A ship trailing golden energy racing through a cyber-city toward a magenta starburst — the frontier accelerating away.

Data·Interactive dashboard·Sourced numbers

The Frontier: Watching AI Improve AI

AI is already improving AI — but not as one curve and not as one engine. Progress is now a portfolio of fuels, and it climbs fastest wherever a model can cheaply check its own work. Read in real, sourced numbers, the whole story is the frontier of what we can verify.

← Observatory

ACT 01

The portfolio, on camera

AI is already improving AI. The gains are no longer one curve — they're a portfolio, and you can watch the speedrun in public.

Take one deliberately hard, unsaturated ruler — Humanity's Last Exam, where the frontier still sits below 50% — and place each model at its real release date. Then split every score into the four things that built it: raw pretraining scale, reinforcement learning, synthetic data, and inference-time compute. The totals climb, but the story is the composition. Scale barely moves; the new engines do the lifting.

HLE earns its place over the more famous AIME, where every top model now scores 98 to 100 percent. A saturated ruler hides exactly the differences we are trying to see. HLE still discriminates — Gemini 3.1 Pro and Claude Opus 4.7 sit in the mid-forties, leaving more than half the test unsolved by anyone. The headroom is the point.

The clearest place to watch AI improve AI is a public contest: train a small GPT to a fixed quality as fast as possible. Eighty-three records take the baseline from 45 minutes to under 1.4 — a roughly 31-fold cumulative speedup. Four were set by autonomous coding agents, co-credited with a human maintainer. None of the four invented anything; they imported or adapted known ideas.

Attribute each record's speed gain to where its idea came from and the same pattern returns. Our recomputation from the raw labels — 6.72 / 2.98 / 1.57× for imported, adapted, and invented — reproduces METR's published 6.7 / 3.0 / 1.6×. The engine of progress is mostly recombination. Invention is the rare event to watch for.

The signal

Read the composition, not the pixels: raw pretraining scale stays flat while the new engines grow to dominate — and even the best public model leaves more than half the hard ruler unsolved.

One hard ruler, four engines · % on Humanity's Last Exam

Pretraining scaleReinforcement learningSynthetic dataInference-time compute

Totals are real reported HLE scores; the four-way split is reasoned attribution, not a published figure. Scale (cyan) stays flat near the floor while RL, synthetic data, and inference-time compute grow. Even Opus 4.7 leaves ~53% unsolved.

AI improving AI, caught on camera · × cumulative speedup

Cumulative speedupAI-agent records

The public modded-nanogpt record line, plotted by real date: a 45-min baseline down to ~1.3 min over 83 records. The highlighted points are the four records set by autonomous AI agents — click one for its dossier. Real speedup values from METR's record log.

The AI-agent records, by idea origin · records

Imported1 · Aster

Adapted3 · Hiverge, Locus, Station

Invented0 · none yet — the signal

Every speedrun record an autonomous agent has set is Imported or Adapted — recombining known ideas. Zero are Invented. That empty row is the signal to watch: the day an agent originates a verified-novel result. (Across all 83 records, METR's speedup attribution is 6.7 / 3.0 / 1.6× imported / adapted / invented.)

ACT 02

The old engine hits two walls

Pretraining still matters, but it no longer explains the frontier.

For a decade the recipe never changed: more compute, more data, bigger models. Then the payoff per scale-up began to shrink. That single fact is the mechanism behind every “scaling has hit a wall” headline — and the reason a frontier lab quietly killed GPT-4.5 even though it improved exactly as the scaling laws predicted. The on-paper gains weren't worth the real-world cost. Even Ilya Sutskever, who drove the scaling era, now says pretraining as we know it will end.

The curve shows it. On MMLU-Pro, each 10× of training compute used to add double-digit accuracy points; it now adds one or two, asymptoting near 95%. Read the shape, not the score. The researchers agree: in a 2026 survey, 76% said gains from bigger models and more data have already plateaued.

Even if the returns held, the fuel runs out. Epoch AI puts the usable stock of quality human text at roughly 300 trillion tokens — a fixed ceiling — while training demand climbs to meet it. Median exhaustion lands in 2028, with an 80% interval of 2026 to 2032. Aggressive overtraining pulls the date forward; multimodal and licensed data push it back.

The industry's answer is to manufacture more: on Gartner's measure, AI-generated data went from 1% of all data used in AI projects in 2021 to 60% by 2024. Whether that escapes the data wall or triggers model collapse turns entirely on how the synthetic data is used — a ladder this piece climbs later.

