Everyone's Cooking, Nobody's Tasting the Product

Spent a chunk of the weekend reading a Reddit thread about whether a new Chinese model was “distilled” from a US one, and by the end of it I genuinely couldn’t tell if I’d learned anything about AI or just watched several hundred people argue about vibes wearing lab coats.

The setup is familiar by now. A new model drops, benchmarks look good, and within about four hours someone in an American AI lab (or someone rooting for one) says “that’s obviously distilled from our model, they couldn’t have done that on their own.” Then someone else points out the timeline doesn’t actually work, because you can’t distil a model from something that’s been public for a week. Then it turns into a proxy war about whose country has more PhDs. Somewhere in there, someone makes a Breaking Bad joke about quantisation, which, credit where it’s due, is the funniest thing to come out of an AI argument in months.

What actually struck me wasn’t the technical dispute. It’s that nobody in that thread, on either side, seemed to trust the companies making these claims. Not the American labs claiming theft, not the users cheering for the open-source alternative, nobody. There’s a comment in there that says the quiet part out loud: people are furious at OpenAI and Anthropic for locking things down and gatekeeping, and then furious at themselves for being furious, because the other option is cheering for a state-backed lab from a country with its own set of problems. You can hold both of those thoughts at once. I do most days.

I work in tech, not in AI research, but I’ve spent enough years around infrastructure to know that “how it actually works” and “the story someone tells about how it works” are two very different documents. The distillation argument in that thread is a good example. Distillation is a real technique, training a smaller model to mimic a bigger one’s outputs, but it’s being deployed here as a rhetorical move, not a technical claim. “They stole it” is a much cleaner story than “they built a competitive research programme with different constraints and got some good results.” Cleaner stories win arguments. They just don’t always match reality.

There’s an Australian angle here, even if it’s a quiet one. We don’t have a horse in this particular race, no local foundation model, no national champion to defend. Which means we get to watch the argument from the cheap seats, and honestly, that’s a decent vantage point. When the American labs lose ground to Chinese open-source models, the response isn’t “maybe our approach has limits,” it’s “they must have cheated.” I’ve seen that instinct in workplaces here too, when a competitor ships something good and the first reaction is to question how they possibly could have, rather than sit with the discomfort that maybe they just did the work.

None of this makes me feel great about where the industry’s heading, mind you. The environmental cost of training these things keeps climbing regardless of which country’s flag is on the data centre, and that’s the bit that actually worries me long after the distillation slap-fight is forgotten. Everyone in that thread is arguing about who owns the recipe. Almost nobody is asking whether we should be baking this much cake in the first place.

I don’t have a tidy answer for that one. I don’t think anyone does yet, and I’m suspicious of anyone who says they do.