[OSGeo-Discuss] AI/LLM discussion thread

[Laurențiu Nicola]

On one hand, I personally find it fascinating, in a train wreck in slow motion sort of way. LLMs are a lot more knowledgeable and smarter than me, more hard-working and apparently even better at googling. I can only doom scroll and watch, terrified, how I'm becoming more and more obsolete.

On the other hand, I'm not convinced by most arguments against LLMs. After doing a bit of research, I believe they rely on cherry-picked, incomplete, misleading or even dishonest information. Yes, they "use" water, but I use more when I flush the toilet than in one whole year of LLM usage [1], not to mention the occasional burger I might eat or the clothes that I wear [2]. Yes, they use energy, but I use AC and I've been on so many flights in the past two months. Yes, some data centers use gas turbines [3], but most of them don't, and the energy usage pales in comparison to anything industrial (although for better or worse, it looks like we've decided to outsource our industry to China). Yes, companies lay off people and say it's because of AI, but I don't believe it, simply because nobody had incredible success using it: do you know any high profile, successful projects developed primarily by LLMs? I don't.

My main worry is that by using AI, we'll end up losing our skills, or never acquire them in the first place. There's also the impact of filing hundreds of automated pull requests against open source projects, which is unsustainable, but easier to deal with: like always, maintainers are under no obligation to review or accept pull requests.

On the third hand, I find most of the anti-AI crowd depressing. As I mentioned, there's a lot of misinformation (probably not even intentional). But for people who present themselves as more ethical than the occasional LLM user, they're all too happy to harass open source maintainers and even resort to death threats and other abuse [4]. If -- like some say -- I can't pick the middle road, I'm afraid I'll go to the "pro" side, where the worst people I see seem mildly delusional, try to sell something, or just pester open source maintainers [5].

On the fourth hand (I have many today), I see a lot of people using LLMs to build cool stuff [6]. There's joy in there and there's empowerment, and I wouldn't try to take that away from anyone. And there are adjacent use cases like machine translation, speech-to-text/text-to-speech and OCR, where LLMs and other recent models completely wipe the floor with the previous approaches.

So yeah, it's complicated and I'm pretty pessimistic. We're living the interesting times in that famous curse and we have no way out. I think AI is here to stay, even if OpenAI, Anthropic and Gemini close shop tonight.

On more practical matters, I try to limit my usage of LLMs. For code, I review, clean up, and try (not always successfully, especially in domains I know little about, or when I don't care) to understand their output. And I almost never let them do stuff unsupervised, except when I don't really care.

Laurentiu

[1]: I think it's more optimistic than my previous estimates but see e.g. https://calculator.tydalconsulting.com/
[2]: making 1 kg of cotton uses roughly 20 tons of water
[3]: correct me if I'm wrong, the impact of those seems to be estimated to e.g. about 0.5 ug/m^3 for PM2.5; my city is at roughly 20 ug/m^3 and nobody cares, some areas in the center of Paris hit 40 today, the WHO guideline is 15, and Knoxville is at 10
[4]: like in the recent rsync case, which wasn't even substantiated
[5]: yes, I've seen https://theshamblog.com/an-ai-agent-published-a-hit-piece-on-me/, I'll take that over death threats
[6]: I bookmarked https://news.ycombinator.com/item?id=48421175 as a recent (but less impressive compared to others) example; as another example, a blind person I know recently mentioned being able to play Civ V thanks to some mods developed using an LLM [7]
[7]: which got dismissed with something along the lines of "blind folks have the right to use LLMs for accessibility purposes, but they actually shouldn't be using them because LLMs are absolutely evil for [one screenful with the same 10 reasons already repeated N times upthread]"