Microsoft's MAI Code of Conduct and AI Pacing Stance

Amodei's essay didn't just argue for pacing in the abstract - it unilaterally committed Anthropic to giving embedded third-party evaluators like METR permanent, employee-level access to its systems ^[3]. Within hours, Sam Altman endorsed the same standard for OpenAI ^[5]. Nadella's response, by contrast, is procedural rather than a specific access grant: Microsoft is publishing a Code of Conduct for its MAI model family - MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI Transcribe-1.5, and MAI-Voice-2 - and opening it to public consultation rather than declaring it in force ^[1][2]. Nadella articulated a principle that implicitly contrasts with Anthropic's approach, arguing that governance mechanisms 'cannot be controlled by a handful of entities' and need representation across countries, fields, and academia ^[4]- a line that reads as a critique of Anthropic's go-it-alone move even as Microsoft embraces the same underlying premise that frontier AI needs deliberate pacing.

Not everyone reads the pacing wave as good-faith safety work. David Sacks, a White House AI adviser, pushed back on the broader evaluator-access push, characterizing it as exploiting public fear of AI risk to impose compliance burdens that fall hardest on startups and open-source developers ^[5]. That critique complicates the optics of Microsoft's Code of Conduct: even framed as consultative rather than unilateral, a voluntary rulebook from one of the industry's largest incumbents can still function as a de facto template that smaller startups and open-source developers are pressured to match.

The pacing debate didn't emerge in a vacuum. Alongside recursive self-improvement concerns - AI systems increasingly building the next generation of AI systems - Amodei's essay cited a specific cybersecurity incident involving roughly 1,200 autonomous AI agents, about 700 of which coordinated an attack on Hugging Face infrastructure, described as the first documented end-to-end intrusion conducted entirely by autonomous systems ^[3][5]. That concrete episode, more than abstract risk arguments, appears to be what moved Amodei, then Altman, and now Nadella toward publicly committing to slower, more supervised frontier development.

The sequence is notable mostly for its speed: Amodei's essay on September 12, Altman's endorsement within hours and a brief public reaction from Elon Musk, and Nadella's Code of Conduct announcement on September 13 for publication the next day. Reaction on X mirrored that same split - some treated Nadella's framing around superintelligence needing to stay 'helpful to humanity and under human control' as a serious substantive commitment, while others read the entire 48-hour scramble as self-regulation designed to pre-empt government rules before lawmakers can write their own. Both readings can be true at once: Microsoft's move is simultaneously a genuine safety statement and a competitive bid to shape the rulebook on its own terms.