I agree with you re water. As long as they're not permitted to use evaporative cooling, the water use is fairly reasonable, in the long term (though it's a huge amount when first loading up—which imo they should be forced to pay above-market rates for and to fill extremely gradually).

Electricity though I'm not so sure. There's one that's going to be built near me (near in Australian terms—it's a 3 hour drive, most of the way at a clean 110 km/h). It's supposedly going to be allowed to plug directly into the nearby existing fossil fuel power station, where it will eat up basically the entire output. But if we set a requirement that they be entirely self-sufficient (or, say, 90% self-sufficient with a significant premium paid for any grid import) using only renewable power, then I agree, the electricity usage problem becomes largely a non-issue.

I'd also extend the environmental concern to add a third category, which is the impact specifically on the local community. The enormous use of what could have been otherwise productive land, the highly disruptive noise given off which disturbs locals, etc.

I also largely agree with you re intellectual property. I actually can't find a moral difference (and the legal difference is yet to be clearly defined and might end up different across jurisdictions) between an AI training on material and a human learning from it. Training on pirated material is a different matter, but training on material available for free online is difficult to call wrong.

I think your point about the social implications and centralisation of control is spot on, but the nuance that I think is important to get out here is the difference between frontier LLMs and self-hosted models. Some people will object to both, but the social argument you made only really works against the former. Same with the environmental impact.

For me, the most important aspect is the economic impact. The reduction in demand for artists and editors. And the damage done by people relying on bad AI-driven information—or worse, decisions. The output is worse art There are ways AI can be used to remove or at least minimise these concerns, but under a capitalistic system that's not going to happen.

That's how all my LLM projects are. Things that just wouldn't happen otherwise. Not that I can't, but just that I don't have the time.

It's not that my brain is tired or even that I trust LLMs, I just don't have time and either an LLM can create the thing I want or I do without. If the project fails, nothing of value is lost.

At home, that's the summary of so many of my hobby projects lately. At work, however, there's one phenomenon that I don't know if we have a name for yet: loss of job satisfaction because a large part of the joy of creative problem solving comes from the process of wrestling with a hard problem, making a breakthrough, and genuinely learning something. LLMs make it too easy. I don't feel like I'm learning anything of value. I'm just regurgitating some version of the LLM's solution back to my coworkers during stand-ups. And so, if anyone really pressed me on the details, I would have a hard time coming up with anything. For this latest release on a project that I've worked on for most of a year and given demos for in the past, we unleashed AI. I'm genuinely nervous to give an upcoming customer demo for this release because I don't really feel like I understand the changes were shipping.

At first, I thought it was just a problem I was experiencing alone. But then I started hearing others express very similar sentiments. At least one person mentioned that they chose to work on hobby projects without LLMs, precisely because they wanted to recapture that joy they lost from their career. Even though I've finally started knocking out quite a few ideas from the backlog that had accumulated over the years, it's hard to really reach that same level of satisfaction.

Is there a balance point where I can stay on top of the backlog while feeling like my brain hasn't fallen out of my head?

Always. There are tens of us though. Tens.