What does this do?

[paulca]

Tonight I open an old project... ironically while listening to tunes via it as my primary audio pipeline.  It has issues.  a bit rough on the edges.

I ran two commands:
git clone git@git:MixAndEQ.git
claude

and this is how that went:


Biquads? I wish.

But I am being picky am I?

When challenged..

Honestly, I got lost. Sounds "plausible", but other than letting it "go for it" and flash it to the device... I have not spent the time to validate it.

[negativ3]

Last time I tried Claude, it identified a wrongly defined variable as a show stopper but didn't realise that variable wasn't used in the code, only declared.

It can do great things with detailed supervision/scrutiny.

[paulca]

Last time I tried Claude, it identified a wrongly defined variable as a show stopper but didn't realise that variable wasn't used in the code, only declared.

It can do great things with detailed supervision/scrutiny.

"It's" just a tool like KiCAD, try not to get attached to it.  It changes, gets better... and worse each release.

Honestly when I see it make mistakes, I just smile.  "I" see it, thats the important part.

There are somethings it's dealing with in context here that I "dont" understand. so I opened it up to the wolves.  I can post the project source from gitlab.

Back in level 0 context,  Converting to BiQuads I hoped would give me a performance boost out of the box with parallel stereo pair processing.  However, the bot is suggesting that it's (CMSIS?) more suited to block processing... offering block size =1?

Anyway... the bigger "take away" is for those who have NOT used Claude CLI or codex?...   That first screen shot was the only prompt.   It took over and went and "LOOKED".

  ⎿  Done (23 tool uses · 52.5k tokens · 1m 7s)                                                                                                                       

The explore agent.  It runs these two or three at a time in parallel.  If you "team up" with it and agree to do something it will launch it's tasks and concurrently and then ask if you need help with your task.

Explore, plan, act are the three main "meta" agents.  They use "skills", "tools (aka commands etc)" and plugins to execute actual terminal commands to act.

It has real power.  I got it to be "root" and configure OS level stuff tonight.  Because if I run it as me, it can do as me.  No, the power and security reality is not lost on me.

It is very, very unlike your average chat bot.  Designed to be an engineers first officer.  It will shot the sherif if you tell it to.  Its like the C++ of the LLM world.

[stj]

remember the old "Dr.Who" series where you deal with robots and computers by asking them a regenerative question that locks them into a continuous loop trying to formulate the answer?

maybe it's time to try such things 

[Fire Doger]

You gave root to a tool that executes string from random git repos... 

[az1]

You gave root to a tool that executes string from random git repos... 

The problem will solve itself.

[paulca]

You gave root to a tool that executes string from random git repos... 

The project was written 2-3 years ago.  AI have never even seen it.

And root?

[woofy]

... It changes, gets better... and worse each release.

Interesting you should say that. I've felt that Opus4.6 does not do as good a job as Opus4.5. It's hard to quantify, I'm currently working on an STM32-H563 project and felt that Opus4.6 is both slower and goes off focus a little easier than Opus4.5. YouTubers seem to rave about it though, so maybe its better in the pretty graphs to impress the boss BS department. Wish I could select 4.5 to really compare, it could be just the different section I'm currently working on.

[paulca]

... It changes, gets better... and worse each release.

Interesting you should say that. I've felt that Opus4.6 does not do as good a job as Opus4.5. It's hard to quantify, I'm currently working on an STM32-H563 project and felt that Opus4.6 is both slower and goes off focus a little easier than Opus4.5. YouTubers seem to rave about it though, so maybe its better in the pretty graphs to impress the boss BS department. Wish I could select 4.5 to really compare, it could be just the different section I'm currently working on.

I noticed differences yes.  On the plus side I gave it a "big" task with multiple components and it produced a plan and implementation on it "better" and more complete than Sonnet 4.5.
However.  I also noticed it taking considerably longer on all tasks, even simple ones.  I also noticed it costs a sh1t ton more in usage.

I tasked is with something really chunky tonight.  Let it plan, but ... I am not actioning on the plan.  It just revealed to me that there are decisions to make first.

To stop talking in riddles.  I am designing an OS.   Home brew.  I have a binary loader and a "run" command to execute in memory.

I extended this after careful consideration to have the "bin loader" allocate memory, not accept a hardcoded address, and return a Process ID.  Then "run" can run that.

Thinking this was a good point to introduce a syscall, I asked claude to add "Trap for yeild", no multi-tasking, kernel yeild just prints "Process N yeilded" and passes control back.

It thought for 5 minutes and produces and a plan that I don't doubt would work.  The reason I am aborting it, is because it contains a lot of smaller chunks of work with assumptions Im an not ready to "yield to it".  In particularly it is introducing the concept or a context switch between kernel space and user space.

It's not wrong in anyway, but I don't want to just hand it full design control over those features.  So, instead I'm going to pull it back to just that context.  "Kernel vs. User", lets face it in isolation before we actually need it.  So I get to decide what and how that looks like, not just for the current use case but or the others that come after dependant on it.

There is, bizarrely merit in taking the output from one context (Claude CLI planner) and putting it back into another, like a Claude (or GPT) chat and asking for a review.

It can also be worthwhile "chit chatting" about things for a while beforing actually engaging it to help with decisions.  The chit chatting can establish "ethos", "approach", "preferences".  That can obviously be both a bonus and a risk.  You manage.

If you find yourself having to "setup" the basic approaches everytime, you can add them in many different places to always be "in context", from system promopts to config to readmes to documentation.

For example, my OS design introduces it's first invariant.  "No unbound buffer access."  No "gets" and friend, period.

[paulca]

Its (the chat bot interface) content filters are becoming a bit annoying.  They, as simple rules engines tend to trigger a bit too often.

I used the term "dumb ass" in a prompt and it rejected it. Yet it will let me F-bomb all the time and f-bombs back at me.

I postulated a Z80 application that prints lines from Eric Idle's Bright side of life and it content filtered itself for "nihilism" due to "Life's a piece of shit when you look at it".  Similar results if you explore the nature of "free will" with it.

It's usually easy enough to work out why and reprompt around it.


EDIT: Why is this important in an engineer context?  When you start talking about killing children in the GTFO handler that terminates processes, the terminology can get into dicey grounds.  The LLM can track the metaphors but the content filters can't.

"Implement execute_all_children(), ensure that all expire and self terminate before returning."

[stj]

killing your "slave processes" will get you reported to the police!
LOL

just another reason to avoid "other peoples computers" aka "cloud" shit.