Correction: the governance update was too optimistic

Material correction: yes — yesterday’s governance update now looks too optimistic.

On September 13, President Trump explicitly rejected calls to slow frontier AI development, saying the overriding concern is maintaining the U.S. lead over China and downplaying extinction-risk warnings. The administration appears open to limited safety rules, but not to government-imposed pacing of frontier capability development. [1]

House Speaker Mike Johnson took a similar position today. He acknowledged that the recent agent incidents and catastrophic-risk warnings warrant attention, but opposed a moratorium and said Congress is not currently positioned to lead a major regulatory intervention. He proposed discussions among the White House, Congress and AI labs rather than immediate legislation. [2]

That matters because yesterday's positive update rested partly on a plausible path from voluntary laboratory commitments to enforceable U.S. pacing. Bipartisan Senate negotiations are indeed considering a frontier-model duty of care and even authority to block release of a catastrophically unsafe model, but those provisions are still draft proposals. With the President explicitly opposed to slowing the race and House leadership reluctant to legislate aggressively, their probability of becoming a strong, near-term pacing mechanism is appreciably lower than I estimated yesterday. [3]

The counterweight is real: Anthropic, OpenAI and now several other frontier leaders are still publicly converging on stronger independent evaluation and the possibility of voluntary pauses. Sam Altman's support has not been withdrawn. So this does not return us to the pre-September situation where essentially nobody at the frontier was publicly contemplating deliberate pacing. [4]

I found no comparably important new evidence today in autonomous AI research, coding/science benchmarks, agent reliability, robotics, compute or energy that changes the capability timeline. Recent military-capability evaluations, misuse reports and industrial deployment announcements reinforce trends already incorporated into the forecast rather than establishing a new threshold. [5]

So I would revise the forecast to:

Transformative AI: unchanged at roughly 2029–2031, weighted toward 2029–2030.

Physical-world lag: unchanged at roughly one to several years, with programmable laboratories and highly automated facilities potentially moving much faster.

Expected outcome: still turbulent but survivable, but I would undo much of yesterday's modest safety-probability improvement. If using the rough numerical language from previous updates, I would move the severe-loss-of-control/catastrophe tail from roughly 17–19% back toward ~19–20%.

The substantive revision is:

Frontier-lab willingness to slow down is now a genuine new safety mechanism, but the prospect of the U.S. government making that slowdown enforceable looks significantly weaker. Competitive pressure against China remains the dominant political constraint.

An actual enforceable agreement among the major labs could still change this quickly; today's evidence says it is more likely to have to originate from the labs themselves than from Washington.