Anthropic says Alibaba illicitly extracted Claude AI model capabilities

So when Anthropic uses millions of copyrighted works to train their model, that's fair use, but when Alibaba uses Anthropic's model to train their own, that's infringement?

The Coca-Cola formula was reverse engineered in 2026 by a sufficiently motivated individual.

Here it is.

Per liter of cola:

104 g sugar

1 mL Flavor Solution A

10 mL Flavor Solution B

Carbonated water to volume

Flavor Solution A (Essential Oils):

Dilute 20–21 mL of the following oil mixture to 1 L using 95% ethanol:

45.8 mL lemon oil

36.5 mL lime oil

8 mL tea tree oil (emulates decocainized coca leaf extract)

4.5 mL Cassia cinnamon oil

2.7 mL nutmeg oil

1.2 mL orange oil

0.7 mL coriander oil

0.6 mL fenchol

Flavor Solution B (Chemical and Color Base):

Dilute the following ingredients to a volume of 1 L using water:

320 mL Shank's caramel color or 190 mL Durkee caramel color

160 g glycerin

45 mL 85% phosphoric acid

10 mL vinegar (5% acidity)

10 mL vanilla extract

10 g wine tannins (emulates decocainized coca leaf extract)

9.65 g caffeine

I'm more excited by open weights models you can't self host and need to spin up on H200s (RunPod or bare metal). This is where the real power lies and is where the open source world will trend.

It's far cheaper to spin up an H200 hourly or to simply consume a managed version of an open weights model than it is to use a proprietary hyperscaler API. And you own the model itself and can fine tune, tweak, lobotomize, etc.

The stuff you can run on your own RTX cards is neat, but it's rather hobbyist. The real power is in the cloud. Renting cloud hardware is fine, because the core problem is ownership of the weights, not the server rack or ISP fiber lines. Those are already commodity.

Big businesses will eventually run open weights models in the cloud, and it'll be a rather large part of the future AI economy.

The horse has bolted some time ago on this; the "frontier" is not as inaccessible as it once was, and open models, once out there, can't be put back in the bag.

Even if the US bans opens models, the Chinese and Russians will still have them, along with the rest of the world including cybersecurity attackers, and that's probably the worst-case scenario for the US.

The only way forward now is open models and how we restructure society around them.

That's not how it works though. When you prepare the conversations for distillation, it's the most trivial and obvious first step to replace "Qwen" with "Claude" and vice versa. I doubt they'd simply forget to do it.

A model may misidentify itself due to the surrounding context. When a model is about to answer "I'm ...", what follows is a sorted list of probabilities for what the next token should be. In most models it's usually a list of popular model names: say, in the list, first comes Claude, then Qwen, then ChatGPT etc. Usually the "Claude" token would be the most probable token, say 70%. But if the surrounding context is in Chinese, the embeddings for "something to do with China" may nudge the combined embedding of the output token towards the "Qwen" embedding more ("China+Claude=Qwen" in the embedding space). Say, the probability for "Qwen" now becomes 60% instead of 10%.

If we also use high temperature for more "creativity", the token sampler now may choose "Qwen". It's not the most probable token still, but it was chosen because selecting the 2nd most probable token once in a while usually allows a model to explore unexpected "creative" paths, and 60% probability is good enough compared to 70%. It's basically a hallucination.

I once made an experiment: if I ban the word "Qwen" in the inference engine entirely, and ask Qwen "which model are you?", it happily starts announcing it's Claude 100% time, simply because "Claude" is the next most probable token after "Qwen" in this context.