[himata4113]

Never claimed that it is irrelevant. And kimi k3 is on the same level and sometimes outperforms fable 5 - that cannot be explained by distillation. The reason why gap is not closed is simply the fact that fable was trained months ago so in theory the frontier labs are still 1 (small) step ahead.

Although I will reiterate the fact that distillation is not the primary reason why these models are performing so competitively.

[atleastoptimal]

Closed-source models have to deal with the current frontier being heavily regulated. Fable, at its old level, was "too good" to be released and they had to add an additional safety layer to sanitize the outputs. Lowering the quality of the models so they are safer and more steerable has been something all the closed-source models have been doing for a while, a requirement that many open source models don't need to deal with.

If Kimi k3 really were above Fable 5 then there invariably the USG would have to consider their restrictions on model capabilities excessive, or one would have to admin closed source models are held to more restrictive safety standards than open source models.

>Although I will reiterate the fact that distillation is not the primary reason why these models are performing so competitively.

How would you know this? How could you ascertain exactly how much performance is attributable to their unique engineering/research? If they really were so competitive they could surely make a model that isn't dependent on distilling Fable or other frontier models.

[himata4113]

I recommend reading some of their research it's honestly astonishing how intelligent some of their solutions are.

Kimi specifically relies heavily on reasoning traces which is largely due to their training strategy and will perform poorly when thrown into a conversation from another model. Another fun advancement is that they simply ctrl+c ctrl+v'd attention which means that the model can steer where to look in the context window without ever producing an output token increasing token efficiency and attention accuracy as a side effect you end up with weaker prompt adherence.

None of these 'issues' manifest in US models which proves that kimi has diverged and is achieving these capabilities seperately from the architecture that US labs rely on.

I would agree with you during the Deepseek R1 era, but US labs were heavily inspired by open research at that point as well so I wouldn't give them too much credit.

[spwa4]

You still left out that it doesn't matter anymore, just like Anthropic/Facebook/OpenAI only really needed to read massive amounts of copyrighted data only once (and of course, they all did this illegally, which makes their current complaints more than a little ...). Once they have a large model trained on the data, they can just retrieve reasoning traces and copyrighted data from the previous model. In fact that is a training technique long used because it has better results that directly training on the original data.

In other words: even if the US (somehow) denies them access to the current OpenAI/Anthropic models, they'll be able to improve based on what they already have.