~~Where's Christoph?~~ We found him and now we talk about AI

[iamlamprey]

It's very bittersweet: on the one hand I love the current technological renaissance we're in, on the other I feel bad about ditching HISE since it's where I first cut my teeth learning to code, and Christoph has obviously put years of his life into it (as well as answering my dumb as shit questions on the forum).

As for the LLM part of the discussion - the harness is 80-90% of the output quality, I "vibe-coded" my own harness (I'm sick of that term at this point) that has my C++ syntax force-loaded as a skill, as well as a working (bot-detection avoiding) web browser and other goodies. I use Deepseek Flash and the combination works as well or better than any standard Claude / Codex setups at 1/100th the price. It also uses grill-with-docs and ponytail liberally, two of the most popular coding skills - in fact my harness doesn't even have "editing" tools, it just writes python files and executes them to modify code, which works WAY better (at least for my setup).

Last thing I'll touch on is the idea of breaking things into reasonable chunks, the more tasks you give the model, the less effort it spends on each (I guess that's where subagents come in but I haven't needed to use them at all yet). A small list or one large feature per prompt / run, and make sure the agent cleans up after itself (diff checks, delete temp scripts etc).

Caveat in case someone thinks I'm stealing their .ch1 files with my "ch1 extractor" - I'm building out an MVP to see if I can convert my existing sample libraries into a synth and I can't be bothered downloading the original sample files lol

[David Healey]

@iamlamprey I Claude built a ch extractor, it's in my HISE repo somewhere on one of the Claude branches.

Here it is: https://github.com/davidhealey/HISE/tree/claude/monolith-ch-decompression-l1xy04

This was vibe-coded so I haven't upstreamed it, I just needed it for a one-off, but it worked.

[dannytaurus]

@Lindon said in Where's Christoph?:

I recall there was some stuff mentioned a few months back but Im pretty sure its out of date now...

That's why I'm hesitant to post anything along the lines of guide or instructions.

Things are moving so fast that anyone reading after a few weeks will be out of date.

But yes, a high-level config/setup would be good. Things like CLAUDE.md files, memory, etc. have always been around and will likely remain in place until the next big shift.

[dannytaurus]

@iamlamprey said in Where's Christoph?:

I use Deepseek Flash and the combination works as well or better than any standard Claude / Codex setups at 1/100th the price.

Is that a subscription or API calls?

You can't really compare DeepSeek API costs with the token usage you get bundled in a Claude or Codex subscription, especially if you're driving it hard.

Comparing directly to Claude/Codex API isn't fair either, since the vast majority of heavy Claude/Codex users are on a subscription and wouldn't dream of paying API prices for those models/harnesses.

[dannytaurus]

@iamlamprey said in Where's Christoph?:

Caveat in case someone thinks I'm stealing their .ch1 files with my "ch1 extractor" - I'm building out an MVP to see if I can convert my existing sample libraries into a synth and I can't be bothered downloading the original sample files lol

Don't worry - code and digital asset privacy is a thing of the past too.

It's always been possible to do all the reverse engineering that's all the rage now - the decompile libraries have been around for years. It's just that now LLMs bring us two things we didn't have before: speed and accuracy, and both at a much lower cost than hiring a human.

[Lindon]

@dannytaurus said in Where's Christoph?:

@Lindon said in Where's Christoph?:

I recall there was some stuff mentioned a few months back but Im pretty sure its out of date now...

That's why I'm hesitant to post anything along the lines of guide or instructions.

Things are moving so fast that anyone reading after a few weeks will be out of date.

But yes, a high-level config/setup would be good. Things like CLAUDE.md files, memory, etc. have always been around and will likely remain in place until the next big shift.

yeah thats what everyone says..so we all end up with nothing. Which is WORSE than out of date

[dannytaurus]

@Lindon said in Where's Christoph?:

yeah thats what everyone says..so we all end up with nothing. Which is WORSE than out of date

Well, since you asked so nicely I'll write something up to post later.

[David Healey]

@dannytaurus I'm still using Claude code in the browser, is there any advantage to accessing it from a desktop app, other than local file access?

[dannytaurus]

@David-Healey Given how it sounds like you use Claude, might not be too different.

If you only use Claude to propose/write/update code that you then read and edit, doing it in the browser works fine.

But working in a browser means your agent is running just in the repo, on an isolated cloud instance.

Using the desktop app (or the CLI as I prefer) is all local, so Claude can:

• launch HISE itself, run the hise-cli, Projucer, pluginval, auval, even launch a DAW
• run the local MCP for API lookups
• run the edit/recompile/screenshot loop with the HISE REST server for UI tweaks
• work in any folder on your machine, not just the project repo: the samples, AppData folder, etc.
• use your existing local credentials so you don't have to put duplicates in the browser

[iamlamprey]

@dannytaurus said in Where's Christoph? We found him and now we talk about AI:

Is that a subscription or API calls?

API, I also have strata running the Qwen flash 110b locally but my system is the "bare minimum" so it's not super reliable.

The Chinese models are all distilled from Claude / GPT anyway, they just focus more on optimization via KV cache tweaks or attention trickery, hence why they're so cheap. Deepseek recently published their entire inference engine for running models at scale, and I think the Kimi team posted the full training run of their latest model, including the costs which is pretty cool.

I've also been training a tiny (<10m param) dual-mamba model designed to fetch information at runtime instead of storing it in the weights, getting it to form coherent sentences for the first time was cool to see :D