A pause (for now) on AI and the environment posts, and a bounty for mistakes

A retrospective

Over the last half year or so I’ve been running through a lot of deep dives on AI and the environment. I was motivated by a few key points:

• I was going completely crazy with the number of everyday people I was meeting who suddenly all had wildly inaccurate confident beliefs about AI, energy, and water. All of these beliefs were easily disprovable with simple easy-to-find statistics and comparisons, but I wasn’t finding anyone online doing it. Every piece of news coverage I was reading had ridiculous comparisons (the first one I read literally said that ChatGPT is now using more than twice as much energy as a whole person) and no one doing the simple David MacKay move of actually putting the numbers in context.

• I had a decent background in the general facts involved after following climate change for 15 years and teaching physics for 7 (if you’ve enjoyed my explanations here you can find my full explanation of all high school physics on my YouTube channel) and it was satisfying to use my stored-up knowledge. It was nice to get out some really fundamental ways I think about climate to a big audience.

• I was getting a lot of great feedback. My original two posts (here and here) have been collectively read 230,000 times, and posting the first was the reason my blog took off:

  

  It’s put me in touch with a lot of really cool people in tech and journalism and environmentalism, and has been extremely fun. I’ve developed a reputation as the AI water guy.

  

• This was one of the first topics I was able to do very rapid, complex research using chatbots as aids. A lot of people who don’t use them much don’t know how easy it’s become to just have them make huge catalogs of relevant sources and to double check all the sources they give. Here’s one of many examples:

  

  I feel like I’m writing with a whole team of researchers by my side now. I owe a lot of this blog’s success to AI chatbots. At some point I’ll write an update to how I use them.

Now I’ve developed a pretty huge catalog of takes on AI and environmentalism more broadly. I always wanted a topic where I could imitate Piero Scaruffi and leave a big collection of takes important to me people could explore for themselves, and now I have that.

I’ve run up against a big wall of reasons to take a break from the subject:

• Everyday people are converging on the correct objections. Much more than before, I’ve noticed a lot more people saying “AI water use is mostly fake, personal prompt costs are mostly fake, but AI as a whole is going to put a huge strain on our electrical grid going forward that’s going to matter a lot for the green energy transition, local electricity prices, and air pollution.” To be clear I don’t think I caused that, it’s just that the conversation has found its footing more than it had a year ago.

• I’m working on AI and the environment in other places. I have a few pieces forthcoming in more established places than this blog and want to give them my full “AI and the environment” energy right now.

• I’ve said most of what I want to say about it. Some of the new objections to AI data centers are getting weirder and weirder. Some people are now saying the land use is an issue, which is so wild that I don’t even think it deserves a post on its own (they’re so incredibly compact relative to what they do, and such tiny tiny parts of America’s full land footprint. We have a lot of land!). I’ve addressed what I think I can address, and other stuff (like household electricity prices) are complex enough that I feel like I can’t do them justice with the limited time I have right now. I’ve really really really been thorough in addressing every last point. Every now and then some new big article will come out with some massive incorrect claim about AI and the environment, and every single time I feel like my past blog posts have already addressed all the goofy moves the author is making.

  

  I can have a hit tweet on the topic every month or so when a new article comes out, that takes me ~10 minutes to research and type, and reach a pretty large audience. This leads into the next point.

• I’ve gotten most of the “explore” out of the way, what’s left to do is “exploit.” In the explore-exploit trade-off, I feel like I’ve covered a lot of the ground and learned where most of the contours of the debate now. What’s most impactful to do now is mash a button on the most important points.

  

  This is me saying “It doesn’t use that much water tho” over and over again. Because I feel like I’ve found the most important points, I’ve mostly just been saying the same few simple things (listed here) over and over in response to different specific arguments. This can get old, and is starting to not feel valuable.

• My impact on this topic could be bigger elsewhere. I think if I wanted to actually maximally change the conversation on this, I should make some kind of YouTube overview. I don’t have the capacity for that right now, but if I’m being honest and wanted to build the biggest audience and change the most minds, that’s where I’d do it. If you’d like to run with a lot of my points on YouTube, I think there’s a ton of low hanging fruit in the debate there.

• I’m burned out on this topic. Back when I was teaching physics, the word “physics” had become painful for me to hear, because I’d hear it so often every single day. Now typing the word “data center” is equally painful. I kind of just need a break.

