Sacks reads the Kimi K3 cyber report as China winning. Marcus says it's the wrong race.

A preliminary UK and US government assessment of Moonshot AI's Kimi K3 model, published this month, is now being read in two opposite directions. David Sacks sees it as confirmation that China is closing the gap with the United States. Gary Marcus sees the same document as evidence that the US-China framing is the wrong race to run in the first place.

The report is a joint preliminary assessment by the UK AI Safety Institute (AISI) and its US counterpart, the Center for AI Safety Innovation (CAISI), hosted on the National Institute of Standards and Technology website. It evaluates Kimi K3's cyber capabilities: what the model can do, in concrete terms, when given access to software and a target. Sacks treats the assessment as a capability scorecard, a Chinese model, evaluated by Western governments, performing at a level worth flagging on the international ledger.

Marcus does not dispute that reading. His open letter to Sacks, published on Substack, runs in a different direction. He argues that the report's existence, and the political mileage taken from it, are beside the real question: whether the United States should organize AI policy around a sprint to outpace China at all. The disagreement is about which question to ask, and in AI policy the question is often the policy itself.

What "closing the gap" means, in this exchange, is the load-bearing term. Sacks uses it the way a competitor uses a scoreboard: benchmarks move, deployment moves, and a single-model assessment is one data point in a longer trend line. Marcus uses it the way a diplomat would: if AI is a coordination problem, with international safety norms, shared standards, and joint red-teaming, then "winning" is the wrong verb. His earlier essay, "China has all but caught up. The US is not going to win the AI war," makes the same argument without the personal address.

The reaction cluster around the report runs in more than two directions, and the public disagreements help locate the actual fault line. A roundup in The Next Web frames Sacks, Vinod Khosla, and Marcus as a single thread, useful for spotting who shows up and how, but not for adjudicating what the underlying document says. Sacks's post, according to Marcus, was itself a reply to a tweet from Commerce Secretary Howard Lutnick, which means the policy reading arrived in the public conversation through a chain of second-hand summaries rather than from the primary report.

Three signals will tell readers which approach is winning in practice. First, whether the AISI/CAISI joint assessment produces a formal US government response — a policy acknowledgment, a directed review, or a public position from a relevant agency — in the next policy cycle. Second, whether the next round of China-related export controls tightens or loosens the cyber-capability category the report flags. Third, whether the small set of international AI safety talks that have run since 2024 expand or stall before year-end.

A single capability report, in this light, is less a verdict on the AI race than a snapshot, taken at a moment when policymakers and commentators were already arguing about what kind of race, if any, to run.