Hiring a first-rank researcher away from a competitor, alongside further departures from the same team in the following weeks, exposed a rarely discussed feature of the sector: the number of people able to run frontier-scale training is very small.
The entry barrier in artificial intelligence is usually described in capital and compute, which are real and measurable obstacles. The third barrier rarely enters the calculation and is the hardest to overcome with fast money: accumulated experience from people who have trained large models and know what to do when training goes wrong.
That knowledge isn't published. Papers describe architecture and results; what decides in practice is judgement about data, about when to stop and what to adjust, and that's learned by doing.
A trade secrets lawsuit weeks later turned on the identical question, which is why the line between personal knowledge and company asset is so hard to draw.
For the market, the effect is concentration: whoever already has these people holds an advantage that a bigger budget doesn't buy.