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 Fri, 04 Sep 2026 21:33:29 +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. Fri, 04 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. Fri, 04 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. Fri, 04 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. Fri, 04 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. Fri, 04 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. Fri, 04 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. Fri, 04 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. Fri, 04 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. Fri, 04 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. Fri, 04 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. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/it-already-knew-the-order/ https://awarenesssoftwaregroup.com/research/it-already-knew-the-order/ We found the order in which a model learns three related skills, then tried teaching them in that order. Fri, 04 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. Fri, 04 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. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/neither-half-works-alone/ https://awarenesssoftwaregroup.com/research/neither-half-works-alone/ A trained model's machinery splits into directions and strengths. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/not-the-hardest/ https://awarenesssoftwaregroup.com/research/not-the-hardest/ A model that has finished a harder job ought to help with an easier one. It does not. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/one-forward-pass-tells-you/ https://awarenesssoftwaregroup.com/research/one-forward-pass-tells-you/ A single cheap check, fixed in advance, separates every borrowing that helped from every one that cost time, with a clear gap and nothing in between. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/predicting-it-without-watching/ https://awarenesssoftwaregroup.com/research/predicting-it-without-watching/ Our only working early-warning instrument costs 55% extra compute to buy about 23 steps of notice. Fri, 04 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. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/six-numbers/ https://awarenesssoftwaregroup.com/research/six-numbers/ Under two per cent of a trained model carries a third of the head start it gives away. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-experiment-we-could-not-run/ https://awarenesssoftwaregroup.com/research/the-experiment-we-could-not-run/ We tried to test whether our main finding is about how a model is built or about what it is trained to do. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-head-start-closes/ https://awarenesssoftwaregroup.com/research/the-head-start-closes/ We published a break-even calculation one day and tested it across model sizes the next. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-layer-that-changes-most-is-not-the-one-that-matters/ https://awarenesssoftwaregroup.com/research/the-layer-that-changes-most-is-not-the-one-that-matters/ In a two-layer model, one layer has a narrow window in which it must be learning and the other does not care when you interrupt it. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-lean-that-carries-nothing/ https://awarenesssoftwaregroup.com/research/the-lean-that-carries-nothing/ Two of our findings pointed at the same part of a model and suggested an obvious conclusion. We tested it properly and it is wrong. Fri, 04 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-method-that-looked-fastest-was-cheating/ https://awarenesssoftwaregroup.com/research/the-method-that-looked-fastest-was-cheating/ One way of choosing training data reached the finish line 26 steps sooner than the others. It got there by quietly skipping the hard half of the task. Fri, 04 Sep 2026 00:00:00 +0000