We recently released REAPER v7.66, which is slightly interesting in that there was a small bit of LLM-assisted programming in there.

I'm not a fan of LLMs, but I do think what they can do is pretty amazing. Whether it is worth the cost, I have doubts.

Previously, on askjf:

Question: Have you tried Claude Code (CLI or VSC ext)? Opus 4.5 is *really good* (could be useful for tedius/boring tasks)
I haven't. tedious/boring things are good for you. either you figure out ways to speed them up (scripts to accomplish repetitive tasks), or you get the practice in doing these things, improving your skills. You train for marathons by putting in lots of easy miles along with your speedwork. The biggest win of LLMs IMO is discovery of poorly documented things.

...later on...


Question: I saw people pressing you here about supporting new versions of ffmpeg, and you replied it was a pain to do because of their constantly changing API. I just saw that you have done it. Well done! Was it a pain while doing it or it was not as painful as you envisioned??? Also, wouldn't their continous changing of API cause same issue again of latest version not being supported?

Turns out, this is was a task that an LLM could assist on.

Comment: What happened to tedious / boring things being good for you? ;-)

(lol)

My current thoughts:

• LLMs are great as a discovery tool for programming. Google has gone to shit, API documentation is terrible, it's really hard to find answers. LLMs end up indexing a lot of content that would be really hard to find otherwise.
• You can have high quality code produced by an LLM, but it will almost certainly take a ton of work by the human, unless you happen to be writing something that has been totally done before, in which case it will probably just rip that off.
• The better LLMs get at writing code, the more consolidation of wealth will occur, as independent software vendors will not be able to compete with the multinational megacorporations that control all of our infrastructure (read that essay). This will be a very not good thing. :/

My recent experiences, using Claude Opus 4.6, mostly via Claude Code:

• Updating REAPER's FFmpeg loading code. Up until v7.66, we supported FFmpeg v2.8/3.x/4.x. Every new major release of FFmpeg has subtle API changes. Navigating through all of those changes is a pain. I gave Claude Code our FFmpeg-loading code, and very little else, and asked it to support newer FFmpeg versions. It added support for 5.x, 6.x, and 7.x (it wasn't aware of 8.x as I guess it was trained before it was released). It pretty much succeeded; a few little tweaks were necessary, there were a couple of small errors, and it produced some preprocessor logic that was so incomprehensible it took me an hour to untangle. And I had to update it to support 8.x... butit still probably saved me a great deal of time (at least a few hours of very frustrating digging through headers and git histories).
• Updating FFmpeg support to support hardware encoders - it pretty much nailed this one, adding the necessary code which seemed to work.
• Updating our VLC loading code to support AV1 decoding. In this case I gave it only a few files relating to the VLC library loading. It gave me a working implementation, though it added approx 50 lines of code, which after reviewing, was mostly unnecessary, and I determined it would only take changing a few lines to support it properly. Still, super helpful.
• (this one hasn't been released yet) -- figuring out how to get consistent text rendering via CoreText on macOS vs win32. Initially the LLM gave me multiple solutions which were completely wrong (LOGFONT's lfHeight should be at 72DPI for macOS and 96DPI for win32, convert that way...), but there were kernels of things in there that were helpful (e.g. suggested ways of going from a positive LOGFONT's lfHeight to point size). I ended up getting something that almost worked, but I noticed that CoreText was returning obviously incorrect font metrics (particularly the internal leading). Once I pointed this out to the LLM, it suggested querying the OS/2 font table for the font (via a CoreText API I had no idea existed), and extracting the metrics from that. It produced code that almost worked, and from that, I managed to put all of the pieces together.

In all of these, I probably could have iterated with the LLM to get it to do more of the work. I'm not sure what time it would have saved, and even if it did save time, I don't think I would walk away from the experience with the same level of understanding I've attained going this route.

Based on that -- if you're vibe coding, right on. Enjoy it while you can potentially profit from it. Try to learn from it. If LLMs end up as master programmers, then the billionaires of today will be even richer (ripping off anything useful by burning tokens and then forcefeeding it to the world), and the rest of us, we'll just get the scraps.