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 Wed, 02 Sep 2026 23:03:31 +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. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/it-never-becomes-free/ https://awarenesssoftwaregroup.com/research/it-never-becomes-free/ Our one positive efficiency result does not survive a larger model with a matched task. The point past which switching off was free turns out to have been the point past which the model had stopped learning. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/one-number-for-every-size/ https://awarenesssoftwaregroup.com/research/one-number-for-every-size/ We published one figure for what our early-warning measurement costs, then quoted it in eight results. It was computed in a unit that cancels out of its own ratio. Wed, 02 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. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-cliff-survives-matching/ https://awarenesssoftwaregroup.com/research/the-cliff-survives-matching/ We predicted our shortcut's failure at scale would dissolve once the task was made harder alongside the model. It did not. The first effect here to survive that check rather than dissolve on it. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-effect-never-faded/ https://awarenesssoftwaregroup.com/research/the-effect-never-faded/ We reported an effect vanishing as models grow. Making the task harder alongside the model shows it does not vanish at all. The disappearance was produced by how we ran the test. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-moment-marks-nothing/ https://awarenesssoftwaregroup.com/research/the-moment-marks-nothing/ We withdrew yesterday's conclusion the same day. Model size decides whether the shortcut helps; the moment we thought was the dividing line turns out to mark nothing at all. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-rule-caught-it-the-same-day/ https://awarenesssoftwaregroup.com/research/the-rule-caught-it-the-same-day/ We found a genuine signal in a place our earlier negative result never looked, then a rule written hours earlier made us check it against model size. It disappears. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/the-shortcut-reversed/ https://awarenesssoftwaregroup.com/research/the-shortcut-reversed/ A training shortcut that saved 22% of the data at one model size costs 26% more at a larger one. The early warning that told us when to take it is still worth nothing. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/we-tried-to-spend-our-own-result/ https://awarenesssoftwaregroup.com/research/we-tried-to-spend-our-own-result/ The first time this project tried to use one of its findings to make training cheaper. The detector cost thirteen times what it saved. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/where-the-shortcut-stops-paying/ https://awarenesssoftwaregroup.com/research/where-the-shortcut-stops-paying/ The one shortcut we have that saves anything works below about a hundred units of width and not above it. The drop is a cliff, and it lands exactly where extra model size stops paying. Wed, 02 Sep 2026 00:00:00 +0000 https://awarenesssoftwaregroup.com/research/which-side-of-the-moment/ https://awarenesssoftwaregroup.com/research/which-side-of-the-moment/ A grid separating model size, task difficulty and data-per-step. Its negative stands; its headline was withdrawn the same day by a follow-up that untied the confound this record named itself. Wed, 02 Sep 2026 00:00:00 +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