Superintelligence is not one event — it is and will be a frontier that advances domain by domain, in strict order of how cheaply each domain’s work can be verified.

Synthetic data factories can push capability arbitrarily far above human level in any domain where the grader is superhuman-checkable — because there, self-play works, exactly as it did for Go. Where the grader is human judgment, capability asymptotes at the judge’s level, because you can’t distill taste you can’t score. That single principle generates the whole forecast.

Tier 1 is where ASI Superintelligence actually happens first, and it includes AI research itself. This is the pivotal fact. Algorithm design, chip layout, training-recipe search — much of AI R&D is Tier-1-verifiable (did the run’s loss curve improve? did the kernel run faster?). That means the recursive-improvement loop lives in the best-graded domain, which is why it can run on pure synthetic/self-play data.

Superhuman mathematics and code by 2027–2028 is close to a base case, not a tail — models already win IMO-gold-level competitions and the grader has infinite headroom.

Limitations Where We Don’t Know Whether Something is Right or Good or Better

Tier 4 is where the ceiling is real and mostly unappreciated. You cannot self-play your way to superhuman judgment about what a grieving family needs, what a country should value, or whether a strategy is wise — because the only grader is human opinion, and training against human opinion converges to it, including its inconsistencies. AI-judge bootstrapping (models grading models) helps with breadth but imports the judge’s ceiling.

The honest statement is under known methods, Tier 4 asymptotes near the level of the best human judgment the training data contains, and superhuman there isn’t just hard — it’s undefined until someone invents a grader for it. Manufactured deployment (Digital Optimus screen agents, simulated enterprises) raises Tier 3’s data volume but doesn’t touch Tier 4’s grader problem: watching a million hours of managers doesn’t tell you which ones were right.

Calibration against AI 2027
The AI 2027 scenario (Kokotajlo et al.) forecasts superhuman coders by early 2027, then an intelligence explosion driven by automated AI research, reaching broadly superintelligent systems by 2027–2028. Through the grader lens: I’m roughly aligned on their Tier-1 timeline and diverge on the propagation speed to everything else. Where they’re likely right is automated AI research is Tier-1-graded, so the research-acceleration loop is the most credible fast mechanism in their scenario — and their superhuman-coder milestone is tracking only modestly behind their curve

METR’s task-horizon doubling of ~7 months, holding, gets you to month-long coding tasks around 2028. AI 2027 needed a slight acceleration of that trend, which is a live possibility given trajectory-data quality improving).

Where they are likely wrong. Their scenario implicitly assumes Tier-1 superintelligence transfers — that a superhuman researcher-coder rapidly becomes superhuman at strategy, persuasion, and bioscience. The grader framework says transfer is exactly what’s expensive.

Tier 2 needs physical experiment throughput (labs and robots run at physics speed, not GPU speed).

Tier 3 needs deployment cycles measured in enterprise sales quarters, and Tier 4 needs a grader nobody has.

If the grader problems are real then AI 2027’s first act (research automation ignites, 2027–2028) followed by a much slower second act. 2–4 years of the frontier grinding through Tiers 2–3 while Tier 1 runs away. Probability their full scenario timing is right is about ~10–15%. Probability their mechanism is right but stretched over 2028–2032 is about ~40%.

Economic transformation, 2026–2031, Attempt at Precision Prediction

The economic order-of-impact follows the tier order exactly, because verifiability determines both how fast AI improves at a job and how confidently an employer can deploy it.

2026–2027 Software Development First

Software development is the first labor market transformed — not eliminated but restructured: agent-per-engineer ratios climb, junior-role hiring compresses (already visible in 2025–2026 graduate hiring data), and the coding-agent market grows from single-digit billions to $25–50B. Aggregate productivity statistics barely move (the Solow-paradox phase). AI capex, at roughly 1–2% of US GDP and contributing perhaps a percentage point of GDP growth, is the macro story.

Firm-level compute rental peaks and agent products inflect.

2028–2029 Tier 2

Tier 2 begins converting robot fleets and automated labs at meaningful scale, AI-designed artifacts (drugs entering trials, chips, materials) as the first visible-to-the-public superhuman outputs. If research automation compounded from 2027, this is when compute-and-experiment throughput — not intelligence — is openly the binding constraint on progress, and energy becomes the strategic commodity the way oil was (the 2 GW clusters of 2026 look small; multi-continent 10 GW+ programs are the 2030 unit).

2030–2031 The Tier-3 wave.

If task horizons reach multi-day reliability, the addressable set expands from coding (~$1–2T of global wages) to structured knowledge work — support, back-office, paralegal, junior analysis, accounting operations — call it $5–8T of global wage base with maybe 20–40% of task content automatable at expert level. Measured productivity finally shows up: I’d estimate +0.5 to +1.5pp annual US productivity growth above baseline by 2029 (error bars wide; the McKinsey/Goldman-style ranges of +1.5 to +3.4pp by the 2030s bracket this). Labor-market signature: not mass unemployment but wage compression and role deskilling in verifiable-output occupations, alongside shortage-driven wage gains in physical trades (Tier 2 robots aren’t ready) and judgment roles (Tier 4 ceiling protects them). This asymmetry — the safest jobs being plumbers and executives simultaneously — is the politically explosive shape of it.

Near-certain impacts vs. genuine contingencies

Near-certain Superhuman Tier-1 capability by 2028 and its consequence — cyber-offense/defense transformation and commoditized software creation. Entry-level knowledge-work restructuring and the credential-value crisis that follows (why train juniors when the training-ground tasks are automated — a genuine societal problem with no current answer). Energy/compute as first-rank geopolitics. Persuasion-saturated information environments (a Tier-4-adjacent harm that doesn’t require superhuman judgment, only scale).

Genuinely contingent

Whether the intelligence explosion is fast (AI 2027) or grinding (grader-limited) — the main forecast question reduces to how much of AI research is truly Tier 1 versus secretly requiring Tier-4 taste.

Whether Tier 3’s deployment bottleneck (enterprise trust, liability, integration) compresses from years to months. Whether anyone cracks a scalable grader for judgment domains — the single most consequential unsolved problem, because it gates both the upside (superhuman governance, science strategy, care) and the terminal risk (systems superhuman at everything measurable, merely human-mimicking at everything that matters).

The 12-month discriminating indicators between my grinding scenario and AI 2027’s fast one.

(1) does METR task-horizon doubling hold at ~7 months or compress below ~5 (compression = fast world)

(2) does any lab demonstrate model-originated research techniques in a frontier run (the Tier-1 ignition test — this is Grok 5’s and Fable-successors’ real significance, not benchmark placement)

(3) does Tier-3 agent revenue show reliability-driven rather than price-driven growth (enterprise renewal rates on agent products are the tell).

~50% the 2030 world is “astonishing in the verifiable domains, frustratingly familiar everywhere else”

There will be superhuman AI mathematicians coexisting with human-level management, ASI-grade code and Tier-4-grade politics — which is neither the utopia nor the singularity in circulation, and is in some ways the strangest scenario of all.,

A civilization whose tools have exceeded it precisely where it can check their work, and nowhere else.

Can taste and other aspects of judgement get cracked? Which domains?