The signal
Smaller models that rival much larger ones are appearing rapidly, often trained partly on the outputs of leading systems. AI companies are tightening terms of use and technical protections to stop their models being copied.
Where it could lead
If any breakthrough can be copied cheaply within months, the business model of spending billions on frontier AI gets harder to justify. For users it means powerful AI keeps getting cheaper, faster.
Evidence that would strengthen the case
- Legal cases or rules that define whether copying a model’s outputs is allowed
- AI providers shift to protecting data, distribution and services rather than the model itself
- Distilled models become the default choice for most business applications
What could prevent it
- Technical watermarking or access controls make distillation impractical
- Courts or regulators treat it as clear theft, chilling the practice
- The gap between frontier and distilled models widens again