Y Combinator Startup School

Michael Kratsios on Why AI Rules Shouldn't Be Set in Stone

Michael Kratsios· Director of the White House Office of Science and Technology Policy at White House Office of Science and Technology Policy
·~40 min·English·Y Combinator
PolicyOpen SourceAI Company
TL;DR

The White House's top science and tech advisor argues that in a field reinventing itself every six months, U.S. AI policy should back both open and closed models, avoid firm rules that age out, and keep the lane clear for startups over incumbents.

01How Washington Works

There Is No Technology Department

<strong>There is no U.S. technology department</strong> — AI policy is hammered out across multiple agencies with different equities, and the White House brings them together.

these decisions are extraordinarily federated. There isn't just like one or two people in the White House who decide something.

— Michael Kratsios, Y Combinator Startup School
Key Insight
Read against the news cycle, this reframes 'slow' or 'wishy-washy' AI policy as structural: no single official can issue a ruling, so every position is the negotiated output of security, commerce, and science agencies that each hold a veto.
02Open Source

Open and Closed, or Neither Wins

The official U.S. position is not open versus closed: <strong>American AI leadership needs a vibrant open-source and closed ecosystem together</strong>, and that is page one of the strategy.

if the US wants to lead in artificial intelligence, we have to have a vibrant closed and open source ecosystem, and that's the only way they can all work together.

— Michael Kratsios, Y Combinator Startup School
Key Insight
Naming open-source support on page one of the strategy is a hedge against pressure from within his own coalition — it pre-empts the recurring push to restrict open weights on security grounds by making openness a stated pillar of competitiveness, not a concession.
03Regulating a Moving Target

Rules Must Move With the Frontier

Kratsios's core rule for fast-moving tech: <strong>don't draw firm red lines, because rules like the EU AI Act lag the frontier</strong> — drafted before ChatGPT, they scramble to cover models that arrived mid-process.

you don't want to set very firm red lines in the sand because they ultimately don't work.

— Michael Kratsios, Y Combinator Startup School
Key Insight
The deeper claim is regulatory lag: the EU AI Act was proposed in April 2021, ChatGPT launched in November 2022, and the Act was only agreed in December 2023 — forcing lawmakers to bolt on general-purpose-model rules mid-negotiation. Kratsios argues the Biden compute-threshold cap risks the same trap.
04Little Tech

One National Standard, Not Fifty

A fifty-state patchwork of AI rules is survivable for Google's lawyers but fatal for startups, so <strong>the White House wants one national standard</strong> anyone can learn and build against.

we have to have one national standard for AI so we can make it easy for anyone who wants to build an AI company to know what the one set of rules is and build their company like that way.

— Michael Kratsios, Y Combinator Startup School
Key Insight
Preemption is usually framed as deregulation, but here it is pitched as pro-competition — a single federal rule is the one thing that neutralizes big tech's compliance-cost advantage over a two-person startup.
05A Framework for Regs

Born Free vs. Born in Captivity

Every technology is either <strong>born free or born in captivity</strong>, and each kind gets the opposite regulatory playbook — preserve the first, unshackle the second.

the born in captivity technologies are technologies where you're building something but you can't commercialize it or you can't take it out to market unless you get some sort of government approval.

— Michael Kratsios, Y Combinator Startup School
Key Insight
The framework quietly resolves the 'is AI over- or under-regulated?' fight: it depends which kind of technology a given AI use resembles — an ungoverned frontier to protect, or a gated market to pry open.
06Risk

The Same Model Cuts Both Ways

The hardest AI risk has no clean fix because <strong>the same model that can attack a system is the one that can defend it</strong> — you can't ban one edge without losing the other.

the same model that is able to sort of do something nefarious is the same model that can be very valuable in hardening an existing system.

— Michael Kratsios, Y Combinator Startup School
Key Insight
This is why 'just ban dangerous models' is a non-answer: dual use is a property of the capability itself, so prohibition cannot separate offense from defense — the practical lever is test-and-evaluation infrastructure, not a ban.
07The Next Frontier

Shape the Arena, Don't Direct Discovery

In the AI era of science, <strong>the government's job is to shape the arena, not direct discovery</strong> — funding autonomous labs that run experiments on a loop with no human in between.

The idea that you can have autonomous cloud labs running experiments on a loop without human intervention, testing hypotheses, running the experiments, seeing the results, creating a new hypothesis, testing it and running it, and ultimately getting to a conclusion, that is in our sights.

— Michael Kratsios, Y Combinator Startup School
Key Insight
The 1945 Vannevar Bush parallel is doing real work: it signals a once-in-a-generation reset of the federal research contract, prompted less by a war than by the fact that the private sector plus philanthropy now fund roughly 70% of R&D and AI is collapsing the experiment loop.