Predicting the Rational — AI experiments and learnings.https://labs.predictingtherational.com/https://labs.predictingtherational.com/experiments/009-more-context-same-answer/https://labs.predictingtherational.com/experiments/009-more-context-same-answer/Given the right passage, the cheapest models read a regulation correctly 99 times out of 100. The system built around them manages 77. Retrieving more context does not close that gap; it swaps one kind of wrong answer for another. And when the right passage is missing altogether, the models hallucinate a number instead.Mon, 03 Aug 2026 00:00:00 GMThttps://labs.predictingtherational.com/experiments/008-manufacturing-a-compiler/https://labs.predictingtherational.com/experiments/008-manufacturing-a-compiler/Software engineering is the one domain where the error signal came free with the toolchain, and the compiler's real gift was never accuracy but five properties held at once- free, instant, total, independent, deterministic. Anywhere else you have to manufacture a substitute, and no substitute gives you all five. That turns evaluation from a gate you install into a budget you allocate, and the binding constraint on that budget turns out to be organisational rather than technical.Sun, 26 Jul 2026 00:00:00 GMThttps://labs.predictingtherational.com/experiments/006-token-subsidy-is-ending/https://labs.predictingtherational.com/experiments/006-token-subsidy-is-ending/Unit prices for intelligence are collapsing while enterprise AI bills explode. Both are true because agents changed the metric from cost-per-prompt to cost-per-task — and the mathematics of stateless LLMs, context bloat, and GPU memory eviction mean the tokens per task are growing faster than the price per token is falling.Sat, 18 Jul 2026 00:00:00 GMThttps://labs.predictingtherational.com/experiments/007-cache-eviction/https://labs.predictingtherational.com/experiments/007-cache-eviction/Part 3 of this series claimed provider caches get evicted under memory pressure long before their advertised lifetime, so idle agents resume at full price. I measured it across 120 trials. GPT-5.6 honors its documented 30-minute floor on every single trial and drops the entry the moment the obligation ends — a step function, not a decay curve. The older best-effort cache survived roughly half an hour too, against documentation promising five to ten minutes. The eviction claim was wrong, and the guarantee turns out to buy not more time but the elimination of variance.Sat, 18 Jul 2026 00:00:00 GMThttps://labs.predictingtherational.com/experiments/005-coding-was-easy-mode/https://labs.predictingtherational.com/experiments/005-coding-was-easy-mode/The strongest evidence for 'AI will automate everything' is what it did to software engineering. But SE is a structural anomaly: the one domain with peer-reviewed training data at scale, telemetry on every keystroke, and a deterministic compiler to catch hallucinations. That combination doesn't exist anywhere else.Mon, 13 Jul 2026 00:00:00 GMThttps://labs.predictingtherational.com/experiments/004-innovation-plateau/https://labs.predictingtherational.com/experiments/004-innovation-plateau/AI's exponential run is colliding with its own economics. Two old frameworks — Foster's S-Curve and Perez's Technological Surge Cycle — predicted the shape of what comes next: not collapse, but a Turning Point. This anchor post lays out the thesis; two companion posts examine the software engineering exception and the token economics driving it.Sat, 11 Jul 2026 00:00:00 GMT