Q1· lens: The stack

Are frontier model capabilities still improving fast, or has the curve bent?

◇ Neutral question

Nearly every financial and infrastructure bet is an implicit forecast of this capability curve.

◷ Share view, as of June 2026: confidence, rationales and priors are hidden. The questions, stances, claims, tests and evidence stay visible.

Stances · candidate answers

Q1-S1A

Scaling still works, the curve just moved.

Held by lab research leadership; scaling-law researchers

Pretraining returns diminish, but inference-time compute, RL on verifiable rewards, and tool use opened new axes nowhere near saturated.

would move me offTwo consecutive generations from multiple labs show no improvement on held-out, contamination-controlled benchmarks despite order-of-magnitude more compute: field-wide flat output against rising input.

Strongest opposing: Q1-S1B, Q1-S1C

Evidence behind its claims: 170 support · 20 contradict · 100 neutral

Q1-S1B

The economically useful curve has bent, even if benchmarks haven't.

Held by careful capability skeptics; a growing camp of enterprise buyers

Scores rise but marginal real-world reliability (the thing that lets you remove a human) has flattened, and the last 5% is where the money is.

would move me offA documented wave of agentic deployments running unsupervised at low error rates in high-stakes domains (sustained production, not demos), or a trusted reliability-tracking benchmark that climbs.

Strongest opposing: Q1-S1A, Q1-S1C

Evidence behind its claims: 170 support · 20 contradict · 100 neutral

Q1-S1C

We can't tell, because evaluation is broken.

Held by eval researchers / methodologists

Contamination, teaching-to-the-test, and saturation mean the slope is genuinely unknown. Epistemically humble, under-discussed, real.

would move me offMaturation of held-out, continuously-refreshed, contamination-resistant evaluation the field converges on and trusts.

Strongest opposing: Q1-S1A, Q1-S1B

Evidence behind its claims: 20 support · 0 contradict · 22 neutral

The map: how claims bear on stances

supports contradictshover a node to highlight its links

Claims

1.1

Frontier benchmark scores kept rising gen-over-gen through 2025–26.

Q1-S1AQ1-S1B

1.2

Benchmark gains are increasingly contaminated / saturated, overstating progress.

Q1-S1BQ1-S1CQ1-S1A

1.3

Agentic reliability (multi-step completion without human correction) is improving materially, not just peak single-shot.

Q1-S1AQ1-S1B

1.4

Inference-time compute yields gains that don't saturate quickly on hard reasoning.

Q1-S1AQ1-S1B

Stances

Q1-S1A

Scaling still works, the curve just moved.

held by lab research leadership; scaling-law researchers

Q1-S1B

The economically useful curve has bent, even if benchmarks haven't.

held by careful capability skeptics; a growing camp of enterprise buyers

Q1-S1C

We can't tell, because evaluation is broken.

held by eval researchers / methodologists

Claims, each with its falsifiable test

1.1 · Frontier benchmark scores kept rising gen-over-gen through 2025–26.

would falsifyTrajectories flatten despite continued compute scaling.

CapabilityClean to resolve

1.2 · Benchmark gains are increasingly contaminated / saturated, overstating progress.

would falsifyContamination-controlled re-tests reproduce the gains.

CapabilitySlow to resolve

1.3 · Agentic reliability (multi-step completion without human correction) is improving materially, not just peak single-shot.

would falsifyProduction agent error / intervention rates fail to fall across 2025–26.

CapabilitySlow to resolve

1.4 · Inference-time compute yields gains that don't saturate quickly on hard reasoning.

would falsifyReasoning-model gains plateau sharply with more inference compute on novel problems.

CapabilitySlow to resolve

Bridge-claims connecting out

⤧ B1Agentic reliability gains will translate into enterprise ROI.Capability → Economics ⤧ B3Training-efficiency gains will reduce aggregate compute demand (threatening "more compute = more value").Capability → Build-out

Q1 / 4 claims / 3 stances / 2 bridges·confidence moves are made manually; AI proposes only