Reed’s view
Amazon’s decision to scrap its AI leaderboard after tokenmaxxing rapidly drove up costs is the most recent case of sticker shock hitting companies trying to implement AI. Those rising costs have led to a growing chorus of people hawking Chinese open-sourced models like DeepSeek.
But in a recent interview, Microsoft AI chief Mustafa Suleyman told me why that’s a bad idea.
Companies that rely on distillation — the process of using datasets generated by larger AI models from frontier labs like Anthropic and OpenAI — is a shortcut that often leads to a dead end, Suleyman said. “You’ve basically stuffed your model full of somebody else’s knowledge,” he said.
As Microsoft builds toward independence with its own AI models, it is doing so with “zero distillation.”
Because big frontier labs don’t make public the massive data sets they use when they train the world’s biggest AI models, it’s impossible to know exactly what those companies prioritized when they created them. So while distillation can be effective at making small models for specific tasks, eventually models built on a foundation of distilled data will fall behind when applied to a wide range of general-purpose tasks.
Cheap, distilled Chinese AI models haven’t taken over, as many predicted. Demand for the most advanced AI models has been surging far more than demand for open-source versions. The limits of distillation could be one of the reasons.
And if Suleyman is right, there’s an even bigger gap between the big, frontier AI labs and the open-source models than people think. It’s prohibitively expensive for most companies to train such massive AI models from scratch.
And that could have a big impact on the cost companies must pay to implement AI. Free, open-source AI models aren’t a real alternative to the cutting-edge ones on the frontier, at least not for the most useful work tasks.
We’re past the experimentation phase of AI and it is now becoming a crucial tool for many companies. Costs, even for frontier models, are coming down incredibly fast, but not as fast as companies and users would like. It could be that there aren’t any shortcuts for getting there.
Notable
- Microsoft restructured Copilot leadership last month to free up Suleyman to focus on building new models, with Suleyman writing internally that the next phase would be for him “to focus all my energy on our Superintelligence efforts.”