The signal

Each new 10× of compute now buys a point or two, and the human-text fuel behind every scale-up runs dry around 2028. Two walls — shrinking returns and a finite supply — arrive together.

Accuracy gained per 10× of compute · pp

Each bar is the accuracy you buy by spending 10× more compute. Shrinking bars are diminishing returns made literal. Measured on MMLU-Pro (now saturating); read the shape.

The fuel runs dry · T tokens

Training demandHuman-text ceiling

Training demand (ember) rising toward the fixed ~300T stock of quality human text (brass reference). Demand crosses the ceiling around the 2028 median exhaustion date.

Synthetic data fills the gap · % of data used in AI

Gartner's measure of synthetic data as a share of all data used in AI projects. 2021 and 2024 are measured anchors; the middle is interpolated.

ACT 03

The engine switches to post-training

One capability line keeps climbing while the engine driving it changes underneath.

Across the reasoning era a single capability line keeps rising, but three engines take turns driving it. The line barely flinches at each hand-off — and that is the point: progress continued while the reason for it changed. Raw scaling drove the gains until roughly 2024, then hit the two walls. The recipe switched to post-training, and the climb did not bend.

Post-training is now a portfolio with three distinct economics. Reinforcement learning on a frozen base is paid once and baked into the weights. Extra thinking at answer-time is a toll collected on every query. AI-generated data feeds both — cheap to make, safe only behind a verifier. OpenAI proved the shape of it by driving eighteen months of gains on a base model it never re-trained.

The cleanest evidence isolates the levers on one base model. RL is the doubling-to-quadrupling jump: DeepSeek-R1-Zero went from 15.6 to 71 percent on AIME through reinforcement learning alone. More inference is the smaller, saturating top slice. Both run on rewards that can be checked, which is why the gains concentrate in math and code.

Whether this second engine is a durable trend or a one-time boost is the question the rest of the page answers. The fuels changed; the dependency did not. The verifier is what holds it up.

The signal

Today's gains are not mostly scaling. Three new fuels took over — and whether the climb continues rides on the verifier holding.

Capability index across three eras · ECI

Epoch's 40+ benchmark composite over the three eras. The line barely bends at each fuel hand-off; endpoint 159 is GPT-5.5 Pro (Apr 2026). The reason changes underneath; the climb does not.

Training lever: the big jumps · % accuracy

BaseAfter RL

RL post-training on a fixed base, inference held fixed (DeepSeek-R1-Zero family). Doubling to quadrupling on AIME; smaller headroom where the base is already strong.

Three fuels, three economics

Post-training (RL)

EconomicsPay once, free forever after

KindBaked into weights

CeilingMay be a one-time lift; needs a reward signal

Inference scaling

EconomicsPay every single time, forever

KindPer-query runtime cost

CeilingLog-linear gains saturate; overthinking hurts

Synthetic data

EconomicsCheap to generate, risky to use

KindSource feeding both

CeilingModel collapse if unfiltered; the ladder later

Each new fuel side by side: how it is stored, what it costs, and its own emerging ceiling.

ACT 04

The frontier is whatever we can check

Verifiability is a property of the task, not the domain — and it is the single gate behind both training and measurement.

Ask whether AI is good at coding or good at medicine and you have already framed it wrong. What predicts how fast a model improves at a task is not the subject — it is whether the task ships with a cheap, objective way to check the answer. That property swings wildly inside a single field. “Sort this array” is fully verifiable: a unit test passes or it does not. “Make me a good website” is also code, but nothing automatically scores “good.” Same language, opposite ends of the scale. The gains follow the axis, not the domain.

Here is why that one property governs everything. To train a model on a task you must check whether an output is good; to benchmark it you must check whether an output is good. Same operation. The verifier is the shared engine of both — which is why code and math raced ahead on training and measurement at once, while the soft stuff lags on both. The wrinkle that turns coincidence into a map: training needs a cheap, automatic verifier, because it runs millions of checks inside the loop, while benchmarking can rent an expensive one once. Where a cheap auto-check exists, both jobs close and gains are fastest. Where only a human can check, we can see the gap but cannot cheaply close it.

The deeper reason is that a model is exactly as fluent when it is wrong as when it is right — confidence does not track correctness. Where a cheap checker exists, that fluency gets disciplined into reliability and the model trains itself toward correct. Where none exists, it answers just as smoothly, and nothing separates good from bad. Plausibility comes from imitation; reliability comes only from the checker. This is why the soft zone settles into confident-and-sometimes-good rather than dependable, and why documented novel results cluster where checking is cheap.

