AI 2040 supplement: Comparing Possible Plans (Lifland, AI Futures Project, 2026)

This is a supplement to AI 2040: Plan A, written by Eli Lifland of the AI Futures Project, with additional per-author estimates from Ryan Greenblatt and others. Where the main document sets out Plan A and names four rejected alternatives, the supplement scores them against explicit metrics. Lifland states he is "very much not confident in my precise estimates" while being "confident in the conclusion that Plan A is substantially better than Plans B-D."

The plans compared

Estimates are conditional on the government or leading companies attempting the plan; the taxonomy is not offered as exhaustive.

• Plan A — verified slowdown plus total research transparency. A verified international slowdown deal with complete transparency of AI research, scaling capabilities while maintaining confidence in safety and balancing risks from deal decline and covert projects. To be classified as Plan A more generally, an implementation must involve a substantial verified slowdown with continued scaling — not an indefinite halt — and transparency at least as strong as embedded auditors from foreign governments publishing reports to governments with redacted public versions.
• Plan B-kinetic — sabotage China kinetically and burn some lead, with willingness to escalate to drone or conventional missile strikes. Requires intent to use the lead for safety and a takeoff slowdown of at least 3 months.
• Plan B-cyber — the same without willingness to escalate to large-scale kinetic attacks; cyber and supply-chain attacks instead.
• Plan C+ — domestic regulation plus slowing China without sabotage, via export controls and green cards for Chinese talent. Requires at least a 2-month slowdown.
• Plan C — the leading project burns some of its lead on safety, possibly coordinating with other frontier projects, without domestic regulation. Requires at least a 1-month slowdown.
• Plan D — race. Frontier projects run through the intelligence explosion at nearly maximum speed, devoting at least 1% of resources to safety; below that would count as Plan E.

Two simplifying assumptions are stated: that the minimum-length takeoff from Automated Coder to Top-Expert-Dominating-AI is slightly under a year, from the start to the end of 2030, matching the default Plan A trajectory in the AI Futures Model; and that all plans begin implementation in early 2029, which does not matter for Plans C and D since those involve doing nothing until takeoff is under way.

Plan assessment table

Lifland's median-implementation estimates:

The source presents the two Plan B variants as separate columns for the objective metrics but collapses to five columns for the outcome rows, so the outcome percentages align to Plan A, Plan B, Plan C+, Plan C and Plan D; the table above preserves that structure with the unmatched Plan D cells left empty. Lifland notes that plans rank alphabetically on objective metrics with a large gap between A and B, but that C+ and C beat B on some subjective metrics: Plan B's security focus and "wartime vibe" could degrade epistemics, and its likely nationalization or pseudo-nationalization would reduce public transparency and concentrate power.

Plan likelihood

Lifland's decomposition: a 35% chance the government does something at least as intense as Plan B, on the reasoning that government awareness has been increasing and the importance of ASI will become more obvious, though Plan B requires an actual 3-month slowdown; 50% chance of a deal conditional on that intensity, since "deal is clearly rational but there are various blockers in practice"; and 30% chance the deal is Plan A rather than Plan S, a lower-transparency verified slowdown, or a weaker deal. This yields roughly 5% for Plan A and 10% for other deals, with Plan B at 15% and other intense non-deal outcomes at 5%. He rates Plan C+ as constrained by its 2-month slowdown requirement and "a bit of a narrow target," Plan C as more likely than C+, and Plan D as somewhat more likely than Plan C, with "Other" at 25%. Greenblatt records uncertainty about what should count, reading Plan C+ as mostly "worlds with dumb regulation that slows down AI," and notes Plan B would rank higher for him but for the requirement of intent to spend the lead on safety.

Outcome adjustments

Lifland adjusts his conditional p(alignment) figures for the fact that his median takeoff is roughly twice the length assumed, that he expects later implementation than early 2029, and that the assessment used median rather than expected implementations — each pushing estimates toward 50%. The adjusted values are Plan A 72→70, Plan B 50→55, Plan C+ 45→50, Plan C 40→45, and Plan D 25→32, with "Other" rated at 55, the same as B, on the reasoning that roughly 40% of that bucket is a non-Plan-A deal while it also contains Plan E. He notes Plan A is less sensitive to default takeoff length than the others, and that from a covert-project perspective the direction of the effect is unclear because larger effective compute gaps require more software progress.

Greenblatt cautions that "these numbers are very sensitive to how you classify plans," estimating that Plan A as described in the scenario carries roughly three-fifths the risk of Plan A as classified by the supplement, that a well-executed Plan C by the leading company approaches his Plan B numbers, and that a large fraction of "other" is worlds with too little safety effort to qualify even as Plan D.

Extended plan family

The supplement catalogs further plans, including several the authors judge potentially competitive with Plan A:

• Plan S — an indefinite halt on frontier capability progress intended to last at least a few years, with variants differing on the resumption condition (alignment progress, lie detectors, human uploads, or intelligence enhancement). Advantage: a longer expected slowdown and more margin for error on scaling speed. Disadvantage: scaling to controllable AIs within the human range is itself helpful for accelerating alignment and control, epistemics, verification, and general understanding.
• Domestic-first Plan A — regulate domestically enough to reduce takeover risk to acceptable levels, transitioning to Plan A later if other countries do not follow. Advantage: achievable unilaterally by the US, extends the timeline, and builds trust with China. Disadvantage: worse for covert-project detection and verification setup than negotiating Plan A directly, and possibly not politically feasible under race dynamics.
• GPU arms control — international agreements limiting GPU stock or flow, analogous to historical arms-reduction agreements.
• Also named: Plan B-Kinetic, Plan B-Cyber, Plan C+, Plan E, and CERN for AI.

Provenance

Retrieved July 31, 2026 from ai-2040.com using the vault's bin/fetch-source.py; Firecrawl was rate-limited during this cycle. The assessment table on the live page is interactive, with per-cell reasoning revealed on hover and a Median/Competent toggle; the captured values are the median-across-implementations view, and the per-cell reasoning behind each estimate was not retrievable in a static capture.

Relationships

• depends-on: AI 2040: Plan A (AI Futures Project, July 2026) — the main document this supplement scores.
• related: AI 2027 — the earlier scenario Plan A responds to.
• related: Eli Lifland, Ryan Greenblatt, AI Futures Project — the author and contributors.
• related: Compute Governance — the safety-compute and GPU-arms-control metrics are compute-governance instruments.
• related: AGI Timelines — the takeoff-length estimates are the supplement's principal quantitative axis.
• related: AI Race Dynamics — the plan taxonomy is organized around whether and how a race is slowed.