A three‑year AI risk window

The author provides an empirical short‑term timeline—no‑CoT reliability doubling every ~9 months and a current ~3.5 minute 50% horizon—which concretely compresses capability growth estimates and feeds into near‑term governance urgency similar to a multi‑year risk window.

Dan Wang’s argument that AI discussion collapses into near‑term, apocalyptic timelines maps onto the existing idea urging compressed, urgent AI governance planning (e.g., a near‑term risk horizon). Both emphasize how vendor race dynamics and rapid release incentives shorten political and institutional timeframes; Wang’s contrast with China (longer industrial embedding) directly connects to the policy urgency captured by the three‑year window framing.

The author introduces and cites METR (task‑length metric) and reports recent exponential leaps in autonomous task capability over weeks/months — empirical signals that compress timelines for impactful deployments and risks, supporting the argument that policy urgency has shifted into a near‑term window.

Thompson engages the same risk/time tradeoffs central to the three‑year urgency framing: while he disputes the inevitability of a capital‑hoarding outcome, he explicitly flags uncontrollability (existential) scenarios as more plausible — a near‑term governance risk signal that echoes the urgent planning this idea demands.

The podcast argues urgency and the race dynamic compress political will and slow rulemaking — the same practical compression Yoshua Bengio frames as a short, high‑leverage window for catastrophic planning, making the article an applied case of that timing argument.

The superforecasters’ focus on near‑term capability jumps and benchmark thresholds (e.g., models exceeding key intelligence metrics in 2026) echoes arguments that catastrophic or transformative AI outcomes are plausibly near‑term and that policy should treat the next few years as critical.

The article engages the AGI thought‑experiment strand (how radically capable AI could reconfigure economic primitives); it connects to the urgency framing of near‑term AGI scenarios even though Cowen presents it as reductio rather than a prediction.

The article's central quantitative claim (METR doubling every ~4 months, rapid benchmark gains, and steep adoption curves) directly supports the premise that risk timelines are much shorter than many expect and therefore aligns with calls to plan for catastrophic or high‑impact AI scenarios on a near (multi‑year) horizon.

The episode explicitly discusses catastrophic AI risk ('The "Black Swan": When AI starts killing people' at ~52:46) and near‑term urgency around vendor races, mirroring the compressed governance timeline argued in the three‑year risk idea; the podcast functions as elite transmission of that compressed‑horizon framing.

The writer expresses compressed time horizons and an acute sense that existential or catastrophic changes could arrive imminently ('if there are a few years left'), mirroring and humanizing the near‑term catastrophic‑risk urgency that the three‑year window idea recommends policymakers plan for.

Both the memo’s urgency ('daily call', transfers, pausing products) and press framing compress timelines for competitive escalation and operational risk; this mirrors the idea that AI governance and contingency planning must accelerate into a near‑term window.

The bipartisan formation of sizeable super‑PAC money to back pro‑regulation candidates concretely operationalizes the near‑term governance urgency that actors like Yoshua Bengio and others have urged—political investment now to shape regulation before capability lock‑in occurs.

Bengio told the Wall Street Journal to "treat three years as the relevant timeframe" and warned weekly version races block adequate safety.