First drafted April 2026; updated July 2026 (METR v1.1, AI Futures Q1, tracker check, structural rewrite).
Insiders and prediction markets keep landing on 2027–2028. Read AI 2027 in isolation and you’d think one ex-OpenAI researcher is driving the whole narrative — he isn’t. Compute extrapolation, METR task horizons, R&D automation models, expert surveys, macro physical limits, and lab-leader quotes all intersect around the late 2020s through different methods.
That overlap is worth taking seriously. It’s also worth asking whether these tracks share unstated assumptions — and why “people around me aren’t worried” often means you’re not talking about the same AGI.
What “AGI” means in these forecasts
Align definitions before comparing numbers.
| Term | What it measures |
|---|---|
| HLMI (AI Impacts survey) | Unaided machines beat humans at every task, cheaper |
| METR time horizon | Hours of autonomous work 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 |
Comparing “2047” (expert HLMI median) to “April 2028” (Metaculus weak AGI) without naming the bar is how timeline debates turn stupid.
When this post says “AGI is close,” I mean something closer to METR / Automated Coder: reliable completion of hours-to-a-week of real work (coding, research, computer use) — not movie omniscience or “passed an exam.” The public usually imagines the latter; the 2027–28 convergence is mostly about the former. That’s the first reason forecasts sound urgent while daily life feels normal.
What AI 2027 is: nearcasting (scenario), not a crystal ball
Scenario ≠ “when AGI”
Karnofsky’s Most Important Century pioneered a different question: not “which year,” but “if transformative AI arrived in a world like today’s, what follows?” — nearcasting / scenario.
AI 2027 (Daniel Kokotajlo, Eli Lifland, et al.) is the most detailed instance: a month-by-month narrative from 2025 through late 2027 — unreliable agents to superhuman coders, US–China compute race, alignment degradation, loss of control; two endings (race vs slowdown), five research appendices, 53 trackable predictions (compute, agents, economics, governance, geopolitics).
Key points:
- Scenario ≠ point forecast. Research appendices give ASI median 2028.4 with 80% CI 2027.6 – >2100; the narrative reads like “March 2027: algorithmic breakthrough” and makes you forget the wide band.
- Still extremely valuable: forces concrete causal chains — if superhuman coders appear, then open mini versions, then millions of copies doing research, then what? Harder to hand-wave than “AI might be dangerous.”
- Relation to this post: the six methods below extrapolate when; AI 2027 asks what the world looks like if the timeline goes like this. Complementary, not interchangeable.
Reality check: AI 2027 Tracker (June 2026)
Johannes Haus’s independent tracker scores all 53 predictions (not affiliated with AI Futures Project):
| Metric | Value |
|---|---|
| Speed ratio | 0.70× (reality ≈ 70% of scenario pace) |
| Confirmed | 16 / 53 |
| Ahead of scenario | 3 |
| Behind | 4 |
Running hot: agent shape (unreliable but useful, expensive, sold as assistants), METR horizon doubling (scenario assumed ~4 months; reality Ahead), coding-agent revenue, Cybench-style security benchmarks, AI-for-AI-research as industry mantra.
Running slow or unverified: 10²⁸ FLOP training run (frontier still ~10²⁶·⁵–10²⁷), R&D multiplier 3×+ (1.5× on track; 3×/4× not yet testable), “superhuman coder” and 200k parallel coding agents, +30% market / $3T valuation targets Behind.
Directionally right; uneven clock. At constant 0.70×, scenario takeoff shifts from late 2027 to mid-2029; Kokotajlo’s post–Q1-update Automated Coder median ≈ mid-2028 (if reality stays ~65% of scenario pace).
Filters I keep when reading the scenario (no separate critique post needed)
What AI 2027 does best for me: concrete alignment-failure mechanisms (Agent-2→4 degradation), operational US–China detail, how one decision forks endings. I read with these filters:
- Narrative turns conditionals into fate. Appendices give probabilities on unintended goals; the main text reads like deception and loss of control are guaranteed.
- Economics and social psychology barely appear. Superhuman coders in 2027 would move white-collar labor and asset prices before 2028 — Ghost GDP-class transmission is almost absent.
- Human psychology may move earlier. Pedersen fits five indicators: four capability curves look roughly linear; only human attention/panic is hyperbolic — while firms already cut jobs on potential not performance (HBR 2026).
These blind spots don’t zero the scenario’s value; they explain why “living on the AI 2027 calendar” and “living on the METR curve” can feel like different dailies.
Six quantitative tracks (extrapolating “when”)
1. Compute / OOM extrapolation
How many more orders of magnitude of effective compute from GPT-4 to AGI?
Aschenbrenner Situational Awareness (2024): ~0.5 OOM/year in raw compute, efficiency, unhobbling → AGI by 2027 “strikingly plausible.”
Cotra bio anchors (2020): community median drifted ~2050 → ~2035–37.
Epoch AI: frontier ~10²⁶·⁵–10²⁷ FLOP; >2×10²⁸ may cap efficiency.
Crux: 2024–25 pretraining slowdown — blip or ceiling?
