I've been ruminating on this.

It's not that I don't think a machine can generate code. We've been doing that since the 50s at least. Hell, clang generates better assembly language than I ever will.

In fact, it's not even that I think code-specific LLMs like Claude or Codex are inherently bad at their jobs.

They're as good or bad as their training data, and no better than the user lets them be. You ask for innovation and it's going to disappoint you. Ask for a solution that's tried and tested and in common use, and it'll shine.

That also means it can find common bugs, since the bugs and their fixes were in the training data too. I don't have a problem with that.

What I have a problem with is that some people are just letting the models run rampant, doing essentially whatever, and acting like this is proof of sapience. The mental damage being done here is undeniable, even if not everyone is vulnerable to it.

Personally, I think LLMs are very likely more suited to generating code than they are English prose. And the compile/error/fix loop provides feedback that the machine can use to catch its own errors, something it can't do in English prose.

But untrained, inexperienced users who don't or won't understand what's really going on here are being stunted in their growth by overusing and misusing these things.

The truth is, I think some people will be more productive using an LLM as part of their toolchest than without. And others will not be.

I don't need one, but you could probably convince me to at least try it if doing so didn't give the worst people on earth more money and power. I'll try almost anything, and give it my best attempt at objectivity.

The copyright argument falls flat on its face with me, because I believe copyright to be a capitalist scam to benefit corporations and not actual creatives.

The environmental argument holds a lot more weight, and we should force these companies to address the issue before it's too late.

The copyright argument falls flat on its face with me, because I believe copyright to be a capitalist scam to benefit corporations and not actual creatives.

The environmental argument holds a lot more weight, and we should force these companies to address the issue before it's too late.

@mos_8502 OK, so what's the solution re: copyright?

I can get behind the “copyright is a scam” concept if:

(a) all LLMs are public domain, no IP protections of any kind whatsoever
(b) all output from LLMs are also public domain
(c) all hardware designs and foundry processes used to produce the machines (CPUs, GPUs, etc.) are also public domain, as are the OSes, UIs, etc.

Since none of that is happening or will ever happen, I don't see how your argument holds any water.

@jaredwhite Legally it doesn't -- but it's my position that copyright should be abolished entirely, replaced with mandatory citation for all derivative works. Because I hold that position, arguing the copyright angle against LLMs would be hypocritical.

I've been ruminating on this.

It's not that I don't think a machine can generate code. We've been doing that since the 50s at least. Hell, clang generates better assembly language than I ever will.

In fact, it's not even that I think code-specific LLMs like Claude or Codex are inherently bad at their jobs.

They're as good or bad as their training data, and no better than the user lets them be. You ask for innovation and it's going to disappoint you. Ask for a solution that's tried and tested and in common use, and it'll shine.

That also means it can find common bugs, since the bugs and their fixes were in the training data too. I don't have a problem with that.

What I have a problem with is that some people are just letting the models run rampant, doing essentially whatever, and acting like this is proof of sapience. The mental damage being done here is undeniable, even if not everyone is vulnerable to it.

@mos_8502 Just so we're clear (not trying to be disagreeable, simply curious), are you saying you think all IP protections should be abolished?

@mos_8502 the funniest thing is that it can find common bugs if you tell it specifically what to look for. It'll even find those bugs in code it just wrote, if told to do so. The horror is that most vibe coders don't understand that the common bugs, including common security bugs, are gonna creep in and you're going to have to hunt them and protect against them.

I've come to the conclusion that the job of a vibe coders is policing the AI via overwhelming use of verification tools.

@swelljoe I feel there's a useful and productive distinction to be made between a "vibe coder" and someone just integrating it into their workflow as one more tool in the toolbox.

In my experiment, I found that ChatGPT Codex does a wonderful job working with Git in fairly complex ways, for example.

@jaredwhite I would personally replace current copyright and patents with mandatory donation to the public domain, with legal requirements for the release of source code and schematics of any work or device - coupled with universal basic income and lifetime bonuses paid to creatives for their output being used.

@mos_8502 Interesting! I would love to live in that world. 😅

All I've been saying is a world where LLMs live as if copyright & IP protection isn't a thing (they can steal it all!) whereas the rest of us get thrown in jail for committing a similar offense is the worst of all possible worlds.

Personally, I think LLMs are very likely more suited to generating code than they are English prose. And the compile/error/fix loop provides feedback that the machine can use to catch its own errors, something it can't do in English prose.

But untrained, inexperienced users who don't or won't understand what's really going on here are being stunted in their growth by overusing and misusing these things.

@mos_8502
There's an awfully big difference between code that compiles and code that *works*.

@swelljoe I feel there's a useful and productive distinction to be made between a "vibe coder" and someone just integrating it into their workflow as one more tool in the toolbox.

In my experiment, I found that ChatGPT Codex does a wonderful job working with Git in fairly complex ways, for example.

@mos_8502 yeah, the machines are now better than me several categories of problem, and I'm using them a bunch...but, I'm still keeping a gun by the computer in case it makes any weird noises.

@swelljoe "Go digging in this codebase and find some bugs" is somehow a reasonable prompt now.