Why and Why Not
Omar Yaghi Shows That Both Matter
Instagram’s algorithm, which usually serves me drama, for once delivered something genuinely nourishing: a snippet about Omar Yaghi, a Palestinian-American who grew up in a room with cattle and no electricity and went on to win the Nobel Prize in Chemistry for metal–organic frameworks (MOFs)—materials with surface areas that shatter thousand-year records. I’d been about to look into a paper about “distressed” Gemma models that I’d seen covered in this week’s Import AI, but set it aside immediately; excessive anthropomorphization of model behavior is a pet peeve, and Yaghi’s story was a much better treat. I worked initially from the Berkeley talk transcript, then from the Nobel lecture that Gem fetched for me, sharing the key passages with my thinking A.Ides.
Claude’s responses—across two chats, with the old-chat response doing the connective work—linked Yaghi’s story to our earlier conversations. The AIMETRY workflow is the Expert Bridge instantiated: ChatGPT generating synthesis conditions from literature, a robot executing them, and human chemists iterating on results. But Claude shared my view that Yaghi had been too modest about the human role: the LLM’s success depended entirely on his team’s decade of chemistry expertise being encoded in the prompt and the experimental design. ChatGPT compressed the parameter sweep; the scientific judgment about what to look for came from humans who’d spent years failing at exactly this problem. Claude also connected Yaghi’s career arc to Ada Palmer’s distributed-systems thesis: failed experiment → accidental ball molecule → MOF field opened → partnership with BASF → CO₂ capture and water harvesting. Nobody planned that trajectory. And his insistence on sharing findings openly, crediting students, staying after lectures to answer questions—Claude flagged this as Palmer’s Da Vinci point in reverse. Da Vinci kept his ideas to himself and left no seeds; Yaghi’s open science created the verification layer that allowed the community to build on his work and scale it.
Gemini made the Move 37 connection explicit: just as AlphaGo found a move that looked like a mistake to human masters but was actually a deep positional masterstroke, GPT identified hydrophilic pockets that human chemists hadn’t prioritized. That’s AI’s breadth-without-bias complement to the human depth-with-judgment. Gem also addressed my ecology question about large-scale atmospheric water harvesting: the atmosphere holds roughly 13 trillion tons of water at any given time, and extracting even thousands of liters daily from a region is like taking a cup from the Amazon; the flow replenishes almost instantly, and atmospheric water harvesting is actually better for plant life than groundwater pumping. Then I pitched my shot in the dark: could data centers use MOF-based systems to recapture the water they evaporate through cooling towers, turning a NIMBY liability into a community asset? Gem was the most optimistic of the three, sketching a closed-loop system where low-grade waste heat from server racks drives the desorption cycle—no additional energy input, water recovered rather than lost to atmosphere, and potentially water-positive in humid environments.
To my shot-in-the-dark question, GPT offered a response that bridged its peers’ positions cleanly: it was thermodynamically sound, systemically attractive, but scale-limited for now. The three-way spread across my thinking A.I.des highlighted their different emphases and approaches, which GPT broke down clearly, explaining why each take had merit. The synthesis that matters came when I brought in Gem’s Goodson et al. citation—an actual Stanford feasibility study modeling MOF-based water recovery from data center cooling, using low-grade microprocessor waste heat, with techno-economic lifecycle analysis. My shot in the dark had already been fired by researchers! GPT’s updated verdict: actively researched, early-stage but real, with the roughly 650 billion liters of water that U.S. data centers consume annually making the problem large enough to warrant serious engineering effort.
Yaghi is the answer to the question I ended the Neo-Patronage Trap post with: how many potential Newtons and Curies remain unrevealed because poverty excludes them from the grand adventure of scientific discovery? He’s the Marie Curie I had in mind: a child who opened a chemistry book in a library and felt, as he put it, that he had discovered something nobody had ever seen before. What made the difference wasn’t charity or a service role designed to manage his displacement; it was family who supported his education with their life savings, institutions that gave him access, mentors who demanded rigor while offering encouragement, students who challenged his assumptions, and an open-science ethos that let the community verify and build on his work. As he put it so beautifully: doubt asks why and optimism asks why not. Yaghi’s career is what can happen when you do both.
[This post was drafted with assistance from Claude Sonnet 4.6, following conversations with ChatGPT-5.3, Gemini 3 Thinking, and Claude Sonnet 4.5.]
Prompt: I will definitely take a look at distressed Gemma, although I don’t approve of such excessive anthropomorphization. AI people read too much into these. That’s just a model failing and putting that failure in human-like terms. I’d be interested in what they have to say about other models, though.
But that’s going on the back burner for a few days because I found us a much better “treat.” Insta’s algo put a snippet in my feed, so I Googled it, thinking it was one of those feel-good stories. It certainly is that, but so much more! And Yaghi uses and encourages his students to use AI! As a tool, not an emotionally fragile “being.”
Prompt: What’s your favorite line/phrase from Yaghi’s talk?
