kherud

So far I haven't seen a single model succeeding at transcribing sheet music, but I just tested it again with 5.6 Sol and it nailed the small test case. Fluently reading music requires multiple years of training for most people, but I feel like accurately following the horizontal lines trips up vision models in particular.

Imagine what amazing SVG generators we could have if Simon had randomized the target image from the start (and companies wouldn't just overfit on pelicans).

Feedback about creatine often seems mixed. Many love it, but many also report problems. Just in this thread there are people talking about heart palpitations and sleep problems. I have the theory that these side effects and mixed experiences come from impurities of lesser quality products. Apparently there are mostly two sources: Creatine mono-hydrate which mostly comes from China and "creapure", which is a patented formula known for its purity. Does anyone have insight into how substantiated this theory is? How likely are negative effects because of pollutants/impurities?

LLMs are extremely capable at problem solving. Presumably because you can autonomously learn a lot of it. But can you somehow account for things like long-term maintainability and code quality (whatever that means) or do you always have to rely on either existing high-quality code-bases (pre-training) or human curated datasets? Since you can't really quantify these properties (as opposed to: the problem is either solved or not), does this restrict autonomous improvement in this area? Are there benchmarks that consider this? Could Claude Mythos create an ultra-quality version of Claude Code or would it still produce something similar to earlier models, which are already over-sufficient in individual problem solving capability.

SQLite seems very powerful for building FTS (user enters free text, expects high precision/recall results). Still, I feel like it's non-trivial to get good search quality.

I think the naive approach is to tokenize the input and append "*" for prefix matching. I'm not too experienced and this can probably be improved a lot. There are many settings like different tokenizers, stemming, etc. Additionally, a lot can be built on top like weighting, boosting exact matches, etc.

Does anyone know good resources for this to learn and draw inspiration from?

That was my association as well! Dune even uses similar vocabulary. For example someone mentioned "pranayama" in this thread, which sounds a lot like Dune's "Prana-bindu". Really makes me wonder about Frank Herbert's experiences about all of this.

Aren't LLMs much more limited on the amount of output tokens than input tokens? For example, GPT-4o seems to support only up to 16 K output tokens. I'm not completely sure what the reason is, but I wonder how that interacts with Chain-of-Thought reasoning.

Interesting, thanks for sharing! Do you have an explanation or idea why compilation slows some architectures down?