K3 is a 2.9T parameters sparse model that cost $15 per 1M output token or far less on subscription.
Fable is a 6T parameter (or 10T depending on how you count) dense model that cost $50 per 1M output tokens and can only be accessed via API.
You have to completely ignore Anthropic being an order of magnitude behind in the singular most important metric in performance to say Kimi is 2-3 month behind. The better question is how long will it take for Anthropic to make make a Fable that's actually usable in term of cost and efficiency.
K3 basically just destroyed the last bastion of American AI where they try to charge 100x for marginal capability delta, now they have to try charge 100x for no capability delta, or for regular users, for all practical purposes, 10x for an inferior product in Opus.
To put it another way, I can right now on Kimi subscription go to down and use K3 for everything without a single thought on token usage, if you do that with Fable you'd go broke in an instant. The real metric is, between two companies both with $100 budget, how much can each build if one use Fable and one uses K3. It's not even a competition.
Yup, agree completely.
People seem to overlook a basic point, especially those in the West.
Every single LLM that is competitive was provided to the public by a for profit corporation.
Simply put, AI in the form of LLM currently, is still a business.
The other important point that people do not talk enough about, is distillation, because everyone does it.
So, what does that mean? That clearly suggests the models will converge in terms of performance, because everyone is copying everyone else, to stay ahead.
So, therefore, if the AI current LLM product is always converging in capabilities, all that matters is price. Because it is still a business.
This reminds me of that expressions from the lawyers. This is what they say.
If the law is on your side, pound the law. If the facts are on your side, pound the facts. If neither the law or the facts and on your side, then pound the table.
Those American AI people, they are starting to pound the table, when they start to suggest is Chinese AI is really Chinese Yellow Peril AI.
That is not going to work, because this LLM is still a business, and too many mainstream and famous American companies are embracing Chinese open source AI due to the cost factor.
After all, it is still a business.
That is why I think we read too much bullshit all the time in the media.
If distillation is a normal practice in the development of AI models, then what is the outcome there? Convergence!
If frontier models are defined by compute power, then what is the outcome there? Most costs! Due to brute force methods!
Can LLM brute force its way to AGI? Notice how no one really talks about AGI anymore. It was a smokescreen all along.
