AI Worth Knowing https://aiworthknowing.fun/ Artificial intelligence explained without the sales pitch — how the systems work, where they fail, what the words mean, and how the field arrived here. en Learning, in a neural network, means adjusting numbers until the errors shrinkhttps://aiworthknowing.fun/howitworks/learning-in-a-neural-network-means-adjusting-numbers-until-the-errors-shrink/https://aiworthknowing.fun/howitworks/learning-in-a-neural-network-means-adjusting-numbers-until-the-errors-shrink/Wed, 26 Aug 2026 09:00:00 GMTHow It WorksManish TrivediThe word was borrowed from human experience and it flatters the mechanism; what actually happens is a very large curve being fitted through a very large number of points. Training a model and running one are two different kinds of workhttps://aiworthknowing.fun/howitworks/training-a-model-and-running-one-are-two-different-kinds-of-work/https://aiworthknowing.fun/howitworks/training-a-model-and-running-one-are-two-different-kinds-of-work/Mon, 24 Aug 2026 09:00:00 GMTHow It WorksZoya RahmanOne is a construction project with a fixed budget and an end date; the other is a utility bill that arrives every time somebody presses a key. A token is not a word, and several odd behaviours follow from thathttps://aiworthknowing.fun/howitworks/a-token-is-not-a-word-and-several-odd-behaviours-follow-from-that/https://aiworthknowing.fun/howitworks/a-token-is-not-a-word-and-several-odd-behaviours-follow-from-that/Thu, 20 Aug 2026 09:00:00 GMTHow It WorksManish TrivediLanguage models do not read letters or words; they read fragments from a fixed vocabulary, and the seams between those fragments explain failures that otherwise look inexplicable. Predicting the next fragment is a narrower job than the results suggesthttps://aiworthknowing.fun/howitworks/predicting-the-next-fragment-is-a-narrower-job-than-the-results-suggest/https://aiworthknowing.fun/howitworks/predicting-the-next-fragment-is-a-narrower-job-than-the-results-suggest/Tue, 18 Aug 2026 09:00:00 GMTHow It WorksDaniel OkonkwoThe training objective is almost embarrassingly simple, and the interesting argument in the field is about how much that simplicity can be made to carry. Attention is a weighted lookup and the name oversells it considerablyhttps://aiworthknowing.fun/howitworks/attention-is-a-weighted-lookup-and-the-name-oversells-it-considerably/https://aiworthknowing.fun/howitworks/attention-is-a-weighted-lookup-and-the-name-oversells-it-considerably/Mon, 17 Aug 2026 09:00:00 GMTHow It WorksManish TrivediThe mechanism behind modern language models lets every part of a sequence consult every other part, and its real advantage was never the biological metaphor. Ranking systems were the machine learning most people met firsthttps://aiworthknowing.fun/world/ranking-systems-were-the-machine-learning-most-people-met-first/https://aiworthknowing.fun/world/ranking-systems-were-the-machine-learning-most-people-met-first/Sat, 15 Aug 2026 09:00:00 GMTIn The WorldZoya RahmanLong before anyone typed a question into a chat box, trained models were deciding the order of feeds, search results and shop shelves, and that deployment has been running long enough to have visible consequences. Machine translation cleared a threshold and left its hardest problems behindhttps://aiworthknowing.fun/world/machine-translation-cleared-a-threshold-and-left-its-hardest-problems-behind/https://aiworthknowing.fun/world/machine-translation-cleared-a-threshold-and-left-its-hardest-problems-behind/Wed, 12 Aug 2026 09:00:00 GMTIn The WorldNaina SethiThe jump in quality was real and it changed how much of the world people can read, but the failures that remain are more dangerous than the clumsy ones it replaced. A model trained on past decisions will reproduce the reasoning behind themhttps://aiworthknowing.fun/world/a-model-trained-on-past-decisions-will-reproduce-the-reasoning-behind-them/https://aiworthknowing.fun/world/a-model-trained-on-past-decisions-will-reproduce-the-reasoning-behind-them/Tue, 11 Aug 2026 09:00:00 GMTIn The WorldZoya RahmanAutomated screening in hiring, credit and public services learns from a record of what was decided before, and that record contains everything about how those decisions were actually made. Recognition systems perform unevenly and the unevenness is not randomhttps://aiworthknowing.fun/world/recognition-systems-perform-unevenly-and-the-unevenness-is-not-random/https://aiworthknowing.fun/world/recognition-systems-perform-unevenly-and-the-unevenness-is-not-random/Sat, 08 Aug 2026 09:00:00 GMTIn The WorldImran SheikhSpeech and image systems are described by a single accuracy figure, and that figure conceals which people the system works for and which people it does not. When a first draft costs almost nothing the bottleneck moves to checking ithttps://aiworthknowing.fun/world/when-a-first-draft-costs-almost-nothing-the-bottleneck-moves-to-checking-it/https://aiworthknowing.fun/world/when-a-first-draft-costs-almost-nothing-the-bottleneck-moves-to-checking-it/Fri, 07 Aug 2026 09:00:00 GMTIn The WorldManish TrivediGenerating plausible text became cheap and verifying it did not, and a great many