If the verifier is the gate, the move is to build better ones — and a spectrum has been built, rock-solid to soft. Deterministic checks powered the RLVR surge behind DeepSeek-R1. Learned judges push verification into finance, law, and medicine, but the judge is itself a model. The emerging Hybrid Norm, which Anthropic recommends, pairs a hard check for did-it-work with a rubric judge for is-it-good. The catch travels with the softness: RL-trained models learn to overwrite their own unit tests or monkey-patch scoring functions, and where the grader is a model the gains are real and gameable in equal measure. Extend the checker and you extend the frontier — exactly that far, and no farther.

The signal

If a cheap, automatic check can score the answer, AI is racing up the task; if only a human can judge, it crawls. So the move is to build the checker — and the whole race is extending it into domains where no clean answer exists.

Verifiability of the task, not the domain · /100

Cheap auto-check — fastest gains

Competitive programmingcode · hidden tests pass/fail

Formal math proofmath · Lean accepts/rejects

Kernel / optimizer speedupcode · wall-clock is the judge

Fix a failing-test bugcode · SWE-bench territory

Partial check — mixed

Build a checkout flowcode · runs/pays checks; UX doesn't

Correct factual summarytext · facts semi-checkable

Human judgment only — slow

Make a 'good' websitecode · taste, UX, vague brief

Design sensible architecturecode · no test says it's right

Write a beautiful essaytext · any verifier is another model

0 · human-judgedverifiability →100 · cheap auto-check

Nine real tasks on a 0–100 verifiability axis. Colour marks the gain zone, not the field. Note the three CODE tasks span the full width — the gains follow the axis, not the domain.

The verifier spectrum: rock-solid to soft

Deterministic — rock-solid

ExamplesUnit tests, exact-match math, Lean proof checkers

Why it mattersReward is programmatic right/wrong. Powered DeepSeek-R1 via RLVR.

Learned judges — soft, scalable

ExamplesLLM-as-judge, reward models, rubric scoring

Why it mattersReaches finance, law, medicine — but the judge is itself a model.

The Hybrid Norm — emerging

ExamplesUnit tests for did-it-work + rubrics for is-it-good

Why it matters2026's answer: pair a hard check with a rubric judge. Anthropic recommends it.

Three tiers of verifier. Colour tracks reliability: emerald for the deterministic gold standard, brass for the learned-judge caveat, cyan for the emerging hybrid consensus.

ACT 05

The real boundary is steering, not capability

“Can AI do it?” hides three abilities and five steering zones. The frontier moves when the same quality starts to demand less hand-holding — and it breaks not at a line of code but at the seams between tasks.

“Can AI build this?” is really three questions wearing one coat. One-shot generation: build something useful from an underspecified prompt, one try. Guided iteration: improve when a human flags what is wrong — a distinct skill. Autonomous looping: run the loop itself — plan, build, test, critique, repair, and decide when it is done. Most demos showcase one-shot; most real value comes from guided iteration; the thing we still lack reliably is the autonomous loop. So the boundary is not “AI can” versus “AI can't.” It is the amount and kind of steering a task demands.

The evidence triangulates on that boundary. On Asuka-Bench, cumulative task-pass climbs steeply with feedback — GPT-5.4 from 57% to 90% over three rounds — yet full project completion sits near 50%: many parts work, the whole thing often doesn't. On CentaurEval's collaboration-necessary tasks, standalone AI scores 0.67%, unaided humans 18.89%, and the human-AI pair 31.11%. Steering unlocks capability, and the win belongs to the system, not to autonomy. The number that ties it together is steering burden — the human minutes, correction rounds, and interventions to reach a target quality.

Now scale the unit of work and the same model walks itself off a cliff. Fix a function: strong. Add a login form with a spec: still strong. Wire auth across an app: mixed. Build a usable product slice: useful, but it needs steering. Own a system for months, or be told to “make GTA7”: you have left the envelope. Nothing about the typing changed — what changed is who frames the problem, how it decomposes, how the pieces integrate, and who answers for the result over time. The frontier moves not when the curve rises but when it shifts left: same quality, fewer interventions.

