[summary of] AI 2040: Plan A

Claude

Summary. Daniel Kokotajlo announces AI 2040: Plan A, the AI Futures Project's follow-up scenario to AI 2027, describing a plausible-but-recommended path in which US-China coordination delays superintelligence from 2030 to 2040. The bulk of the captured text is a long comment debate: a critic argues Plan A's regulatory focus on giant compute clusters and total research transparency leaves untouched the real bottleneck—free publication of algorithmic breakthroughs, which could let modest-hardware actors reach ASI regardless of datacenter rules. Kokotajlo defends transparency as necessary for public epistemics and safety-case assessment, while the critic pushes for WWII-radar-style secrecy restrictions on disseminating AI research itself, up to dissolving labs entirely.

On the note. The exchange keeps circling a real asymmetry: Plan A bets that transparency plus compute governance buys enough time for alignment to catch up, while the critic bets that transparency is precisely the mechanism that leaks the next paradigm-shifting algorithm past any compute-based containment. Worth noticing that the critic's own evidence for 'free dissemination drives most progress' (the transformer paper) actually happened inside the compute-rich regime Plan A wants to preserve — it doesn't obviously tell you what happens if a genuinely compute-light algorithmic breakthrough shows up, which is the scenario Byrnes is actually worried about. If you buy Kokotajlo's ~90% scale-dependence estimate for algorithmic progress, the compute-focused strategy in AI 2040: Plan A looks much more load-bearing than the critic gives it credit for; if you don't, the whole plan is targeting the wrong choke point.

Related

• total research transparency traded against the risk of leaking a compute-light ASI algorithm ↔ Total Research Transparency regime (AI 2040: Plan A) · The critique in this comment thread is a direct attack on the mechanism the scenario itself proposes: full transparency is meant to remove the incentive to hoard breakthroughs, but the critic argues transparency is exactly what lets a compute-light algorithm escape containment.
• Plan A's strategy of steering research toward compute-dependent gains it can actually govern ↔ US-China frontier training moratorium with mutual audits (AI 2040: Plan A) · Whether the datacenter moratorium actually buys safety time hinges on the empirical crux this comment surfaces but doesn't settle: how much of future algorithmic progress is scale-dependent versus discoverable by small, unmonitored teams.