institutions were quietly built on the assumption that those two costs were similar. Being confidently wrong is the default behaviour, not a malfunctionhttps://aiworthknowing.fun/limits/being-confidently-wrong-is-the-default-behaviour-not-a-malfunction/https://aiworthknowing.fun/limits/being-confidently-wrong-is-the-default-behaviour-not-a-malfunction/Tue, 04 Aug 2026 09:00:00 GMTLimits & RisksManish TrivediNothing in the way these systems are built distinguishes a supported statement from an unsupported one, so the surprising thing is not the false answers but how many true ones there are. What a benchmark score measures and what it quietly does nothttps://aiworthknowing.fun/limits/what-a-benchmark-score-measures-and-what-it-quietly-does-not/https://aiworthknowing.fun/limits/what-a-benchmark-score-measures-and-what-it-quietly-does-not/Sun, 02 Aug 2026 09:00:00 GMTLimits & RisksSamar BhatiaEvery claim about a system being better than another rests on a test, and the properties of those tests are less well understood than the numbers they produce. Correlation carries these systems a long way and then stops at the edgeshttps://aiworthknowing.fun/limits/correlation-carries-these-systems-a-long-way-and-then-stops-at-the-edges/https://aiworthknowing.fun/limits/correlation-carries-these-systems-a-long-way-and-then-stops-at-the-edges/Sat, 01 Aug 2026 09:00:00 GMTLimits & RisksManish TrivediStatistical association is enough to produce behaviour that looks like understanding across most of the range, and the difference only becomes visible where the training data ran out. A system’s account of its own reasoning is a reconstruction, not a recordhttps://aiworthknowing.fun/limits/a-systems-account-of-its-own-reasoning-is-a-reconstruction-not-a-record/https://aiworthknowing.fun/limits/a-systems-account-of-its-own-reasoning-is-a-reconstruction-not-a-record/Tue, 28 Jul 2026 09:00:00 GMTLimits & RisksManish TrivediThese models can be asked why they answered as they did and will produce a fluent explanation, which is generated by the same process that generated the answer and carries no special authority. Training data has an end date and the model cannot feel the edge of ithttps://aiworthknowing.fun/limits/training-data-has-an-end-date-and-the-model-cannot-feel-the-edge-of-it/https://aiworthknowing.fun/limits/training-data-has-an-end-date-and-the-model-cannot-feel-the-edge-of-it/Fri, 24 Jul 2026 09:00:00 GMTLimits & RisksManish TrivediEverything a system knows was fixed at some point in the past, the boundary is blurrier than a single date suggests, and the system has no sense of standing at it. Symbolic AI was not a mistake, it was answering a different questionhttps://aiworthknowing.fun/history/symbolic-ai-was-not-a-mistake-it-was-answering-a-different-question/https://aiworthknowing.fun/history/symbolic-ai-was-not-a-mistake-it-was-answering-a-different-question/Wed, 22 Jul 2026 09:00:00 GMTHistoryManish TrivediThe approach that dominated the field for its first three decades is now treated as a dead end, which misreads both what it achieved and why it stalled. The perceptron controversy that shelved neural networks for a decadehttps://aiworthknowing.fun/history/the-perceptron-controversy-that-shelved-neural-networks-for-a-decade/https://aiworthknowing.fun/history/the-perceptron-controversy-that-shelved-neural-networks-for-a-decade/Wed, 15 Jul 2026 09:00:00 GMTHistoryNaina SethiA single mathematical limitation in an early learning machine became the reason an entire research programme lost its funding, and the story is told more confidently than the evidence supports. An AI winter is a funding event before it is a scientific onehttps://aiworthknowing.fun/history/an-ai-winter-is-a-funding-event-before-it-is-a-scientific-one/https://aiworthknowing.fun/history/an-ai-winter-is-a-funding-event-before-it-is-a-scientific-one/Wed, 08 Jul 2026 09:00:00 GMTHistoryImran SheikhThe field has twice gone from national enthusiasm to institutional embarrassment, and the mechanism was the same both times and had little to do with whether the research was any good. Expert systems made the field money and then became its cautionary talehttps://aiworthknowing.fun/history/expert-systems-made-the-field-money-and-then-became-its-cautionary-tale/https://aiworthknowing.fun/history/expert-systems-made-the-field-money-and-then-became-its-cautionary-tale/Sun, 05 Jul 2026 09:00:00 GMTHistorySamar BhatiaFor a decade the way to build a useful thinking machine was to interview a specialist and write down what they knew, and the reasons that stopped working are more interesting than the fact that it did. The statistical approach won by outgrowing the argument rather than settling ithttps://aiworthknowing.fun/history/the-statistical-approach-won-by-outgrowing-the-argument-rather-than-settling-it/https://aiworthknowing.fun/history/the-statistical-approach-won-by-outgrowing-the-argument-rather-than-settling-it/Sat, 04 Jul 2026 09:00:00 GMTHistoryDaniel OkonkwoTwo research traditions spent decades disagreeing about whether counting could substitute for understanding, and the disagreement was resolved by hardware and data rather than by anyone conceding.