Kratsios on U.S. AI Strategy: Innovation Over

There is a particular genre of Washington reassurance tour, and Michael Kratsios delivered a polished version of it at Y Combinator's Startup School. The director of the White House Office of Science and Technology Policy sat down with YC's head of public policy, Luther Lowe, to tell a room full of founders that the administration is on their side — against heavy regulation, against state-level patchworks, against anything that hands incumbents a moat. The message was consistent and well-rehearsed. What makes it worth examining is where the seams show.

The Open-Source Moment

The session opened with a live policy question, not a softball. In the days before the event, the startup world had worked itself into a low-grade panic over rumors of an executive order that would restrict open-weight AI models. YC had helped organize a letter to the White House; larger companies sent their own. Lowe — sitting a few feet from Kratsios onstage — noted he'd been one of the recipients on the Wednesday letter and wanted to know how tethered to reality the Twitter anxiety actually was.

Kratsios's answer was calm and direct: "What the secretary said yesterday is the same policy that we had on page one of our AI action plan that was released last July." He described the plan as co-authored with David Sacks and Secretary Rubio — an attribution that, per reporting from the Revolving Door Project on Sacks's role in Trump administration AI policy, tracks with the known principals involved — and said chapter one opens with a commitment to open source on the grounds that a healthy AI ecosystem requires both open and closed models working in parallel.

The anecdote is clarifying for reasons beyond the open-source question itself. It shows how policy signals get distorted in transit — rumor to Twitter to lobbying letter to onstage clarification — and why Kratsios argues that weeks of industry pressure are actually useful rather than annoying. "If DC is in a vacuum and isn't hearing anything from the startup ecosystem," he said, "we can't make the best decision." That is either a genuine acknowledgment of how policy should work or a savvy reframe of industry lobbying as civic participation. Possibly both.

Don't Draw Lines in Sand

The White House's core regulatory philosophy, as Kratsios articulates it, is essentially: don't make the EU's mistake. The EU AI Act — finalized after years of deliberation — was written before large language models existed as a commercial reality. Everything that has happened since ChatGPT's release now has to be reconciled with a framework designed for a different technological moment. Kratsios cited this as the cautionary tale against hard thresholds.

He applied the same critique to the Biden administration's compute threshold — the specific FLOP count above which AI developers faced mandatory government disclosures. "Once [government] sets a line it's very hard to reset it," he said, "so we are very cautious in trying to set these hard thresholds." The logic is sound as far as it goes: technology evolves, fixed numbers calcify, and a threshold that seemed meaningful at one compute scale looks arbitrary at the next. What the framing sidesteps is the alternative — if you don't draw lines in sand, you don't draw lines. The question of what oversight mechanism does work for fast-moving technology got less airtime than the critique of the ones that don't.

On the risk side of the ledger, Kratsios flagged cyber-capable models — specifically the dual-use problem, where the same model that could do something harmful is also the one best suited to hardening existing systems. He was more skeptical about biological risk, describing it as a concern that has been discussed seriously in policy circles for years without materializing into a demonstrable crisis. His position: the infrastructure for proper test-and-evaluation needs to exist and scale with the frontier, but the alarm level at present is higher than the evidence warrants.

Born Free vs. Born in Captivity

The most analytically useful framework Kratsios offered was a taxonomy of technologies by their regulatory inheritance. "Born free" technologies — like the early internet — arrive without pre-existing regulatory apparatus, and the policy instinct should be to preserve that openness rather than rush to impose structure. "Born in captivity" technologies — commercial drones, supersonic flight, autonomous vehicles — require government approval before they can reach the market at all, and the policy task there is to clear the runway.

AI, in Kratsios's view, lands in the born-free category. The implication is that introducing new regulatory frameworks is the move that requires the highest justification threshold. This is a coherent position, and it explains why the administration's legislative asks to Congress have been relatively narrow: preemption of state AI laws (to prevent the compliance-cost patchwork that favors large incumbents over startups) and some statutory clarity on model outputs and intellectual property.

On IP specifically, Kratsios drew a firm line at model outputs — if a model produces Mickey Mouse, that is a Walt Disney intellectual property problem that statute should address clearly — while suggesting that the training-data side of the debate might be better settled by market mechanisms than legislation. Licensing deals between AI companies and content owners are already emerging, and his read is that the market will mature faster than Congress can act competently.

Science at Scale, and Who's Not at the Table

Kratsios also discussed a roughly 150-page report his office recently released — the administration's attempt to rethink how the federal government engages with the science ecosystem. The framing draws on the post-World War II moment when federal investment dominated the R&D landscape. Kratsios argued, using figures he attributed to the report, that the composition of who funds American R&D has shifted dramatically — the federal government now accounts for a minority share while the private sector and philanthropy have grown to dominate — and that this inversion requires a new theory of the government's role.

His answer is the "shape the arena, not direct discovery" formulation: fast grants decided in under a month, prize structures designed to attract private co-investment, AI agents running autonomous cloud lab experiments on hypothesis loops. The vision is genuinely ambitious, and the Genesis Mission's specifics — including 90-day agency action plans and a three-to-one private leverage target — reflect a real attempt to translate ambition into mechanism.

But the framing raises a question Kratsios didn't fully engage: when the government funded the majority of basic research, it also set the direction of basic research. The questions that got asked were, at least in part, publicly accountable questions. When private capital drives the bulk of R&D, it asks the questions that have commercial payoffs, or that align with the strategic interests of the entities doing the funding. "Shape the arena" is an elegant phrase for a government stepping back — but who fills that space, on whose agenda, isn't a question the arena-shaping metaphor answers. Basic research into diseases without large wealthy patient populations, or materials science with long time horizons and uncertain commercial applications, has historically depended on public funding precisely because the market won't chase it.

The Little Tech Problem

The session's most politically interesting tension wasn't between the White House and big tech — Kratsios was notably warm about ongoing antitrust cases against Google, Apple, Facebook, and Amazon, crediting them as a bipartisan project that the current administration has continued. The more interesting friction is structural: the argument that a single national AI standard is better for startups than a state-by-state patchwork is correct on its face, but a national standard also requires Congress to pass something. Congress has not demonstrated notable competence on technology legislation. Executive orders, as Kratsios cheerfully acknowledged, are revocable the moment the administration changes. The durability of "little tech's seat at the table" depends entirely on which administration is running the table.

That's not a criticism unique to this administration. It's the baseline condition of technology policy in a system where the technology moves faster than legislatures, and where the institutional knowledge required to write good rules tends to exist in the very industry being regulated. Kratsios is a genuine believer in the project — his biography, moving back into government at real personal cost, makes that credible — but belief and mechanism are different things. The White House can want startups to thrive. Whether the regulatory environment it builds outlasts this particular set of principals is a harder question.

"The government can either help or they can unfortunately kind of screw things up," Kratsios said at the outset. That is, stripped of the optimism, a fairly accurate description of the range of outcomes. The current approach is betting heavily on the former. The architecture to ensure it is still being built.

Samira Barnes covers technology policy and regulation for Buzzrag.