Agreed, and from the little experience I have, this is my primary issue with full and/or automated LLM contributions (no matter it be OSS or in business).
100%
I’ll bring the marshmallows
I haven’t watched it, but I’ve definitely seen it pop up in various communities, a video caled “The Infinite Software Crisis” by Jake Nations (Netflix). Regardless of what’s stated in the video, I’ve heard murmurs of a similar concept, that if LLM coding gets good enough, who’s going to buy any software? Certainly don’t see that being the case for a long time, since the only way to make it viable for everyone is that a model that could generate purpose-built code on the fly as necessary without any user guidance. The generation would have to be really good too, otherwise all the previously mentioned issues would still prevent this reality.
If we accept the hypothetical premise of a reality where LLMs were so capable there was zero user effort needed, and infinitely small probability of significant errors, for any software to be written at all, the idea becomes more important. Perhaps it would be best to leave this for a new topic discussion, as I worry it would distract from the wider OSS conversation, but I think it’s an interesting question, especially b/c of the subjectivity in what’s “acceptable” means that “perfection” may not be necessary. I’m going to mull this over and if no one else posts a topic about it first, once I’ve collected some thoughts, I’ll try to start a new topic around this.
Now, back to the point you actually made. It would be quite fantastic and entertaining to watch large organizations, especially monopolistic ones, get usurped by someone slinging code via LLMs. Of course there’s lots of ways this could go sideways quickly, especially for a browser, but it could make it a bit more entertaining too.
Not sure if the idea was a Musk comment made the headlines, but at a certain level of advancement, software itself we we know it today becomes obsolete. The interface transitions from one of clicking/tapping, filling out forms, or even being aware of any underlying structure to one of just natural language exchange. For a layperson, interacting with a machine becomes indistinguishable from interacting with a person—behind a screen, over the phone, or in the “flesh.” All the rigorous logic to fulfill a natural language request is block-boxed. We don’t speak in binary or character sets, so the machines will have to become closer to us to close the gap. It’s a logical end to moving machine and humanity as close together as possible until somebody figures out a way to physically merge the two.
I’m not saying we’ll get there in this lifetime (if at all) or with the current technological trajectory, but at least some people have had these lofty thoughts and it doesn’t seem to be a stretch that it’s even possible.
It’s a scary thought for a nerd though. We’re the type to find meaning in discovering structures and tapping into raw underlying interfaces. We’re the kind of people who will customize a restaurant order by specifying the individual ingredients (e.g., “no onions”) and cooking methods (“medium rare”) versus those who pick from a numbered photo or ask the server for a recommendation.
I’m not even sure this is true anymore, atleast for LLM’s. If ChatGPT is a indicator, from my experience it peaked at 3/4 ish and has been progressively gimped since.
It also important to note that since it was originally released, it wouldn’t surprise me if it has inadvertently been trained on its own output which poisons the entire model
Yes, I have been encountering this a lot in OSS repos. Human-written code + AI-written issues/PR descriptions. Also Copilot reviewing every commit. Usually what is being said makes sense and matches the work being done, but it takes many, many more sentences and unnecessary headings to get there. Or restates what has already been said. What was once 4 sentences or bullet points is now 4 paragraphs of fine prose.
Right and that’s fine, but I can see that myself. What I can’t see and what they don’t express because they can’t know is the intention behind a change.
I can see what was changed but I want to know why it was changed.
The descriptions being overly long is honestly probably just an issue with the prompt, LLMs are able to write bullet points too.
I’m a very experienced engineer, with over 40 years working for some of the biggest and most prestigious banks in the world. I use vscode+copilot. I’m fine with using AI myself, because I’m on top of it and would stake my reputation on how it speeds up my work. I have a deep hatred for vibe-coding though and can understand if projects want to ban it completely. In my case, you wouldn’t know that I’d used AI assist other than the fact that I seem to get a lot done in very little time. Tools like the context7 mcp server have really sped things up a lot, and I’d not want to work without it now. Banning AI is ok, but banning humans who can show real, high-quality, acceleration, is silly IMO.
Ha! Well, most people are senior to me at least, and I guess really they never claimed to be senior. I called them senior since they’ve been programming longer than me, but as with many things, more experience doesn’t necessarily imply greater skill.
My understanding, or at least my interpretation as I continue to reflect on this discussion, is that the ban is to stop slop code and minimal effort attempts at contribution (including comments). I suspect that anyone who can use the tools effectively, intelligently, and with consideration to the maintainers would be welcome to contribute (assuming all other contribution guidelines were followed).
I’m going to mull this over and if no one else posts a topic about it first, once I’ve collected some thoughts, I’ll try to start a new topic around this.