graphitout

For a chatbot like claude code, since the instruction prompt part (including descriptions of tools) is relatively constant for multiple users and over a long time, a lot of optimizations can be made. Even basic prompt caching gives a lot of speed and cost reduction.

We are not there yet. A bit too optimistic for today. May be in 1 year.

I added types to a python codebase (50k) using claude code and it still took some effort to make mypy happy.

Very nice. Just the thing I was looking for.

Yes. Most of the MCP based search lookup now happens over the text stage.

voicevoyage.io looks interesting. Will keep an eye.

Quite funny to see that coming from OpenAI. Nevertheless, how can they prove that?

What is the strategy to deal with pay-as-you-go services used in vibe coded apps? I am always worried about the risk of some bad loop consuming thousands of dollars in a day.

Very interesting. The moment I saw the repo I looked for a python reference implementation and found:

https://github.com/antirez/voxtral.c/blob/main/python_simple...

Thanks a lot! Let me see if I can get it to work with pure numpy while keeping the input audio part streaming.

Claude gave me $50 yesterday with the release of opus4.6 and I burned my session quota + $30 yesterday itself. The code changes were very reasonable.

It is good for handling all code changes where there are no security aspects involved.

The RAG was setup on a bunch of documents, most of them were manuals containing steps about measurements, troubleshooting, and replacing components of industrial machines.

The issue was that most of these steps were long (above 512 tokens). So the typical chunk window wouldn't capture the full steps. We added a tool calling capability by which LLM can request nearby chunks of a given chunk. This worked well in practice, but burned more $$.

I do not follow. Could you clarify please? Who is going to fund and what is the criteria they are using?

optimal chunk size is strongly query-dependent - very true.

Faced similar issues. Ended up adding a agentic tool call layer on the top to retrieve the nearby chunks to handle a case where a relevant answer was only partially available in a chunk (like a 7 step instruction in which only 4 were available in a chunk). It worked ok.

Nice visualization. Would be nice to also see state variables (like stack or queue) along with the graph.

You are right.

- llm_generate() - this is the core function which calls ollama API

- get_related() - this will give a sequence of related topics for a given topic and description

- llm_get_aspect_query() - this is tricky. The interface has a fixed set of aspects for every response. Like history, related ideas, people etc. I wanted a way to create new query based on an existing query and a given aspect. This func does the rephrasing of the query.

- App.vue: handle_related() - there is a tiny bit here. When going to another "suggested" topic, it seems we have to give it some context.

Very interesting. Eight healthy subjects used for testing. I am guessing that the effect is not limited to humans alone. Shouldn't it be easier to do a larger study on animals?

Is there an example NNET to see how it is used?

First focus on inference. It is a lot costlier to have GPUs for inference.

Additive synthesis was a pain to configure. I had to model and estimate the behavior of each harmonic. It was too much effort.

Subtractive synthesis didn't work that well for me either. I believe my implementation of the variable cut off lowpass filter had something to do with it.

FM Synthesis worked reasonably well without much effort. Implementation was rather straightforward.

You are right.

I needed that image to explain the parameters in the ADSR configuration I used in the project. Mainly the slevel1 and slevel2 which captures the slow decay during sustain. Most ADSR diagrams show sustain as a constant which didn't really sound that natural when I tested for my projects.

Sure. Audio files generated have no license. Do whatever with it.

Only the code has license, but I have no intention to enforce it.

I am playing gooberdash now.

I had some fun with the strings in the karplus-stress-tester. Thank you for that link.

I am thinking of moving the core to WASM. But I may have to drop the python part and make the calls directly from javascript.

Glicol looks impressive. I will try it out.

I like the tool. Is there a way to make the app move to the middle. Currently all the content appears on the left side.

I will try it out. It looks straightforward.

For UI, I used a tiny tool I wrote: https://github.com/charstorm/linkitall/

It is the same in software development as well. I would rather take review comments from an automated AI system than from my colleagues.

Thanks. This will come very useful.

I did a quick test. It worked well.

Is there also a way to remove lip & mouth sounds?

This is very interesting. I am working on a similar project related to generating dependency graphs (basically only type of connection is A depends on B). I tried to get the common chatbots to generate the data in the form of dependencies, but I didn't have much luck.

Example graph (hand written): https://charstorm.github.io/class-11-12-india/class11/maths/...