There is an important distinction between using AI to create content (writing, images, etc.) and using AI as part of a scientific workflow (e.g. to speed up coding and data analysis). I definitely share some of your concerns around the former.
I'm with you on everything but "4) AI slop is soulless content."
AI will keep improving at generating "clever" and/or engaging content. All it needs is honesty, sincerity, authenticity. Once it can fake that, it's got it made.
"To put this in perspective, showering for 10 minutes (which is rather frugal)...."
----
Ah, someone who hasn't lived in a water-restricted area.
I remember the days when I had the water running continuously the whole time I was in the shower. Now I take "space station" showers (wet down, turn off water, scrub everything, then rinse off).
I am in this boat. I wanted to be a scientist growing up. I took engineering, computer science classes and eventually got a degree in Earth and Space science education. I am not a software developer but I have played around with vibe coding and now the possibility of me actually creating something that contributes to my mission of Environmental Sustainability Education seems very doable. I have looked into ArcGIS and Ersi and taught myself the basics of Python from my understanding of C++ and Visual Basic. I am teaching an intro JAVA programming class now and AI makes me feel like I can build something bigger than what runs through the compiler and console. I think technology and AI are great TOOLS but that cannot take the place of the problem finder or sharing the findings to create change. You also mentioned that the data centers take a lot of resources, to me that make the sustainability issues more pronounced. We need alternative renewable energy sources powering these new technologies. We must balance the advancements we have sustainably so we do not cause a deficit in the distribution of energy resources. We have so many on one extreme or the other. We need more people that want to work on the integration and balance of technology advancement and environmental sustainability. Agricultural technology is a great example how automation increases efficiency yet most farmers care about their land as it is their livelihood.
"I think technology and AI are great TOOLS but...."
----
As an Old Person, part of my job is to complain about how fast things are changing, but now I find myself warning people that the rate of change may be beyond what even thirty-somethings can handle: geopolitics (China is racing ahead), discoveries of climate science, the climate ·changes· themselves, AI capabilities, cars driving better than people, drone warfare, etc.
(And that's not even counting the insane rate of damage to the US in one year, with the collapse of the Constitutional Order and systems decades in the making being wiped out overnight.)
I am not speaking solely on the US. I am speaking globally. We are much more interdependent on other countries after the internet made rapid communication of large data possible over long distances. Environmental Sustainability is a World issue that needs equity to progress. Unless a new worldview of environmental policy happens, we will continue to deplete resources without addressing the resource requirements of the future. The decisions of a few countries that contribute to increased pollution rather than reducing environmental impacts may alter the future for all. The world needs to wake up and realize we only get one Earth to take care of. The talk of space colonies is still science fiction as they have not reached a self sufficient sustainable design. Therefore we cannot just relocate and restart. We destroy the planet and it is Game Over for humanity.
Glad you are on the AI train Dr Hausfather! It looks like Clarke's magic. The only caveat we programmers have in all s/w is beware of GIGO. Good Data is the golden juice.
As a recovering software engineer, I'd say that Good Specifications have to come first. An economist's program that calculates the impact of climate change on GDP better not assume that it won't affect people who work indoor jobs.
[Disclosure: this comment is wholly organic. No AI assistance whatever. -MA]
Thanks Zeke! I emailed your post to a retired scientist friend, who's been "skeptical" (i.e. negative) about the social cost of LLMs. I was too, until I started experimenting with them. I've gotten used to the hallucinations, and always check the human-generated sources it links. As you say, it's clearly improved in the last couple of years:
"Today the tools are much better than they were in 2023. Hallucinations still exist, but they are much less frequent. As someone who has used these tools more than most in the scientific community, I have a good sense of what they work for and what they do no do well today. The tools I primarily use now are Claude Code (Opus 4.6, via my terminal) and the web-app for Gemini (3.1) for projects where integration with my email, Drive, and other parts of the Google ecosystem is helpful."
I'm a retired systems administrator, who did a lot of ad-hoc coding in Perl but never got past procedural Python. I haven't coded anything in eight years, thankfully. I'm still not willing to pay more for *correct* facts, but my experience with Gemini 3 'fast' (i.e. "free") is otherwise similar to yours.
I've been using the free (i.e. no extra charge on top of my other Google services and my user eyeballs) Gemini 3 web app for advice on writing substack comments! I try to be scrupulous about disclosing how much of my comment is AI-assisted, however. I like to draft a reply to a decarbonization obstructionist's comment, then ask Gemini to evaluate both the comment and my reply for motivated cognition.
