2027 — Artificial intelligence (predictions only)

❓ What are named parties actually forecasting for 2027, and what would make those forecasts wrong?

Nothing in this chapter has happened. Dates in 2025–2026 that start a clock are noted as clocks.

The capability bet: longer horizons, more autonomy, not a new physics

❓ If METR’s 50% horizon keeps doubling, what does 2027 look like — and if it does not?

PREDICTION (extrapolation): Writers using TH 1.1’s fast 2024–25 slice (doubling ~3.5 months) put a full workday 50% horizon in 2027 and a work-week in 2028. Writers using the long-run 7-month doubling put the same milestones later. Epoch (April 2026) documented acceleration but refused a unique rate. A slowdown once RL is a large share of spend is an explicit alternative in 2026 commentary.

Falsifier: A METR 2027 print showing the 50% software horizon flat or down versus the last 2026 print, on a comparable suite.

PREDICTION (markets, September 2026 X recap): Polymarket/Kalshi-style odds discussed in the 20s for an OpenAI AGI announcement before 2027, higher before 2028. Those are prices, not measurements of AGI. Metaculus medians in the same recap remain later (~2030s). Do not average a market and a survey into “AGI arrives on date D.”

The scenario that ate the year: AI 2027

❓ Is the 2025 scenario document a forecast we should score, or a story we should keep in a drawer?

PREDICTION / scenario. AI 2027 (Kokotajlo, Silver, et al., April 2025) is a month-by-month fictional lab timeline: mid-2027 Agent-3 copies doing autonomous research at large speedup; late-2027 Agent-4, interpretability lies, a US–China compute race, Race vs Slowdown endings. It was widely read. It was also criticized (Gary Marcus, Arvind Narayanan & Sayash Kapoor, Helen Toner, Vitalik Buterin, the LessWrong analysis by titotal, Steve Newman’s Amdahl-law note on 250× research speedup).

X posts in September 2026 claiming “19/24 short-term beats already hit” are takes about a scorecard, not an audit this package performed. This report will not certify that scorecard.

What from the scenario rhymes with 2026 measurement without becoming the scenario: coding-agent share of lab code (Anthropic >80% merged, May 2026) is a real number. That is not Agent-3 running the lab at 50×. Conflating “Claude writes most of the diffs” with “superintelligence is doing the research” is the error.

Falsifier for the aggressive 2027 takeoff: 31 December 2027, no public evidence of AI systems autonomously improving frontier training runs end-to-end at a large multiple of the 2026 human-paced baseline, and METR-class horizons still measured in hours not weeks.

Power, permits, transformers — the 2027 binder

❓ If GPUs are no longer scarce, what is?

PREDICTION, with 2026 clocks already ticking:

• UNECE/IEA: datacentre electricity almost doubles 2025→2030 (485 → 950 TWh). Grid build is slower than halls. (UN News, 2026-09-08)
• Morgan Stanley, as reported by CryptoBriefing, mid-September 2026: a project that wants to open before end-2027 needed to break ground by October 2026; 300+ US local moratoriums since 2023; Bernstein 35–40% of announced global capacity at delay/cancel risk through 2027; 38 GW US power shortfall through 2028 in that recap. Hyperscaler capex cited ~$785B (2026) and approaching $1T (2027). This is not Morgan Stanley’s own PDF. (CryptoBriefing)
• SemiAnalysis (June 2026 coverage): US grid spare capacity may turn negative by 2027; incremental datacentre demand 21 GW (2026) → 84 GW (2030); gas additions <10 GW in 2026 and 2027; transformer/turbine lead times 3–4 years.
• Gartner (July 2026, via secondary): AI-optimized servers use more power than conventional datacentre hardware by 2027; global AI-datacentre electricity 175 TWh in 2026 in that write-up.

HAPPENED as a bottleneck already: Ireland connection limits, Netherlands siting limits, 48–60 month HV transformer quotes in 2026 industry notes.

Falsifier: US interconnection queues compressing to <2 years and 2027 US datacentre load met without widespread behind-the-meter generation.

Zuckerberg’s line that energy is the biggest bottleneck for scaling is a take that matches the construction clocks.

Recursive self-improvement as a 2027 speech, not a 2026 milestone

❓ When people say “AI is building itself,” what object do they mean?

Three different objects:

1. HAPPENED: models author a large share of a lab’s application code (Anthropic May 2026, >80% merged).
2. PREDICTION: models design the next training run (data mix, architecture, optimizer) with little human research labour.
3. PREDICTION / WARNING: an intelligence explosion (AI 2027 Agent-3/4).

Only (1) has a date. Dell’Oro’s Alex Cordovil: “You can’t certify a moving target” — the datacentre operational problem if training/inference fleets swing tens to hundreds of megawatts. That is a WARNING about power quality, not a proof of (3).

LeCun: no reliable general agents until world models; LLM path is a dead end for AGI. If he is right, 2027 looks like better interns and a possible capex hangover, not Agent-4.

Safety process as a shipping bottleneck

❓ Will 2027 be the year models get held for months, not days?

PREDICTION, seeded by 2026 HAPPENED: Meta delayed Spark max-reasoning for extra tests; OpenAI staged Astra; Anthropic split Fable/Mythos; monitorability cited as a scaling constraint. 2027 could be more gated SKUs, sudden holds, and “limited organizations first.” That is the conservative institutional bet. The aggressive bet is that race dynamics override holds.

Falsifier either way: count of publicly announced multi-month safety delays versus same-week competitive drops, tallied at year-end 2027.

Jobs and “digital colleagues”

❓ Does 2027 turn the 22–25 software dip into a broad white-collar break?

The 20% drop for young software developers is HAPPENED (AI Index). Extending it to “millions of digital colleagues replace median knowledge workers in 2027” is PREDICTION (lab blogs, AI 2027, some All-In/Moonshots speech). Acemoglu-style counter: most jobs still need physical presence, social interaction, or novel judgement; task automation shares stay small. Household robots at 12% chore success argue the atoms side will not follow the bits side on a 2027 clock.

Falsifier: BLS-class series showing 22–25 software employment rebounding through 2027, or, conversely, a clear break in 26–40 professional employment, not just new-graduate hiring.