditional cybersecurity teams can’t match.
But the administration is worried that Anthropic doesn’t have enough “compute,” industry lingo for computing power, and that the company might have to throttle the use of Mythos even for high-priority users. In a crisis, some officials fear, the government could be locked out of the tools it needs to patch its software systems.
That’s where Google (and OpenAI and others) come in. Even as the administration appears to be easing tensions with Anthropic, it’s hoping to use Google’s Gemini, OpenAI’s GPT models and other bleeding-edge tools to decrease its reliance on Claude.
Separately, the Pentagon last week said that it had reached deals to use several more models beyond Claude on classified work.
Google does face a big challenge, DealBook understands: Some of its A.I. processing chips, known as TPUs, can’t be used in some classified contexts. Anthropic became an easy option for the government because its models largely run on Amazon Web Services, Amazon’s cloud computing platform, which has spent billions to achieve so-called Impact Level 6 security certifications.
The government is seeking ways to help accelerate similar clearance for Google TPUs.