Technology

Which just shows how fucking stupid this current LLM-based AI approach is. There isn't a way to differentiate between data and meta data or instructions. It all just gets shoved into a prompt that might end up the length of a short novel by the time all the context has been added and read operations have finished. A tool so sensitive to its input that adding a period at the end of an instruction could completely change the output it generates, even with temperature (randomness) set to 0.

I'm not even sure this can be fixed. Like, even if they they try separating the instruction input from the supporting data input, LLMs don't follow instructions in the first place, they just predict text and having instructions in the context can strongly affect the output it generates. Meaning there are no instructions to separate from the data; it's ALL just data and platforms like Claude Code just give it the ability to do things with that predicted text that hopefully follows your instructions and uses your data rather than the other way around.

I think we're stuck in a local minimum of an optimization problem for AI because an LLM is much easier to make than a more reliable form of AI. You mainly need to throw a lot of text at it to train. There's probably other tweaking that goes into it, like a way to do more training using user thumbs up/down feedback, but it's just the big data approach of soaking up all the data they can find and just throwing it at a blank statistical model and see what it spits out.

If we want something like the Star Trek computer, I'm pretty convinced at this point that it's going to take a completely different foundation, but the industry is currently stuck on improving LLMs.

To a developer, "jqwik tests and code" doesn't mean jqwik itself. It means the tests and code written using jqwik.