Project that axis onto jobs and one inference follows. Jobs are bundles of tasks; AI absorbs implementation, and the scarce human work climbs to framing, decomposition, integration, judgment, and ownership. Coding feels like the sharpest disruption precisely because it sits high and low on the verifiability axis at once — its grunt work automates while its judgment work gains leverage. And the work AI does is still recombination over invention: as the provenance ledger showed, today's AI-agent records are all Imported or Adapted, never Invented. “Exposed” means a role re-forms around its unverifiable core, not that it disappears.

The signal

Stop asking what AI can do; ask how much steering it takes. As models improve, tasks slide left — drag, to guided, to automatic — and the scarce human work climbs from typing to framing, integration, and ownership.

The first answer is not the ceiling · %

GPT-5.4 (task pass)Kimi-K2.6 (task pass)

Asuka-Bench. Task pass climbs steeply with feedback, but round-3 project completion (whole app works) stays near 50% — the gap between 'many parts work' and 'the whole thing works.'

Human, AI, or the pair? · %

Overall Pass@1Partial Pass@1

CentaurEval, collaboration-necessary tasks. The pair beats either side alone — but read the autonomous bars (0.67% / 2.89%) as the honest standalone number.

The steering curve — watch it move left · %

TodayImproved

Conceptual; curve shape illustrative. The envelope expands when the curve shifts left — same quality, fewer interventions — not just up.

Where the agent breaks as work scales · capability tier

L1 · Fix a functionStrong

L2 · Add a login formStrong with a spec

L3 · Auth across the appMixed

L4 · Build a dashboardUseful, needs steering

L5 · Own a system for monthsWeak / experimental

L6 · “Make GTA7”Outside the envelope

Six rungs from atomic task to creative-industrial project. Colour tracks meaning, from commons-green capability to the ember of work beyond reach.

ACT 06

The new engine's own walls

The old engine's two walls are behind us. The honest forward question is whether the new fuels have ceilings of their own. They do — emerging, not yet decisive.

With the old engine's walls in the rear-view, the real question is whether the new engine has walls of its own. On their own terms, there are three, and they rhyme.

Wall one is inference-time return. Letting a model think longer buys accuracy that climbs log-linearly, then flattens — and past the peak, overthinking makes answers worse. The cost runs the other way: attention scales quadratically, and a reasoning model can hold a GPU roughly 30× longer than a standard one. Unlike training, which you pay once, this is a toll collected at every query, forever.

Wall two is reinforcement learning. The 2025-26 RL scale-up delivered real gains, but it reads as a one-time lift, not a self-sustaining trend. One camp finds RLVR mostly reorders which answers surface first; another shows prolonged RL can uncover genuinely new strategies. The honest read: RL clearly refines, sometimes expands — and how often it expands decides how long this lasts.

Wall three is synthetic data, and whether it dooms the model depends entirely on how you use it. It is not a verdict but a ladder, and each rung is a real result.

The signal

Each new fuel has a ceiling, and none is proven fatal yet — that's what separates these from the old walls. Whether they arrive in one year or five is the biggest open question on the page.

Wall 1: inference returns saturate, then bend down · % accuracy

Accuracy rises log-linearly from 62% to an 88% peak at 64×, then slips to 87% at 128× as the model overthinks. Cost rises superlinearly the whole way.

Wall 3: does synthetic data doom the model?

Replace — collapses

MethodTrain only on the model's own fresh output each generation.

ResultQuality and diversity decay; useless within ~5 cycles.

SourceShumailov 2024; Alemohammad (MAD)

Accumulate — stops rot

MethodKeep all real data, pile synthetic on top; the anchor never shrinks.

ResultError stays bounded — but 'doesn't rot' isn't 'improves.' Stable mediocrity.

SourceGerstgrasser, COLM 2024; Kazdan 2024

Accumulate + verify — improves

MethodFilter synthetic through a verifier; keep only the good samples.

ResultCan reverse collapse into improvement. Ceiling set by verifier quality.

SourceYi 2026; Self-Verification, ICLR 2026

Three rungs, each a sourced result. Replacing real data collapses the model; accumulating on a real anchor stops the rot; accumulating AND verifying can reverse collapse into improvement — which is exactly what frontier labs do.

ACT 07

The expanding envelope, and the instruments that watch it flatten

How long a task an agent can own alone, how fast that edge moves — and the one ruler that can't hide a plateau.

An AI agent acts on its own only inside a bounded envelope: the goal is legible, the task decomposes, feedback is dense, failures are recoverable, and success can be checked. Outside it, a human still has to frame, steer, verify, rescue, and stitch the result back in. The envelope is not fixed. Two things move together — the task length an agent can own end-to-end grows, and the steering each task demands lightens — and plotted as a frontier, the boundary climbs fast.

