DGX Sparks (Was Re: DX / FM1 Librarian and Editor)

[Jay Vaughan]

> I have my agents really dialed in on classic software and data science practices. They run in a very old school, gated design, dev, test, release cycle. I've found it works very well compared to other in the org who are more vibe focused.

Very good.


> It's funny that being away so long was an advantage when I came back. My stuff works, is documented, and maintainable and observable. The same can't be said for lots of the other Ai or AI assisted projects in the enterprise. 


Yes, we old guys do still have some tricks to teach the newbies and LLM freakzoidicalbuttmonkeys making insane things.

LLM-produced code is only an issue if your existing software methodologies are already broken.

Which is to say, if you don't already (pre-LLM) have a proper software production workflow, internally, you will have to fix that before you can comfortably apply LLM-driven development to your organization.

That means, proper software coding practices, proper review processes involving real humans, proper code-reading sessions, proper automated testing and test-driven development phases, proper release engineering (coordinated with marketing schedules), and so on.

Not a lot of LLM-neurotic hackers have these workflows in place - but it is also the case that not a lot of professional software developers have these processes in place, either.

Quality is a result of constant review.  Who the heck reads LLM code?  "Just ship it" is the scourge of the software engineering world, and has degraded the subject for decades.  Fix that, hackers.

> 
> I've mainly been building heuristic services that then feed the AI analytics so it's good, verifiably correct data going in.

Right, you’ve got an analysis pipeline working.

> What I really like is having my ai dev team but also an AI engineer managing my servers. It's funny when they complain about each other. 

Yes indeed, it is often amusing to pitch one model against the other - something I like about having the local DGX environment is the ease with which I can pitch models against each other… its just a script.sh to switch things up, and this is of course driven by existing well-polished CI workflows, etc.

Crazy times, honestly.  SUPER fun to see the music technology world getting all lit up with a sudden influx of new firmwares for old things.  From the DSP emulators capable of running the SYSEX firmware files directly, to the $60 ultra-cheap device suddenly capable of eating every other synthesizer manufacturers lunch, it sure is going to be WILD to see what the incumbents do, in response.

I know a few folks in the German synthesizer industry who are having conniptions about having to change their product strategy for next year, given the M-VAVE FM1 situation .. and in the Japanese quadrant, there are huge waves too.  Let’s see how KORG responds to this “People Will Just Use AI/ML If The Manufacturer Doesn’t Release an SDK” threat that is very, very clearly upon us.

Someone should get all the devs from all the incumbents, and train a model out of them, lol ..

j.