Long Yi
← Back to all writing

Why everyone converges on 2027–2028 (and why that might not mean what you think)

April 1, 2026

AI SafetygovernanceAI Economy

First drafted April 2026; updated June 2026 with METR v1.1, AI Futures Q1 update, and AI 2027 tracker data.

Read AI 2027 in isolation and you’d think one ex-OpenAI researcher’s scenario is driving the whole “AGI by 2027” narrative. It isn’t. Compute scaling, agent capability curves, R&D automation models, expert surveys, economic growth limits, and lab-leader judgment all land somewhere around 2027–2028 right now — through different methods.

That overlap is worth taking seriously. It’s also worth poking at, because a lot of these tracks may share the same unstated assumptions.

I started pulling this together in April as research notes and turned it into a blog post once the tracker and author updates made the picture clearer. What follows is the map: what each method actually predicts, where they fight, and how reality is tracking against the most detailed scenario we have.

”AGI” is not one milestone

Before the methods — the word is doing too much work.

TermWhat it measures
HLMI (AI Impacts survey)Unaided machines beat humans at every task, cheaper
METR time horizonHours of autonomous task completion at 50% success
Automated Coder (AI Futures)Lab would rather fire all human SWEs than stop using AI
Weakly general AI (Metaculus)Passes 4 public criteria including some robotics
Minimal AGI (Shane Legg)Most typical human cognitive tasks

These aren’t the same bar. Comparing “2047” (expert HLMI median) to “April 2028” (Metaculus weak AGI) without naming the threshold is how timeline debates turn stupid.

Eight methods (not eight people)

1. Compute / OOM extrapolation

Question: How many more orders of magnitude of effective compute from GPT-4 to AGI?

Aschenbrenner’s Situational Awareness (2024) splits progress into ~0.5 OOM/year in raw compute, algorithmic efficiency, and “unhobbling” (chatbot → agent step-changes). GPT-2→GPT-4 was ~4.5–6 OOM in four years. Run the same rate → AGI by 2027 is “strikingly plausible.”

Cotra’s bio anchors (2020) anchor training FLOP to brain-scale estimates. The merged distribution spans 10²⁴–10⁵⁰ FLOP — twenty orders of magnitude. The community median has drifted from ~2050 (2020) toward ~2035–2037 as industry overshoots the lower anchors.

Epoch AI tracks what labs actually spend: frontier training is ~10²⁶·⁵–10²⁷ FLOP as of mid-2026 (~30–100× GPT-4). No confirmed 10²⁸ run yet. Epoch’s newer work suggests data movement bottlenecks may cap efficient scaling past ~2×10²⁸ FLOP on current hardware.

Crux: Is the 2024–25 pretraining slowdown (GPT-5 pretrain ≈ GPT-4.5 scale) a blip or a ceiling?

2. Capability curve extrapolation

The most empirically grounded track. METR’s time horizon benchmark measures how long an agent can work autonomously at 50% success.

Time Horizon v1.1 (Jan 2026):

WindowDoubling time
All-time (2019–2025)~196 days (~6.5 months)
Since 2023~131 days (~4.3 months)
Since 2024~89 days (~3 months)

Naive extrapolation → AI completes a ~40-hour work week by ~2028. You don’t need monthly dates for that to matter.

80,000 Hours’ AGI guide makes the same argument in plainer language.

Crux: Is the 2024+ acceleration a durable RL/agent regime, or a one-time unhobbling that saturates?

3. R&D automation feedback loop

AI 2027’s distinctive move: model not just compute or task length, but AI automating AI research → compounding speedups.

The follow-up AI Futures Model (Dec 2025) initially pushed timelines 3–5 years later than AI 2027 — the authors were less bullish on pre-automation R&D speedups. Then the Q1 2026 update pulled back ~1.5–2 years after METR v1.1, Opus 4.6, and Claude Code’s revenue run:

ForecasterAutomated Coder median (Dec 2025 → Apr 2026)
Daniel Kokotajlolate 2029 → mid 2028
Eli Liflandearly 2032 → mid 2030

Their own check: if reality runs at ~65% of AI 2027 pace, Automated Coder lands ~2028.

METR’s simplified 8-parameter model is more conservative: 2032 for >99% R&D automation.

Crux: Is the R&D multiplier actually compounding, or are we mostly seeing parallel coding uplift while research taste stays human-gated?

4. Expert surveys and prediction markets

Outside view. AI Impacts 2023 ESPAI (n=2,778): HLMI 50% by 2047 — 13 years earlier than 2022, but still the most conservative formal track. 10% HLMI by 2027.

Metaculus (Jun 2026): weak AGI median April 2028; full general AGI median October 2032 — four years apart for stricter criteria.

RAND’s AGI forecasting synthesis (2025): the infrastructure is immature; the policy question is preparing for a range of futures, not picking a year.

Crux: Crowds and insiders moved fast since 2023; expert survey tails (90th percentile ~2150) barely budged.