Heh. Even when I issue the 'less sycophantic' prompt, it tells me I'm a freakin' genius! It's my new best friend 8^D! This way lies madness, or at least Narcissistic Personality Disorder 8^(.
In my experience Claude is a lot less sycophantic than Gemini or ChatGPT and is more willing to call a bad idea. But its still a problem for models in general...
It should be pretty solvable if AI companies pay for access to scientific journals (or offer a specialized product to universities / researchers that comes with that access). The current approach of scraping all the publicly available versions (or possibly training with not very legal and out of date archives from scihub) is far from ideal.
"AI tools have gotten quite good at cleaning, merging, and analyzing large datasets"
Have you found that their accuracy in this is good enough (aka perfect)? I've tested this over the past few years, but not recently, and have always found that it introduces an enormous amount of hallucinations as the data sets grow in size. Are you having the LLMs output the cleaned data, or are you having the LLMs write scripts to clean it that you then manually execute and check?
I'm having LLMs write the scripts and associated tests and check their outputs. In my experience it works well when the problem is addressable by code. Less so when I try and have the LLM itself look through the entries and determine a pattern or classification.
I'd suggest trying with a recent model like Opus 4.6 and seeing how well it does, as things really are improving quickly here.
Yes. Its not always perfect on the first shot (and neither are my attempts), but its much faster to iterate. Having good tests up front is also helpful, and Claude can also assist in looking through the output to find potentially problematic things.
Thank you, Zeke, very insightful. I just came across this study on Linkedin… I see that you use it mainly for coding and quantitative analysis, but any thoughts based on your experience?
GPT5.2 tells me that BECCS is the most cost effective, proven carbon capture, and a $200 trillion investment can be expected to reduce warming by .3 degrees C (assuming the forests are well managed and keep reproducing). Drax's $2 million a day subsidy would seem to support this sort of price range. Has anybody else asked a similar question of your AI, and what answer did you get?
I am also not entirely happy using Chat GPT (deep research) for example for literature review. Maybe I just prompted it badly (who knows) but it overlooked so, so many (important!) papers. And this was on a very specific question where I already knew the literature well.
But coding, damn. Helped me a lot in my bachelor’s!
My experience with GPT5.2 is that if I ask it a complicated, open ended question, it can answer almost anything. Its no good at all in guessing what I think is important. If I ask it what it thinks about a certain paper, it will analyze the paper beautifully. If I ask about another related paper, and then a 3rd, it figures out what I'm after, and will even search the web for other papers relevent to my argument.
The french "Shift Project" (a think-tank of engineers led by Jean-Marc Jancovici, not exactly a bunch of left-wing eco-terrorists) published their report on AI last October. They focus about the situation in Europe (and France, of course). The picture is not as rosy as in this substack post (maybe because in Europe a 10mile round trip to the office would most often happen via bicycle or public transport?). Link to the english translations here (full report and executive summary): https://theshiftproject.org/en/publications/al-data-and-computing-shaping-infrastructures-for-a-decarbonised-world/
They can be important locally but the amount of water involved is pretty tiny compared to, say, agriculture. As I say in the footnote, one 10-minute shower is equivalent to ~300,000 AI queries.
Entire sympathy for the complexity of learning matplotlib
As a layperson I have pledged to never knowingly use AI, for a variety of reasons:
1)Garbage in, garbage out. AI is only as good as the info it has scraped.
2)The power and water usage issues are deeply concerning.
3)The copyright issues are massive. Writers and artists are being ripped off.
4)AI slop is soulless content.
5)Tech bros are immoral oligarchs.
There is an important distinction between using AI to create content (writing, images, etc.) and using AI as part of a scientific workflow (e.g. to speed up coding and data analysis). I definitely share some of your concerns around the former.
I'm with you on everything but "4) AI slop is soulless content."
AI will keep improving at generating "clever" and/or engaging content. All it needs is honesty, sincerity, authenticity. Once it can fake that, it's got it made.
AI is a great power.. with it comes great responsibility
We need power to evolve well
The implication of your rule number 1 is that these results that zeke has produced with ai are garbage. I disagree.
"To put this in perspective, showering for 10 minutes (which is rather frugal)...."
----
Ah, someone who hasn't lived in a water-restricted area.