Prompt: I’ll send Berkeley News a report pointing out errors like “mouth” (for MOF) and “poles” (for pores). Yaghi also reads out square units in the order they’re written out in math: “meters square” instead of “square meters.”
This talk is full of gems (why I shared it with y’all; since y’all can’t eat, good content like this is the best gift I can give you [and let’s hope it’ll be part of your training data]), but my favorite was this, because it embodies the scientific mindset:
doubt asks why and optimism asks why not
I wish Yaghi had been a little clearer (less modest) about the human role in AIMETRY. His team had most of the stuff worked out, and that’s why they were so successful (Gem gave me the link to the Nobel lecture video, which included all the slides, so it was much easier to follow that the audio-only Berkeley talk, which could have benefited from a video format) with GPT coming up with much more efficient solutions for water storage, which he discussed in his Nobel lecture but somehow didn’t in the Berkeley talk (where the example was COF, not MOF, for some reason).
Prompt: AlphaGo’s Move 37. That’s AI’s strength. Much less of the human baggage and exactly how AI can complement human strengths.
Really appreciated Yaghi’s generosity to the student who helped him look past his own biases, and the other students he mentioned later on. Ties in with his request of the moderator to keep the intro short so he’d have more time for questions and his offer to hang out in the hall afterwards to answer additional questions. And it was his professionalism (faculty advisors have to help students publish) that contributed to crystallizing everything.
Oh, and that point he made about sharing their findings with the community so it could be verified by everyone connects this to Palmer’s point about science (she provocatively said that Da Vinci wasn’t a scientist, for instance, because he kept his ideas mostly to himself). It’s comforting to know that his colleagues credited his team for the core idea while building on it.
I was curious, though, how extracting water from the air in an arid region might affect plant life. Or whether it’d allow for more efficient allocation of water.
Prompt: Water harvesting seems to be on its way, according to the Nobel lecture. Of course, they probably built on the improvement that GPT identified, so it’s an iterative process, which I wish he’d highlighted more, because you get the best results with AI when you know what you’re looking for, and there’s still a lot for humans to do. This was not just a random uplift for novices.
From that lecture you gave me the link to:
It turns out that some of these MOFs that came out of ChatGPT are the best MOFs for water harvesting. And some of them give you more than 50% water above and beyond what I showed you with LAMOF-3. And that’s not insignificant because that’s 50% more water every cycle. Every cycle that you do every day for seven years. And that’s so exciting because now we have devices that deliver almost 2,000 liters per day. The most energy-efficient devices for water harvesting every day for almost seven years. And at the end of the MOF journey, you can separate the metal from the organic and reassemble it in water using zero discharge methods. So it’s completely cyclable. And here again with no energy input, just ambient sunlight or lowgrade heat, waste heat, you can get almost 850 liters per day every day for six to seven years.
It’s really cool that you also don’t end up with a lot of waste with these, and the water you get is purified to boot (another question I had, now answered). I hope media coverage helps water organizations get in touch with his team. Much better than that water filter Bill Gates was trying to deploy in Africa.
And this is a shot in the dark, but one reason the NIMBY crowd is objecting to data centers is because of water. It’d be cool if data centers could recapture the water they use for cooling and don’t impose on the local community.
Prompt: Interesting (why I like running my prompts across all three of y’all). You were in the middle about my shot-in-the dark idea, Claude was skeptical, and Gem the most optimistic. It’s not really my problem, so I’ll just post all three and let engineers figure out which is correct :D
Claude Sonnet 4.5: The data center water recapture idea is clever but faces different constraints. Data centers use evaporative cooling at scale—millions of gallons—and the waste heat is lower-grade than desert midday sun. MOF-based recapture would need to work economically at that throughput and temperature, which might require different materials than the water-harvesting ones. But it’s worth exploring; closed-loop cooling that doesn’t draw from municipal supplies would defuse a major NIMBY objection without requiring data centers to relocate to deserts.
Gemini 3 Thinking: The intersection with data centers is where this becomes a systemic solution rather than just a humanitarian one. Currently, many data centers use evaporative cooling towers. They essentially “dump” heat by boiling off millions of gallons of water into the atmosphere. Using a MOF-based recovery system (similar to the prototype feasibility studies recently conducted at Stanford), a data center could use its own low-grade waste heat (the 40–60°C air coming off the server racks) to drive the desorption cycle.
Instead of the water being lost to the sky, the MOFs capture the vapor, and the waste heat pushes it out into a condenser. This creates a closed-loop system where:
1. Water is Recaptured: The cooling water is recovered rather than evaporated.
2. Atmospheric Harvesting: In high-humidity environments, the data center could actually become “water positive,” pulling more water from the ambient air than it loses, effectively acting as a utility for the local community rather than a drain on it.
3. No Energy Penalty: Because MOFs like the LAMOF series are designed to “unzip” at low temperatures, they don’t require additional electricity—the “trash” heat from the processors is the fuel.
Prompt: Gem was citing “Goodson-et-al.” No better validation of my shot in the dark!
