The three companies spending the most money to build the most powerful AI models just announced they're teaming up — not to build AGI together, but to catch people who try to clone it.
What is model distillation, really?
Think of it like this: you spend years developing a world-class chef, pouring resources into training their palate and technique. Someone else then watches your chef cook, takes notes, and opens a competing restaurant using what they learned. That's distillation in a nutshell — using another model's outputs to train a cheaper knockoff.
The practice sits in a gray area. It's technically against most AI companies' terms of service, but it's hard to prove and even harder to stop. When your model outputs are public anyway, the argument goes, isn't it fair game to learn from them? The Frontier Forum members don't think so.
They're not worried about fair competition. They're worried about losing control of their moats.
The timing matters. As inference costs drop and open-source models improve, the incentive to distill frontier models grows. Why pay premium prices for GPT-5 when a distilled version performs at 90% for 10% of the cost? That's the existential question these companies are trying to answer — collectively.
This is weird. Why now?
A few factors converged:
- Open-source pressure: Meta's Llama and other models have closed the gap. Distillation is becoming the shortcut around expensive training runs.
- Regulatory heat: The UK is actively courting Anthropic. Sharing intel on bad actors looks proactive to regulators.
- The China factor: If you're worried about your model weights ending up in Beijing, you need ways to detect when they're being extracted.
The interesting tension here: these companies can barely stand each other in public. Yet they've found common ground in protecting their most valuable asset — the training data and model weights that took billions to build.
What this means for the rest of us
If you're building AI products, this matters in two ways. First, the era of "just use the API and fine-tune" is getting stricter. The terms of service actually mean something now. Second, expect more walled gardens. The frontier models will increasingly lock down their outputs while the open-source world keeps building in parallel.
The collaboration itself is notable — they're treating model protection like a shared defense pact. Whether that's enough to stop determined distillators is another question. But for now, the AI giants are speaking the same language.
The real question isn't whether distillation happens. It does. The question is whether the frontier players can make it expensive and risky enough that most people decide it's not worth it.