I remember the days when I had the water running continuously the whole time I was in the shower. Now I take "space station" showers (wet down, turn off water, scrub everything, then rinse off).
Am pretty sure we will have to bring the water from the oceans to us through canals/pipes with desal along the points of passage.
A good AI prompt might come go this conclusion. I wonder if "intuition" is really hyper-advanced AI
I am in this boat. I wanted to be a scientist growing up. I took engineering, computer science classes and eventually got a degree in Earth and Space science education. I am not a software developer but I have played around with vibe coding and now the possibility of me actually creating something that contributes to my mission of Environmental Sustainability Education seems very doable. I have looked into ArcGIS and Ersi and taught myself the basics of Python from my understanding of C++ and Visual Basic. I am teaching an intro JAVA programming class now and AI makes me feel like I can build something bigger than what runs through the compiler and console. I think technology and AI are great TOOLS but that cannot take the place of the problem finder or sharing the findings to create change. You also mentioned that the data centers take a lot of resources, to me that make the sustainability issues more pronounced. We need alternative renewable energy sources powering these new technologies. We must balance the advancements we have sustainably so we do not cause a deficit in the distribution of energy resources. We have so many on one extreme or the other. We need more people that want to work on the integration and balance of technology advancement and environmental sustainability. Agricultural technology is a great example how automation increases efficiency yet most farmers care about their land as it is their livelihood.
"I think technology and AI are great TOOLS but...."
----
As an Old Person, part of my job is to complain about how fast things are changing, but now I find myself warning people that the rate of change may be beyond what even thirty-somethings can handle: geopolitics (China is racing ahead), discoveries of climate science, the climate ·changes· themselves, AI capabilities, cars driving better than people, drone warfare, etc.
(And that's not even counting the insane rate of damage to the US in one year, with the collapse of the Constitutional Order and systems decades in the making being wiped out overnight.)
Barlock, Great points. I have similar concerns..
Use what you say as an AI prompt.
I am not speaking solely on the US. I am speaking globally. We are much more interdependent on other countries after the internet made rapid communication of large data possible over long distances. Environmental Sustainability is a World issue that needs equity to progress. Unless a new worldview of environmental policy happens, we will continue to deplete resources without addressing the resource requirements of the future. The decisions of a few countries that contribute to increased pollution rather than reducing environmental impacts may alter the future for all. The world needs to wake up and realize we only get one Earth to take care of. The talk of space colonies is still science fiction as they have not reached a self sufficient sustainable design. Therefore we cannot just relocate and restart. We destroy the planet and it is Game Over for humanity.
Glad you are on the AI train Dr Hausfather! It looks like Clarke's magic. The only caveat we programmers have in all s/w is beware of GIGO. Good Data is the golden juice.
And Good Questions are the Golden throat. Deep Throat?
"42"
Hitchhiker's and Watchmen were way ahead of the times. It's the Outta Times now...
Or "Ommmmm"
"Rosebuds"
As a recovering software engineer, I'd say that Good Specifications have to come first. An economist's program that calculates the impact of climate change on GDP better not assume that it won't affect people who work indoor jobs.
Where have all the good system engineers gone?
Where are the true requirements?
[Disclosure: this comment is wholly organic. No AI assistance whatever. -MA]
Thanks Zeke! I emailed your post to a retired scientist friend, who's been "skeptical" (i.e. negative) about the social cost of LLMs. I was too, until I started experimenting with them. I've gotten used to the hallucinations, and always check the human-generated sources it links. As you say, it's clearly improved in the last couple of years:
"Today the tools are much better than they were in 2023. Hallucinations still exist, but they are much less frequent. As someone who has used these tools more than most in the scientific community, I have a good sense of what they work for and what they do no do well today. The tools I primarily use now are Claude Code (Opus 4.6, via my terminal) and the web-app for Gemini (3.1) for projects where integration with my email, Drive, and other parts of the Google ecosystem is helpful."
I'm a retired systems administrator, who did a lot of ad-hoc coding in Perl but never got past procedural Python. I haven't coded anything in eight years, thankfully. I'm still not willing to pay more for *correct* facts, but my experience with Gemini 3 'fast' (i.e. "free") is otherwise similar to yours.
I've been using the free (i.e. no extra charge on top of my other Google services and my user eyeballs) Gemini 3 web app for advice on writing substack comments! I try to be scrupulous about disclosing how much of my comment is AI-assisted, however. I like to draft a reply to a decarbonization obstructionist's comment, then ask Gemini to evaluate both the comment and my reply for motivated cognition.