5. Economic growth / physical limits

Karnofsky’s Most Important Century (2021): ~2%/year GDP growth can’t persist millennia (Roodman 2020). The 21st century has to be stagnation, explosion, or collapse. Karnofsky’s probabilities: >10% transformative AI by 2036; ~50% by 2060.

Different path to similar urgency — economics, not AI insider mechanism reasoning.

6. Nearcasting / scenario

Not “when,” but “if TAI arrives in a world like today’s, then what?” Karnofsky pioneered this; AI 2027 is the most detailed instance — monthly narrative, 53 trackable claims, two endings.

Scenario ≠ prediction. AI 2027 gives ASI median 2028.4 with 80% CI 2027.6 – >2100. The narrative makes you forget the CI.

I wrote a critique of AI 2027 on alignment-as-premise and missing economic psychology. This post is the complement — the whole forecasting landscape, not one scenario’s blind spots.

7. Lab leader insider judgment

Qualitative, but it moves capital. Late 2020s convergence:

FutureSearch’s timeline tracker shows these forecasts oscillating with news — pushed out in 2025, pulled in after Anthropic’s agent progress in early 2026.

8. Takeoff speed (after AGI, not when)

Yudkowsky vs Hanson (2008): weeks-to-months FOOM vs decades of gradual multipolar growth. AI 2027 assumes ~1 year from superhuman coder to ASI. Separate axis from arrival date, but it sets how much urgency you feel.

Convergence vs correlation

The late-2020s intersection might just mean everyone assumes:

  1. Scaling + unhobbling continue (agents, RL, tool use)
  2. Capital deployment continues ($1T+ AI capex trajectory)
  3. Capability can decouple from raw pretraining FLOP (2024–25 evidence)

Break any one and the intersection falls apart. Acemoglu’s macro view is the cleanest pushback: ~5% of tasks profitably automated over the next decade, ~0.05%/year productivity — “imminent AGI” as VC hype, not macro reality.

Reality check: AI 2027 tracker (June 2026)

Johannes Haus’s independent tracker scores 53 AI 2027 predictions. I maintain a full scorecard in my notes (notes/ai_2027_tracker_scorecard_2026-06.md).

Headline numbers:

MetricValue
Speed ratio0.70× (reality ≈ 70% of scenario pace)
Confirmed16 / 53
Ahead of scenario3
Behind4

By category:

CategoryWhat happened
Agent AutonomyStrong: 4/6 confirmed; METR doubling ahead of scenario
Economic ImpactCapex/revenue on track; $3T valuation and +30% market behind
Model CapabilityContinuous training confirmed; 10²⁸ FLOP not yet testable
TakeoffAI-for-AI-research confirmed; 3×/4× R&D multiplier not yet testable
CodingAgents valuable; superhuman coder / 250k parallel agents not yet testable
GeopoliticsExport controls confirmed; China gap larger than scenario predicted

The split I keep coming back to:

Running hot                          Running ~70%
───────────                          ────────────
METR horizons (~4 mo doubling)       10²⁸ FLOP training run
Coding agents / Claude Code          R&D multiplier 3×+
Cybench, OSWorld                     RE-Bench 1.3 target
Cyber capability (ahead)             Market valuation targets

Directionally right. Uneven timing. The takeoff thesis — the part that actually matters for x-risk — is still unproven.

At constant 0.70× pace, takeoff shifts from late 2027 to mid-2029; Kokotajlo’s updated Automated Coder median is mid-2028.

What I’d actually take away

2027–28 is a real intersection of methods, not one blogger’s fanfic. METR, Aschenbrenner, insiders, and Metaculus weak AGI land nearby through independent paths.

But they disagree on definition. Expert HLMI @ 2047 and weak AGI @ 2028 can both be “correct.”

AI 2027 reads best as a stress test. The AI Futures Model is what the authors actually update — and it oscillates.

As of mid-2026: agents and deployment are ahead of pure compute scaling; the full R&D feedback loop isn’t verified.

RAND’s framing sticks: don’t pick a year, prepare across a range — including Acemoglu-slow and AI-2027-fast tails.

Where I land (for now)

I trust METR time horizons more than anything else on this list. It’s measured, updated, falsifiable. Extrapolate it naively and you get “AI does a work week autonomously by ~2028” — transformative even if HLMI is decades out.

I’m skeptical of precise monthly dates from any source, AI 2027 included. The tracker confirms the scenario’s shape more than its clock.

My crux right now: are 2024+ capability gains mostly compute-efficient agent paradigms that keep compounding, or a one-time unhobbling that saturates? METR v1.1 and Claude Code point repeatable; Epoch’s 10²⁸ bottleneck and RE-Bench lag point saturate. I genuinely don’t know.

If you forced me to bet: automated coding at “company would rather lay off SWEs” level ~2028–2030; HLMI-style full labor automation noticeably later; takeoff speed still the biggest unknown after the first coding threshold.

That’s a distribution. Not a date.