• I’d like more people to be doing this. I really think what I’m doing here is straightforward and a lot of people could be doing the same. I’m flattered to get people messaging me about this.

  

  But I’d also like more people to run with these basic moves I keep making over and over and apply them in new places! I’ve gotten some nice messages like this:

  

  I shed a single tear

  and I’d like more people to do the same. Almost no one anywhere is actually writing about this thoroughly for a broad audience. It’s either dense academic papers of mostly correct takes, or wildly off coverage from people who seem to mostly be following whatever others have already written.

• I should prioritize more important stuff. I am into and a representative of effective altruism, a key tenant of which is that most of your impact in life comes from what you choose to prioritize, and most people very quickly get obsessed with a very local maxima of impact they happen to discover and lose the opportunity to do a lot more good elsewhere. Posting about this has been a fun hobby for me, but I think I need to orient my hobbies in more impactful directions. I’d like to focus more on:

  • Misuse risks from advanced AI systems.

  • How seriously to take the basic x-risk case from advanced AI (my “p(doom)” has always been low but nonzero, and I need to see how justified that is).

  • Animal welfare and ethics.

  • The general DC AI policy space.

  • AI and China, and China in general.

  • General lifestyle posting (I think this short post is one of the best things I’ve written and I’d like to write more like it).

  • Ideas from philosophy that have been important to me.

• I need to give more time to big work projects over the next few months. In the words of one of my closest friends:

  

I’m willing to bet that you haven’t read every last post I’ve written on this topic. If you’d like more posts from me on this, first circle back and read the other ones. Maybe upload them all into an AI model and have them give an AI take from my perspective on a new issue you’re trying to understand. Or best of all, just start doing these deep dives yourself! I’m mostly using high school-level math and physics here.

Thank you to my legions of new followers. I’m so excited to be participating in the debates with all of you about this amazing, wild, sometimes terrifying new technology and its implications for society and the future. It’s crazy to have such a big audience from all over and I’m excited to keep this adventure going. If you ever want to chat, your bar for reaching out to me at AndyMasley@gmail.com should be low.

As an aside, if you know anyone in Nevada or South Dakota…

A bounty for mistakes

In writing this blog, I have also developed many critics. While some have been really useful, others make what I think are wildly false claims about what I’m doing here. Take this one:

All my claims are based on external trustworthy sources I link, and if I link myself it’s only to show an argument I’ve made somewhere else using external sources.

I stand behind everything I’ve posted on this blog. To demonstrate, I offer a simple bounty: I will send you $50 if you can find:

• A legitimate source for a statistic I cite where I lie about the number it’s giving or I get it wrong, which invalidates a core argument I’m making.

• A source for a number I cite that’s illegitimate, and that most legitimate sources disagree with, and this updated information invalidates a core argument I’m making.

If I did a calculation wrong and 0.0000102% should actually be 0.0000103%, I’m not going to consider that to be invalidating. But if I say the median chatbot prompt uses 0.3 Wh, and it actually uses 300 Wh, that does invalidate a core claim I make. A few more rules:

• All claims need to be analyzed by when I said them. If I make a claim about ChatGPT in June, and Sam Altman comes out with a new model in December that’s 100x as energy intensive, my claim in June isn’t wrong, just outdated.

• I’m vetoing very specific disagreements of what counts as “using” water. Whether to count rainfall on crops or into lakes dammed by hydroelectric power stations turns out to be very relevant for the water debates. The way both are reported leave huge error bars in both directions and I’ve tried my best to take a middle ground on them. I’d be interested if you disagree with my takes, but I worry that both extremes have such reasonable people on them that I could go back and forth between sending both $50 as I oscillate between two extremes.

• I have a few pieces I typed fast (like this one) where I haven’t circled around to cite some of the stats in them, because I had gone over them in previous posts. I ask for a reprieve if I don’t cite a source in a newer post but I had cited the source for it previously, and the source is legit.

• I used some rough back of the envelope math for the graphs in my very first post that I want to revisit, so I’m walling those off specifically. Sorry fam.

If you find something, email me at AndyMasley@gmail.com. I’ll be happy to have a back and forth with you if I disagree with your take and why. You can feel free to publicly shame me if I’m evasive. If you’re right, I’ll send you $50 and post the update on a wall of shame. I might be a little slow to get to you, ping me if I don’t respond the first time, publicly shame me if you don’t hear back in a few weeks.