Zeke Hausfather

Clawed On Anthropic and the Department of War

Highly recommend this Dean Ball piece on the Anthropic / DOD imbroglio over the weekend:

Important point: fuel costs go up, inflation goes up, interest rates go up, RE deployment suffers

For more (and daily updates to the data), check out the Climate Dashboard: dashboard.theclimate...

For 2026 annual temperatures, we are currently on track to see the 4th warmest year on record at 1.47C above preindustrial levels, but uncertainty remains large enough that it could end up anywhere between the warmest and 5th warmest year.

February 2026 was the 5th warmest February on record, at 1.5C above preindustrial levels in the ERA5 dataset.

This is not that surprising given weak La Nina conditions at the start of the year; February tends to be one of the months most sensitive to ENSO.

An obituary for the short-lived DOE Climate Working Group, from Andrew Dessler over at The Climate Brink. It arose with a bang and died with a footnote in the EPA's endangerment finding repeal. www.theclimatebrink....

Definitely one of my favorite lyrics.

Hmm… www.nytimes.com/2026/03/01/t...

📉 Global Temperature Dashboard - The Climate Brink / @hausfath.bsky.social
2️⃣ bsky.app/profile/haus... 
Explore our article: climate.copernicus.eu/c3s-data-pow...   

⬇️

Statement from Dario Amodei on our discussions with the Department of War A statement from our CEO on national security uses of AI

Well well well. Be very interested to hear more from folk who are in the loop with this.
www.anthropic.com/news/stateme...

We need a bigger scale!

I do think there is a Jervons paradox risk here, where AI work becomes so cheap (compared to humans) that the amount of work being done dramatically increases. Its less us using it as a tool to enhance scientific output (in my case) and more swarms of autonomous agents doing tasks.

Mathematics in the Library of Babel — Daniel Litt Mathematics isn't only about saying true things. It's about asking the right questions, being confused, stumbling about, getting distracted, being wrong, recognizing when you're wrong, being stuck. Mo...

Yeh, I was considering adding a section on its performance in mathematical proofs, but my knowledge of that is mainly second hand. This was a really good discussion of what it can (and cannot) do there: www.daniellitt.com/blog/2026/2/...

AI has an impact, and we need policy to ensure that it is primarily powered by clean energy. And its still helpful to contextualize it relative to other things that we do in our lives.

All reasonable points, though I'd suggest that the median day with Claude Code here is a bit of an outlier (as it represents using the AI tool non-stop all day with many agents running), and even in that case uses less energy than their refrigerator.

Thats a good analogue. One thing I want to work on is improving the planning process to get better results; I was reading some folks at Anthropic who recommended having it ask a lot of questions when developing the plan rather than making assumptions solely based on the initial prompt.

How well it performs will depend on the problem, but I've found that it has gotten much better in recent models (and I myself am not always very good at this!).

Its usually a process of both reviewing the code it writes and looking through the resulting merged dataset for errors or remaining issues. Though I've increasingly started working with Claude Code to develop tests prior to running the analysis to catch potential issues.

GitHub - hausfath/SimMod: Python Simple Climate Model Python Simple Climate Model. Contribute to hausfath/SimMod development by creating an account on GitHub.

I remember back in grad school spending weeks creating my own version of a simple climate model you and Nathan Myhrvold developed: github.com/hausfath/Sim...

It would be trivial to do with AI today, but you'd miss any understanding of "how" that (for me at least) was the most important outcome.

I agree, the impacts that this will have on the broader workforce (and early career folks) is really worrying. And I'm not comforted by vague platitudes from AI executives about future universal income / post-scarcity socialism. The more likely outcome is ever growing inequality...

Agreed; I worry that over reliance on these tools will lead folks to skip learning why and how things actually work. Thats the knowledge thats essential to determine if the AI tool is actually doing the analysis you want (vs its oft-incorrect interpretation of what you want).

Frugal compared to some people I know 😉

The AI-Augmented Scientist The promise and pitfalls of using AI tools to boost my capabilities as a scientist

For more, including a discussion of AI energy and water use impacts and an applied example of where AI worked (and didn't work) on a real science problem, check out the full post over at The Climate Brink:

The context that a specialist in a field has about what is important and how it fits into the broader field’s needs is one of the most important reasons that, at least in science, humans will remain the key driver of scientific innovations for the foreseeable future.

AI has a bit of a conservative bias (small-c, not in the political sense) where it will tend to go with prevailing conventions represented in its training data and discount newer studies. And it will generally not do a good job of coming up with original research ideas.

Exercises like the writing an IPCC report or a thorough review of a climate tipping points requires both familiarity with the literature and the ability to make expert judgements on how to weight the credibility and importance of different sources, something I find lacking in AI-based assessments.

Similarly, AI still has some shortcomings when used for synthesis or assessment (rather than a simple review) of scientific findings.

But these tools lack full access to the peer-reviewed academic literature (much of which remains behind journal paywalls), and often will provide a less thorough (albeit still useful) assessment of topics that I personally know well.

Research using AI is in my experience good but not great. I find tools like Deep Research in Gemini helpful to provide a good overview for topics where I personally do not have deep expertise.

Despite trying lots of difference experiments over the years – trying to have the AI analyze past writings to learn my style – I still find AI writing a poor simulacra of my own. There is a certain style of AI writing that is distinct and a bit soulless.