On September 12, Silicon Valley's fiercest rivals did something rare: on the same day, they stood on the same side.

Anthropic CEO Dario Amodei published a 5,000+ word essay on the company website — "We Must Pace the Frontier." This isn't a research paper or a vision statement. It's an action plan with specific execution blueprints. Within hours, OpenAI CEO Sam Altman publicly endorsed it, xAI (now SpaceXAI after SpaceX's acquisition) and Microsoft followed suit. The same day, OpenAI announced it will not IPO in 2026, explicitly tying the decision to its safety commitments.

In a single day, frontier AI safety went from "everyone agrees we should do something" to "someone has actually started doing something." That shift deserves a close look.

Three Steps: From Paper to Badges

Each step of Amodei's three-step plan is more concrete, harder, and more important than the last.

Step One: Embedded Evaluators. This is the most operationally concrete piece of the entire plan. Amodei proposes that AI labs grant independent third-party evaluators employee-level access — including badges, laptops, internal system permissions. These evaluators can move freely inside the lab, audit safety processes, probe model behavior, and publish their findings publicly at any time.

This is fundamentally different from existing "red-teaming." Red-team exercises are typically time-bound, scope-limited projects that end when the test ends. Embedded evaluators represent a long-term, continuous, independent oversight mechanism. Their very presence is a deterrent — if a lab knows independent evaluators are watching its code and decisions daily, behavior changes on its own.

Anthropic committed to unilaterally accepting embedded evaluators first, meaning it won't wait for industry consensus — it'll experiment on itself. This is a clever strategy: if the experiment succeeds, other companies face enormous public pressure to follow. If they don't, the optics are devastating.

Step Two: Democratic Industry Coordination. Amodei proposes tighter safety information-sharing and coordination mechanisms among frontier AI labs — but only among labs in "democratic nations." The core obstacle here is U.S. antitrust law. Competitors sharing safety information and coordinating policy could easily be construed as anticompetitive "collusion." Amodei explicitly calls on Congress to provide antitrust exemptions for AI safety coordination.

It's a bold but necessary demand. Without antitrust exemptions, safety coordination among frontier labs remains "voluntary" — toothless and unenforceable. Amodei is essentially saying: if you truly care about AI safety, carve out a legal exception.

Step Three: Global Coordination (Including China). This is the most controversial and most visionary step. Amodei acknowledges that excluding China from AI safety discussions is unrealistic — China's AI research capabilities mean any global safety framework without China is incomplete. He proposes exploring a coordination mechanism that includes China, even though this is extraordinarily difficult in the current geopolitical climate.

OpenAI's Response: Not IPOing Is Safety Commitment Monetized

OpenAI's response on the same day was equally striking.

Sam Altman publicly endorsed Amodei's three-step plan on social media, calling it a "constructive path forward." But the substantive moves were two simultaneous announcements from OpenAI:

First, OpenAI declared it will not IPO in 2026. On the surface, this is a delay. But in the context of Amodei's plan, the meaning shifts. An IPO means profit-maximization pressure, quarterly earnings pressure, shareholder primacy. A company mid-IPO can't make bold safety decisions — because safety spending is a cost with no short-term payoff. By not going public, OpenAI preserves room to "do the right thing" on safety.

Second, OpenAI simultaneously released "A Framework for Safety Requirements for Frontier AI". This document attempts to define: what kind of AI models need what kind of safety review? Where are the thresholds? Who enforces? Details still need fleshing out, but as a substantive response to Amodei's call, it's worth infinitely more than "we support safety."

Why This Time Is Different: From Voluntary to Committed

The AI safety field has never lacked pretty words. Every few months brings another "open letter," "pledge," or "summit." But this time has key differences.

Difference One: Anthropic is acting unilaterally first. Past safety initiatives were always "let's all do this together" — meaning no one was responsible for failure. Anthropic committing to embedded evaluators first is a verifiable, trackable promise. If Anthropic actually brings in independent evaluators and allows public reporting, it becomes the industry benchmark. If Anthropic backtracks, that too will be swiftly exposed.

Difference Two: OpenAI is using its IPO as collateral. The IPO is an enormous temptation for OpenAI — billions in returns. OpenAI choosing to delay and tie it to safety means it's backing its safety commitment with real capital opportunity cost. This isn't vocal support; it's financial skin in the game.

Difference Three: The plan includes specific mechanisms. Embedded evaluators get badges, laptops, access, and publish rights — that's concrete. Antitrust exemptions require Congressional action — that's concrete. Global coordination includes China — that requires diplomatic breakthroughs but at least defines a target. Compared to the vague "we will commit to safety" pledges of the past, every step in this plan has an actionable pathway.

Challenges Ahead: How Far Can This Plan Go?

Even under the best circumstances, Amodei's three-step plan faces massive execution challenges.

The embedded evaluator program's first problem is funding. Who pays the evaluators? If labs pay, how is independence guaranteed? If government or an independent fund pays, where does the money come from? Amodei doesn't provide clear answers.

The second problem is evaluation standards. Evaluators can "publish findings" — but what counts as a "finding"? What constitutes a "safety risk"? Who defines these standards? Without unified criteria, evaluator reports risk devolving into a PR game of dueling narratives.

Antitrust exemptions face enormous uncertainty. Congress can barely reach consensus on any AI-related legislation, let alone greenlight "collusion" among competitors. Even under a broad framework where both parties support AI regulation, antitrust exemptions will face fierce opposition — they strike at the core principles of American competition law.

The difficulty of global coordination (including China) is self-evident. U.S.-China AI competition intensifies daily, with fundamental disagreements on AI safety standards. Reaching substantive agreements between the two nations on AI safety may take far longer than Amodei imagines.

But even so, the plan's value lies not in whether it's fully implemented, but in the direction it defines. Past AI safety discussions were abstract — "what should we do" and "what principles do we need." Amodei pulled the discussion into specific mechanisms — "who does it," "how it's done," and "to what degree." Even if the final implementation differs from Amodei's proposal, this plan sets the baseline for all future discussion.

September 12 may not be remembered as "AI Safety Victory Day." But it may well be seen as the inflection point when AI safety moved from slogans to execution. When Anthropic first lets independent evaluators walk through the door, when OpenAI uses an IPO delay to express safety commitment, when industry competitors start discussing coordination instead of going it alone — these details are more persuasive than any manifesto.

The execution era of AI safety has only just begun.

See you tomorrow.