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.

▲

tristanj 5 hours ago | parent [-]

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.

▲

overfeed 4 hours ago | parent [-]

> 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.

▲

tristanj 3 hours ago | parent [-]

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.

	▲	overfeed 39 minutes ago \| parent [-]
		There is a lot of supposition going on your part and mine. IMO, Chinese labs are not dependent on OpenAI/Anthropic outputs; they definitely use the outputs, but along other training/post-training data. Now that Anthropic hides the real thinking tokens in a way that precludes future CoT distillation, we'll find out which side is correct based on whether Chinese AI labs close the gap or not. My bet is they'll close the gap; nothing about frontier AI is magic, once something is shown to be possible, experienced practitioners almost always figure out how to accomplish the same feat, though not always on the same way. This is why frontier US labs keep leapfrogging each other every few months.