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 Sat, 26 Sep 2026 16:48:51 +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-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/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 https://awarenesssoftwaregroup.com/research/depth-costs-most-of-the-budget/ https://awarenesssoftwaregroup.com/research/depth-costs-most-of-the-budget/ A skill built on two layers of other skills took most of the training budget to learn, though alone it was the easiest of five. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/depth-sets-the-order/ https://awarenesssoftwaregroup.com/research/depth-sets-the-order/ Given enough time, a small model learned five planted skills in the order of what depends on what on every run, even though the dependent skills were the Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/does-the-early-signal-generalise/ https://awarenesssoftwaregroup.com/research/does-the-early-signal-generalise/ Two findings that looked independent turn out to have identical boundaries. The early-warning gap appears exactly where the internal reorganisation does. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/does-the-effect-survive-the-optimizer/ https://awarenesssoftwaregroup.com/research/does-the-effect-survive-the-optimizer/ Every result we had used one training algorithm. Changing it at matched steps-to-solve keeps the effect at full size and takes away its timing. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/eight-runs-is-not-enough/ https://awarenesssoftwaregroup.com/research/eight-runs-is-not-enough/ A free experiment produced two statistically significant results pointing in opposite directions. Our first analysis reported one of them as a discovery. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/four-of-them/ https://awarenesssoftwaregroup.com/research/four-of-them/ We counted how many of our own findings that held up have been demonstrated on more than one kind of task. Sat, 26 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/how-fast-it-learned/ https://awarenesssoftwaregroup.com/research/how-fast-it-learned/ Choosing which existing model to build on: how fast each candidate learned its own task ranks the useful ones better than a quick check. Sat, 26 Sep 2026 00:00:00 +0000