Hacker News
points by throwaway85825 3 days ago | 0 comments
It's going to be all about the training set. Theres no moat in models trained on internet data.

There are tons of private datasets where a model will be fine tuned to provide useful domain specific responses.

[–] jacquesm 3 days ago
And this is why I think companies will wise up and stop sending their information to the likes of Anthropic and OpenAI: they will not be able to maintain their edge if either one of them decides to defect and start using all of their uploaded user data to improve the model. So they have a massive incentive to break their own TOS.
[–] throwaway85825 3 days ago
That's the push factor. The pull is, say there's a giant repo like windows or Adobe. LLMs produce spaghetti code because the repo doesnt fit in the context and it wasn't in the training set. If Adobe fine tuned it the LLM would be much more useful.
[–] mattm 3 days ago
This will be the future model for businesses. Buy a generalized frontier model and finetune it and make it available internally. This model offers a lot more business benefits than the current risk.
[–] jacquesm 3 days ago
I'm running a max of 1M context locally. I think the next real move will be continuous learning/infinite context. That will be another game changer.
[–] throwaway85825 3 days ago
Infinite context requires Infinite memory.
[–] jacquesm 3 days ago
It wasn't a theoretical thing, it was a practical thing. Just like when we talk about 'infinite tape' when looking at Turing machines. As opposed to 'clear boundaries to how much context you can have because that is baked into the model'.
[–] user43928 3 days ago
The idea that noisy usage data is worth that much for training as to balance the fatal risk of the crime you describe, imo borders on a conspiracy theory.

I am under the impression that curated training pipelines are driving the gains, not utilizing random usage data in nefarious ways.