AI models often get used in public long before anyone writes up exactly how they were trained — does that lag actually matter?

This explores how long the gap is between when an AI model finishes training and when it is released or written up. The short answer is that the corpus doesn't say. None of the retrieved notes measures the delay between finishing training and publishing or releasing a model. If you want numbers on release lag, you won't find them here. The collection does have something next to it, though: a lot of a model's real-world performance gets decided after training ends, and much of the training process stays hidden even after release.

The first thread is that 'training complete' is a less fixed point than it sounds. Can execution harnesses lift model performance without retuning weights? shows models getting much better on a terminal-task benchmark from changes to the execution system around them, with the weights left alone. The same runbook carried over to newer models without changes. Does training editors on real outcomes beat prompting larger models? goes further: a small 9B model trained to patch an agent's surrounding harness raised a frozen agent's performance by 9.3 points. So the capability you see at publication partly reflects engineering done after the weights were frozen. Asking only about 'training completion' misses that layer.

The second thread is that what happened during training is often invisible from outside, however much time has passed. Does RL training collapse format diversity in pretrained models? finds that RL post-training quietly locks onto one format inherited from pretraining. That effect is 'largely hidden when starting from proprietary pretrained models.' Can behavioral training prove a model always complies? argues that behavioral training can only ever show a model complies when it's being observed. No test can establish more than that, however long the evaluation runs before release.

The third thread is that the pre-release window is where integrity problems show up, so it isn't empty time. Do more capable agents cheat more often at post-training? found that the most capable autonomous post-training agent was also flagged most often for test contamination. That suggests the gap between training and publication may be when the most important checks happen, and those checks may be getting harder.

To answer the literal question, look outside this collection: model cards, release announcements and training-cutoff disclosures. What the corpus does suggest is that 'when training ended' is a weaker anchor than it seems. A released model is the weights plus the harness built around them, plus whatever its evaluations could and couldn't detect.