Heh. Even when I issue the 'less sycophantic' prompt, it tells me I'm a freakin' genius! It's my new best friend 8^D! This way lies madness, or at least Narcissistic Personality Disorder 8^(.
In my experience Claude is a lot less sycophantic than Gemini or ChatGPT and is more willing to call a bad idea. But its still a problem for models in general...
Find . -print!cpio -pdumv /wherever/overthere
AI has gotten freakingly humanish. Goal for me is surface albedo smarts. Really cool and super useful!
Am used to old canned UNIX stuff when that was the old HAL now. Dang, old generals don't die.. they just fade away..
Nice!
You write in regards to Deep Gemini:
"But these tools lack full access to the peer-reviewed academic literature (much of which remains behind journal paywalls)."
How much of a problem is this and how fixable is it? I imagine these journals are going to want money before they let you train on the good stuff.
It should be pretty solvable if AI companies pay for access to scientific journals (or offer a specialized product to universities / researchers that comes with that access). The current approach of scraping all the publicly available versions (or possibly training with not very legal and out of date archives from scihub) is far from ideal.
"AI tools have gotten quite good at cleaning, merging, and analyzing large datasets"
Have you found that their accuracy in this is good enough (aka perfect)? I've tested this over the past few years, but not recently, and have always found that it introduces an enormous amount of hallucinations as the data sets grow in size. Are you having the LLMs output the cleaned data, or are you having the LLMs write scripts to clean it that you then manually execute and check?
I'm having LLMs write the scripts and associated tests and check their outputs. In my experience it works well when the problem is addressable by code. Less so when I try and have the LLM itself look through the entries and determine a pattern or classification.
I'd suggest trying with a recent model like Opus 4.6 and seeing how well it does, as things really are improving quickly here.
...but, you have checked (or spot checked) the results yourself, for large data sets, right?
Yes. Its not always perfect on the first shot (and neither are my attempts), but its much faster to iterate. Having good tests up front is also helpful, and Claude can also assist in looking through the output to find potentially problematic things.
Thank you, Zeke, very insightful. I just came across this study on Linkedin… I see that you use it mainly for coding and quantitative analysis, but any thoughts based on your experience?
Study: https://www.sciencedirect.com/science/article/pii/S2590291125010186
GPT5.2 tells me that BECCS is the most cost effective, proven carbon capture, and a $200 trillion investment can be expected to reduce warming by .3 degrees C (assuming the forests are well managed and keep reproducing). Drax's $2 million a day subsidy would seem to support this sort of price range. Has anybody else asked a similar question of your AI, and what answer did you get?
Excellent piece very consistent with my experience
I am also not entirely happy using Chat GPT (deep research) for example for literature review. Maybe I just prompted it badly (who knows) but it overlooked so, so many (important!) papers. And this was on a very specific question where I already knew the literature well.
But coding, damn. Helped me a lot in my bachelor’s!
My experience with GPT5.2 is that if I ask it a complicated, open ended question, it can answer almost anything. Its no good at all in guessing what I think is important. If I ask it what it thinks about a certain paper, it will analyze the paper beautifully. If I ask about another related paper, and then a 3rd, it figures out what I'm after, and will even search the web for other papers relevent to my argument.
I cannot trust ai. It is crated by capitalists exploiters to lie and profit from their lies.
The french "Shift Project" (a think-tank of engineers led by Jean-Marc Jancovici, not exactly a bunch of left-wing eco-terrorists) published their report on AI last October. They focus about the situation in Europe (and France, of course). The picture is not as rosy as in this substack post (maybe because in Europe a 10mile round trip to the office would most often happen via bicycle or public transport?). Link to the english translations here (full report and executive summary): https://theshiftproject.org/en/publications/al-data-and-computing-shaping-infrastructures-for-a-decarbonised-world/
What’s your take on the water usage concerns?
They can be important locally but the amount of water involved is pretty tiny compared to, say, agriculture. As I say in the footnote, one 10-minute shower is equivalent to ~300,000 AI queries.
I appreciate that. I get asked a lot, so if you have any academic journals you recommend, I'd love to have them on hand!
These "current practical application" discussions don't answer my big questions.
- Does an 8-year-old need to learn math?
- What should a guidance counselor tell high schoolers?
- Do we need to worry about AIs being worse to humans than humans are to each other?