https://awarenesssoftwaregroup.com/research/ Open research on how AI models actually learn. One finding per entry, in plain English first and in full technical detail underneath, including the approaches that turned out not to work. en Sun, 27 Sep 2026 04:08:35 +0000 https://awarenesssoftwaregroup.com/research/flat-through-width-256/ https://awarenesssoftwaregroup.com/research/flat-through-width-256/ Across a fourfold range of model width, an unneeded skill costs the other skills the same share of their learning time. Sun, 27 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/less-competition-at-width-512/ https://awarenesssoftwaregroup.com/research/less-competition-at-width-512/ With its learning rate tuned, the widest model paid a smaller share for an unneeded skill, but its rate changed along with its width, so which one Sun, 27 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/most-of-it-is-pull/ https://awarenesssoftwaregroup.com/research/most-of-it-is-pull/ Teaching an unneeded skill half as often removed about half of what it cost a small model's other skills: the competition is mostly the pull of training, Sun, 27 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/questions/ https://awarenesssoftwaregroup.com/research/questions/ The 10 questions our public AI training research is organised around, each with its current answer and the published results behind it. Sun, 27 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/two-papers-on-interference/ https://awarenesssoftwaregroup.com/research/two-papers-on-interference/ Two 2026 papers on how tasks interfere during training, read against our own measurements: one matches what we found, the other predicts something our Sun, 27 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-boost-off-the-frontier/ https://awarenesssoftwaregroup.com/research/a-boost-off-the-frontier/ Raising the learning rate by half for 40 early steps made small models learn as fast as a higher rate while ending as accurate as the lower one. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-five-second-check/ https://awarenesssoftwaregroup.com/research/a-five-second-check/ Reusing a finished model to speed up a new one is worse than useless if you pick without looking. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-fixed-budget-thinks-better/ https://awarenesssoftwaregroup.com/research/a-fixed-budget-thinks-better/ Penalising a model for thinking cut its thinking by 38% at almost no cost. Always letting it think for three steps beat that, and we found out why. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-free-check-saturates-at-three/ https://awarenesssoftwaregroup.com/research/a-free-check-saturates-at-three/ A one-pass check that tells a helpful trained model from a harmful one gets all of its value from the first three candidates. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-head-start-from-another-model/ https://awarenesssoftwaregroup.com/research/a-head-start-from-another-model/ Dropping a finished model's internal machinery into a fresh one removes about forty per cent of the time it takes to learn. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-quarter-is-free/ https://awarenesssoftwaregroup.com/research/a-quarter-is-free/ Copying the wrong finished model into a new one is the worst thing we have measured. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-queue-not-a-graph/ https://awarenesssoftwaregroup.com/research/a-queue-not-a-graph/ Holding back one skill while a small model trained delayed the skills it normally learns afterwards, including ones that did not depend on it at all. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-ratio-that-could-not-decide/ https://awarenesssoftwaregroup.com/research/a-ratio-that-could-not-decide/ Our test of why skills compete fired its kill test, but on a measurement so unstable it ruled nothing out: one model's baseline ran backwards and the Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-rule-from-1990/ https://awarenesssoftwaregroup.com/research/a-rule-from-1990/ We finally asked what a pre-neural predictor scores on the task most of our results were measured on. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-short-model-is-a-lottery/ https://awarenesssoftwaregroup.com/research/a-short-model-is-a-lottery/ Two of our own experiments reached opposite conclusions from one trained model each. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-shortcut-only-matters-if-there-is-no-other-one/ https://awarenesssoftwaregroup.com/research/a-shortcut-only-matters-if-there-is-no-other-one/ Three of our results seemed to disagree about one piece of wiring. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-small-dose-does-little/ https://awarenesssoftwaregroup.com/research/a-small-dose-does-little/ Mixing only part of a borrowed model into a new one does little, good or bad. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-speed-accuracy-frontier/ https://awarenesssoftwaregroup.com/research/a-speed-accuracy-frontier/ Faster learning rates made small models learn sooner but end less accurate. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-task-battery-with-designed-headroom/ https://awarenesssoftwaregroup.com/research/a-task-battery-with-designed-headroom/ Our quality checks all passed because the benchmark scored 99% and nothing could fail. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-warning-before-anything-happens/ https://awarenesssoftwaregroup.com/research/a-warning-before-anything-happens/ Our early-warning probe's famous head start holds for one task only, and on easy tasks it goes off at the very first step, before the model has learned Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-window-that-parks/ https://awarenesssoftwaregroup.com/research/a-window-that-parks/ We wrote down the practice plan an adaptive teacher discovered, as a two-setting rule. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/a-yardstick-that-does-not-stretch/ https://awarenesssoftwaregroup.com/research/a-yardstick-that-does-not-stretch/ We fixed the measuring instrument that had given us a false result the day before, using no new training at all. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/being-choosy-about-data-costs-more-than-it-saves/ https://awarenesssoftwaregroup.com/research/being-choosy-about-data-costs-more-than-it-saves/ Inspect four batches, train only on the most informative. It sounds efficient. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/boost-then-decay/ https://awarenesssoftwaregroup.com/research/boost-then-decay/ A learning-rate schedule that speeds up early and slows down late made small models learn as fast as our boost and end as accurate as the standard Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/calibrate-after-learning/ https://awarenesssoftwaregroup.com/research/calibrate-after-learning/ Training a thinking model's early answers blocked learning when done from the start. Done after it learned, it helped every run. Sat, 26 Sep 2026 00:00:00 +0000