2. Capability curve extrapolation (my most trusted track)
METR time horizon (v1.1, Jan 2026) — autonomous work duration at 50% success:
| Window | Doubling time |
|---|---|
| All-time | ~196 days |
| Since 2023 | ~131 days |
| Since 2024 | ~89 days (~3 months) |
Naive extrapolation → ~2028 for a ~40-hour work week. Monthly precision unnecessary; still transformative.
80,000 Hours AGI guide — plainer language.
Crux: durable RL/agent regime or one-time unhobbling?
3. R&D automation feedback loop
Not just compute or task length — AI automating AI research → compounding (AI 2027 / AI Futures Model core move).
AI Futures Q1 2026 update: METR v1.1 + Opus 4.6 + Claude Code pulled Automated Coder median from late 2029 to mid 2028 (Kokotajlo).
METR simplified 8-parameter model more conservative: >99% R&D automation by end 2032.
Crux: real compounding multiplier or parallel coding uplift with human-gated research taste?
4. Expert surveys and prediction markets
AI Impacts 2023 ESPAI (n=2,778): HLMI 50% by 2047 (most conservative formal track); 10% by 2027.
Metaculus (Jun 2026): weak AGI median April 2028; full AGI median October 2032.
RAND AGI Forecasting Synthesis (2025): prepare a range, don’t pick a year.
Crux: insiders and markets moved since 2023; expert 90th-percentile tails (~2150) barely budged.
5. Economic growth / physical limits
Karnofsky / Roodman: ~2%/year GDP can’t run millennia → 21st century must stagnate, explode, or collapse. Karnofsky: >10% transformative AI by 2036; ~50% by 2060.
Different path, similar urgency direction.
6. Macro skepticism (anchor)
Acemoglu: ~5% of tasks profitably automated over a decade — “imminent AGI” as VC narrative, not macro reality. Slow-tail anchor, not a METR dismissal.
Two dimensions people skip
7. What lab leaders are saying
Not a sixth quantitative model. No error bars, not falsifiable — but moves capital, hiring, and state strategy:
| Who | Public line (approx.) |
|---|---|
| Dario Amodei | 2026–27, “country of geniuses in a datacenter” |
| Demis Hassabis | 2028–30 |
| Shane Legg | 50% minimal AGI by 2028 (consistent since 2009) |
| Elon Musk | human-level by end 2026 (outlier) |
FutureSearch tracker: dates oscillate with news — pushed out in 2025, pulled in after agent progress in early 2026. Read as weather vane, not instrument.
8. Takeoff speed — how fast after AGI?
Orthogonal to “which year AGI”:
- Arrival: METR / Metaculus ask when systems do ~week-long real work autonomously.
- After: once systems can edit their own code and run their own research, how long to superintelligence?
| View | Takeoff imagination |
|---|---|
| Yudkowsky (foom) | human-level → far superhuman in weeks–months if self-improvement closes |
| Hanson (slow multipolar) | decades of gradual, multi-lab, multi-stakeholder growth |
| AI 2027 scenario | ~1 year from superhuman coder to ASI |
The 2027–28 convergence answers the first question; the second sets how panicked you should be. Tracker says: agents and METR run hot; 10²⁸ FLOP and 3× R&D multiplier unproven → arrival may be near; explosive takeoff still unknown.
Convergence or common cause?
The late-2020s intersection might just mean everyone assumes scaling + unhobbling continue, capital keeps deploying, capability can decouple from single pretraining FLOP. Break any assumption and the intersection falls apart.
Takeaway: why forecasts sound urgent while people around you stay calm
This is what I most want to leave — not another methods recap.
1. Definition mismatch (most common)
Insider “2028” usually means autonomous real work for days to a week; public “AGI” means smarter-than-human, conscious, can do anything. The first is on an exponential METR track; the second isn’t operationalized. Colleagues say “ChatGPT still hallucinates” — they’re grading chat; you’re reading task horizons. Not lying — different nouns.
2. Brains are bad at exponentials
METR doubling ~every 3 months: “1-hour tasks” to “1-week tasks” takes only a few doublings but spans years on the calendar. Intuition defaults to linear — “last year felt similar” — until a jump feels sudden. Exponentials look reasonable on paper and absurd in lived experience.
3. Gradual iteration + habituation
Tech ships as continuous small steps, not a Judgment Day cutscene. GPT-4 shocked once; now it’s tap water. Claude Code felt like magic three months ago; now it’s “another tool.” Brains re-base in days (hedonic adaptation at civilization scale). Macro curves can steepen while subjective life stays flat — until unemployment, regulation, or accidents force the curve into your face (and Pedersen’s frame says panic itself may be the only truly hyperbolic curve).
4. Reactions price potential, not performance
Layoffs, equity narratives, and policy noise often lead METR measurements — capital prices future distributions, not this week’s benchmark. Researchers see curves up; workers see “company says AI is huge but my tools still suck” — both can be true.
5. Watch distributions, not dates
RAND still wins: prepare Acemoglu-slow, METR-median, and AI-2027-fast worlds. Month-precise dates are narrative devices; horizon doubling time, tracker speed ratio, whether R&D multiplier closes are the dashboards worth watching.
Related
- My AI futures forecast — after capability timelines: institutional lag and three outcome regions
- AI 2027 · Tracker
- AI Futures Q1 2026 Update
- METR Time Horizon v1.1