Noam Shazeer puts it at 50/50 that the next major breakthrough comes from Gemini rather than a human researcher. Four days earlier the corpus recorded a model watching its own training runs; this is the same claim stated as a probability.
The counterweight is friction. The Gemini 3 team now rejects performance gains that exceed a complexity budget, protecting the throughput of 150 researchers over isolated metrics — capability traded away deliberately.