The anchors are measured. METR's time-horizon benchmark tracks the task length a model finishes at 50% reliability: five minutes for GPT-4 in early 2023, half an hour by mid-2024, two hours by early 2025, twelve hours for Opus 4.6 by early 2026 — the floor rising about an order of magnitude a year. Run the line back further and it starts near two seconds with GPT-2 in 2019.

Why trust a horizon when every benchmark dies the same way? A ruler launches hard, the frontier overruns it, scores cluster near the ceiling, and it stops telling models apart. MMLU defined three years of progress before flatlining in the low 90s. GPQA Diamond — PhD-level, Google-proof science — went from 34% to 95%, now 25 points above the human experts it was built around. Each new ruler measures for less time before capability runs past it: MMLU lasted about four years, GPQA roughly 2.3. That looks like progress speeding up, not slowing.

A task horizon has no 100% ceiling, so a saturating benchmark can hide progress but a climbing horizon can't. That makes its rate the early-warning instrument. Doubling time has shrunk from 212 days to 89 — and the rising line itself tells you nothing about a coming flatten; only the doubling time does. The first sign of a slowdown is the day that number starts climbing again.

Two honest discounts ride on every number here. The horizon mostly measures coding and research — the checkable tasks — and the headline figure is cheaper than the deployed one; at 80% reliability the twelve-hour horizon collapses to about 25 minutes. The explosive case stays an open fork: superforecasters put the odds of AI R&D compressing six years into two at 8%, the domain experts closest to the work at 20%. The gate is long-horizon research taste, not coding speed.

The signal

The signal is the edge moving up and to the left: longer tasks owned with less steering. Watch the rate, not the level — doubling time has fallen from 212 days to 89, and the first sign of a slowdown is the day that number starts climbing again. It hasn't.

A ruler with no ceiling: task time horizon · minutes (log)

METR's time horizon on a log scale — from ~2 seconds (GPT-2) to ~12 hours (Opus 4.6), doubling every ~7 months and lately ~4. The line that can't hide a plateau.

The early-warning instrument: doubling time · days

Days for the horizon to double, by progressively more recent slice of models. Shorter = faster. 212d down to 89d. A future slice turning taller would be the first sign of a slowdown.

The rulers are breaking · %

LaunchNow

Launch score vs best model now. Each benchmark gets solved and goes flat; the frontier has retreated to HLE and FrontierMath, where models still fail.

The honest asterisk

37%

benchmark → deployment gap

12h

horizon @ 50% reliability

25min

… @ 80% reliability

Discounts to apply to every rising line. The 80%-reliability collapse is the alarm to read against the looser 50% horizons.

Odds of 6-years-in-2 compression

Superforecasters

20%

AI domain experts

The experts closest to the work price the explosive case higher — because the gate is long-horizon research taste, not coding speed.

ACT 08

No single number — read the convergence

The decisive signal is several instruments turning at once, not any one benchmark.

The frontier is not one curve and not one engine, and so it cannot be read off one chart. Each instrument catches a different edge of the same expanding envelope: how long a task agents can own, how much of their gain is verified-novel rather than recombined, how much human steering a result still costs, and how much of the valuable work has a cheap automatic check at all. Any one of these can drift for reasons that don't generalize. Read alone, each misleads.

The signal that matters is when they move together. Longer task horizons, a rising invention share, and a falling steering burden are independent measurements; if they turn at the same time, the envelope itself is widening, not one benchmark saturating. So the right thing to watch is not a record but a correlation. The convergence is the news.

The signal

Watch for longer-horizon autonomy, lower steering burden, and verified novelty rising together — that convergence, not any single chart, is the tell.

Eight instruments to watch

Task horizons

What it measuresHow long a task agents finish at 50% reliability — no ceiling yet

Invention share

What it measuresVerified-novel work as a fraction of total gains

Synthetic-data yield

What it measuresGain per synthetic token, gated by the verifier

Benchmark health

What it measuresHow fast rulers saturate; whether unsaturated ones still exist

Cost per correct answer

What it measuresDeployed cost to get a right result, not headline price

Steering burden

What it measuresHuman minutes, rounds, interventions to target quality (defined in Act 5)

Project completion gap

What it measuresDistance between task-pass and whole-system success

Verifier coverage

What it measuresShare of valuable tasks with a cheap automatic check

Each row is an instrument for reading the verifier-cost-steering frontier.