How would you prove they trained on your data specifically?

Showing that a novel mathematical approach was in your prompts shortly before the model “proposes” that approach to relative amateurs would be ideal, if only we had a case like that…

and still they denied it - it only shows that we have little recourse.

It’s possible that they didn’t train on it, and the approach was derived by the LLM.

What’s a problem is that they haven’t outright denied it. That could be caution and them doing their due diligence first, it could be that it was intentional and they didn’t expect to get caught, or it could be because they have no way of knowing themselves.

It's likely much closer to that price if you don't calculate with subsidized token pricing.

Once a model like Fable/Mythos is trained, that was used for this project, the inference costs are profitable. CapEx vs. OpEx. Every new Zig->Rust port of this caliber using Fable/Mythos costs the standard token price, it's not subsidized. The model already exists for the original CapEx, you can do 1,000 ports of this for standard token pricing.

But when do you get that type error?

This is the important bit.

You get it after the app is deployed, the query is ran and a result is expected.

When do I get a type error from my language if it's statically typed? That's right, before I even deploy.

Can you explain how COT is essential to distillation and why obfuscating it helps defend against distillation?

Or point me in the right direction in terms of what to read.

Because they are an integral part of the solution generation.

By distilling, you are training a model via Reinforcement Learning (RL) to mimic the answer of a bigger model. To do that, you need all the steps that contributed to generating an answer.

To give you a better idea, imagine teaching a student how to solve math problem:

1. You give it the problem and the answer only (no thinking trace) 2. You give it the problem, the intermediary steps and the answer (full transcript)

I think you can agree that the second method is more likely to give a well-informed student.

Isn't distillation done through SFT by having the student model approximate the teacher's output distribution?

This might be helpful https://arxiv.org/abs/2608.09867

I'm getting a little tired of the "X is all you need" formulation

If the performance and accuracy was there, don't you think there would be a market other than memes?

I'm not speaking about that, he made a specific claim that AI videos wont get better and that was completely wrong.

I would argue models have been plateauing for quite some time now.

This is why I believe capitalism cannot survive AGI

I would like to add that for what it's worth, I wouldn't mind slower compilation times if it means we get the type system as described. If it really does end up having a significant impact on performance, would you be open to making compile-time type-checking optional instead of scrapping the whole thing?

If it comes to this, we will certainly need to have a discussion on the topic. So I am definitely open to discussing the idea!

what are the alternatives?