Comment by overfeed
17 hours ago
While it sounds like a lot, do you suppose 3.4 million sessions come even close to being sufficient to train a frontier model?
Assuming each session was 10,000 words each, that's 34 billion words; lets call it 50 billion tokens (0.05 trillion) unfairly pilfered from Claude. That left Moonshot needing to scrounge for the other 14.950 trillion training tokens required for a baseline frontier model.
What do you think those tokens are used for?
Distillation attacks aren't about replacing the entire pretraining dataset with questionably sourced synthetics. It's all about post-training.
Train your own base model - but tune it off Claude output to make it perform more in line with Claude. Yoink the products of Anthropic's expensive SFT, RLHF and RLVR work for yourself by training on the outcomes.
The post-training datasets are small, but they are what controls the final model behavior.
> Train your own base model - but tune it off Claude output to make it perform more in line with Claude
Is that actually genuine distillation though? Distillation suggests the core model is being pre-trained using output from another model. For the above to work, you have to already have all the core intelligence trained into your base model.
If distillation just comes down to post-training then it's tantamount to admitting that the Chinese base models are just as good as frontier US lab models. Because you can't post-train frontier intelligence into a model. It has to be there in the base. Then you can change how that intelligence is expressed through post-training.
What's in the base model is "bits and pieces of intelligence".
You have to bring those bits and pieces together, put them into the right shapes and fill in the gaps to get a model that actually performs. This is what post-training is all about. It's not at all a trivial thing.
Reasoning, tool use, agentic behavior - all of those are post-training performance gains. Getting a good well trained base model is putting your foot in the door of frontier performance - post-training is how you actually get inside.
See: GPT-4.5 vs o1. One went for "build a bigger better more capable base model", the other went for "take the old base and post-train it for advanced capabilities". The results: a wider base with basic post-training loses to a narrower base with advanced post-training. Or, hell: GPT-3 vs GPT-3.5. One was largely a research lab curio, and the other kicked off the AI revolution as we know it.
The gains compound. Getting a better base model with the same type of post-training helps, see: the jump from Opus to Mythos/Fable. But post-training techniques account for a lot of the performance juice.
And yes, reasoning trace post-training distillation is "genuine distillation". As is logit distillation in pre-training. "Distillation" isn't a single training recipe that you have to follow to a tee - it's a large group of training methods. I've seen plenty of wacky things like inverse distillation bootstrap and post-training self-distillation that use distillation in strange ways at different stages of the training run to get results.
How does yoinking outputs from from prior generation Claude model and post raining on them result in a model competitive with the latest generation? That doesn't add up - nevermind Anthropic hasbeen summarizing thinking tokens since January to counter distillation.
Do I really have to explain the shape of AI training pipelines to you?
Train a big, wide base model with a lot of potential. Mid-train or post-train that on Claude Opus 4.5 reasoning/agentic traces (i.e. Claude Code data from Chinese API resellers) to make your model approximate a high baseline of chatbot behavior, reasoning, agentic work and tool use.
Then run your own expensive SFT, RLHF and RLVR on top of that yoinked baseline to dial it in further.
Actually doing RLHF and RLVR is extremely expensive. Distillation gives you a lot of dense, high quality post-training signal for cheap. This can get your model into the basin of "the right way to tackle this kind of problem" without a frontier lab compute budget. It's a big shortcut that gets you closer to the target - you can take it from there and build on top of it with your own work.
Also, it's unclear whether "summarizing thinking tokens" actually ruins distillation, or just makes it work worse. I'd bet on the latter, really. Because it's an approximation game, and summarized reasoning is still a better approximation of true reasoning than most of what you get online and in pre-training datasets.
3.4 million is the number of sessions Anthropic detected. The actual number of Claude sessions trained on is likely >100 million. There are tens of thousands of accounts funneling Claude sessions into Chinese labs https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
They are used for post-training, i.e. calibrating the model to understand and use tools/command line more effectively.
> 3.4 million is the number of sessions Anthropic detected. The actual number of Claude sessions trained on is likely >100 million.
That's an increase of only a single order of magnitude, increasing my estimate of exfiltrated tokens from 0.05 to 0.15 trillion - a far cry from the 15 trillion required.
> They are used for post-training
Possibly - it may be too much data for post-training, unless further curation was done. However, this is not distillation; you know it, I know it, Dario knows it, but "Distillation Attack" is a short, memorable, sciencey-sounding, political sound-bite with enough malevolence to be deployed on the floors of congress, or by the usual fear-mongering newstainment talking heads.
You're conflating pre-training data volume with post-training data volume.
Nobody is suggesting Moonshot used 15 trillion tokens of Claude data to pre-train a base model from scratch. That would be impossible and nonsensical.
This is entirely about distillation, which happens during post-training (alignment and SFT). Here, datasets are measured in millions or billions of tokens, not trillions. 50 billion Claude tokens is far, far than enough to copy Claude's reasoning logic, writing style, and tool-use ability to the pre-trained base model.
> However, this is not distillation
I don't understand how you're so caught up on the term "distillation". Distillation is using a larger model's outputs to train a (weaker) student model. Which is exactly what's happening. It's a standardized term that has been in use for a decade.
1 reply →