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 experiments that went against us. en Tue, 01 Sep 2026 21:44:04 +0000 https://awarenesssoftwaregroup.com/research/a-dial-not-a-switch/ https://awarenesssoftwaregroup.com/research/a-dial-not-a-switch/ We knew this internal change needs a model to be learning something. Now we know it scales with how much there is to learn, almost exactly. Tue, 01 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. Scrambling which unit is which changes nothing, so what transfers is not specific to the donor. Tue, 01 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. It scores perfectly, and the network never catches it. Tue, 01 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. Four candidate replacements, an acceptance rule fixed in advance, and two adoptions. Tue, 01 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. The repair also confirms the result stays withdrawn. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/an-average-is-not-a-guarantee/ https://awarenesssoftwaregroup.com/research/an-average-is-not-a-guarantee/ Two warning signals that fire before a model learns, and why neither can be relied on for any particular training run. Also a correction to our own previous result. Tue, 01 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. Even given perfect information for free, choosing well beat choosing arbitrarily by half a step out of ninety-four. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/close-calls-are-decided-by-the-code/ https://awarenesssoftwaregroup.com/research/close-calls-are-decided-by-the-code/ Two hundred comparisons, each recomputed fourteen ways. When two groups are clearly apart nothing changes; when they nearly touch, one in six changes its answer. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/consolidation-is-a-measurement-frame-effect/ https://awarenesssoftwaregroup.com/research/consolidation-is-a-measurement-frame-effect/ Freezing the reference frame reverses a result we had reported for a year. The internal spread we thought was settling after learning is in fact still growing. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/depth-does-not-add-stages/ https://awarenesssoftwaregroup.com/research/depth-does-not-add-stages/ In a two-layer model the layers learn at almost the same moment, so training them on separate schedules has no basis. But the reorganisation we study nearly doubles in the deeper layer. Tue, 01 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. Tue, 01 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. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/does-the-transition-create-features/ https://awarenesssoftwaregroup.com/research/does-the-transition-create-features/ Freezing a model's memory at its random starting values and training only the two ends. It learns, slowly and partially, and never abruptly. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/does-the-warning-cry-wolf/ https://awarenesssoftwaregroup.com/research/does-the-warning-cry-wolf/ Every reliability figure we have published was measured on runs that succeeded. This asks what our early warning does when nothing is coming. Tue, 01 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. Three times the runs later, we have the real answer. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/five-models-pay-for-one/ https://awarenesssoftwaregroup.com/research/five-models-pay-for-one/ Handing a trained model's machinery to fresh ones saves real time, and we finally worked out what it costs. A donor pays for itself once about five models have used it, and a cheaper donor is better than a thorough one. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/fourteen-copies-of-one-definition/ https://awarenesssoftwaregroup.com/research/fourteen-copies-of-one-definition/ Our archive holds fourteen separate implementations of the same measurement, copied across forty-one experiments. Run against each other on 1,091 committed runs, they agree - and the one that never drifts is the one anchored outside the run. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/how-many-of-our-results-are-close-calls/ https://awarenesssoftwaregroup.com/research/how-many-of-our-results-are-close-calls/ One of our results was overturned just by running it more times. So we measured how close every other published result sits to the line that decides it. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/how-you-cut-a-model-matters/ https://awarenesssoftwaregroup.com/research/how-you-cut-a-model-matters/ Delete three quarters of a model's internal connections two different ways, keeping exactly the same number, and one way costs twice as much as the other. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/is-learning-a-lucky-accident/ https://awarenesssoftwaregroup.com/research/is-learning-a-lucky-accident/ Testing whether sudden learning is randomness knocking a model out of a rut. Injected noise makes the plateau longer and the transition blurrier, so randomness is friction rather than fuel. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/is-transition-timing-controllable/ https://awarenesssoftwaregroup.com/research/is-transition-timing-controllable/ A nudge of one part in a hundred million delays learning by 21 steps. A nudge ten million times larger delays it by 24. There is no dial, and this closed a research line. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/it-does-not-survive-a-second-task/ https://awarenesssoftwaregroup.com/research/it-does-not-survive-a-second-task/ A window we had just shown survives bigger models turns out not to exist on a different exercise. Five of our results depend on it, including our only genuine saving. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/it-is-not-how-much-you-remove/ https://awarenesssoftwaregroup.com/research/it-is-not-how-much-you-remove/ Removing one direction from a model's training signal at the right moment causes most of the damage that stopping the whole component does. Removing three times as much at the wrong moment causes a third as much. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/it-tries-the-wrong-rule-first/ https://awarenesssoftwaregroup.com/research/it-tries-the-wrong-rule-first/ Instead of asking how often the model is right, we asked what it is doing. Early on it behaves like a simpler rule that does not work. Tue, 01 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/latent-knowledge-before-behaviour/ https://awarenesssoftwaregroup.com/research/latent-knowledge-before-behaviour/ The answer is linearly readable from a model's internal state about 23 steps before the model can produce it. An earlier version of the same experiment reported the opposite. Tue, 01 Sep 2026 00:00:00 +0000