🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

A little-known corollary of the second law of thermodynamics is that it is easier to break than to build. Be careful breaking what you have inherited. It might be more fragile and harder to rebuild than you think.

If we are heading towards a future of AI led autonomous science, expect data collection to still be the bottleneck. Your AIs will idly ponder eternities while waiting for new information about the world to trickle in.

Why isn’t thinking interleaved with answering in the current gen of reasoning models? Maybe the context is long enough that they are effectively carrying the information of the answer through the reasoning trace and so there isn’t much benefit. Or maybe they are doing this

I recently read Cat’s Cradle by Vonnegut. This is the first Vonnegut that I’ve read since a compelled Slaughterhouse reading in my youth. At times it reminded me of something Walker Percy might write, but without the stubborn doubt about doubt. '“Beware of the man who works hard

You shouldn’t be making your assignments LLM-proof. That’s not really the point. Would you make your engineering assignments calculator-proof or python-proof? Alternatively, maybe you think you will make the problems so hard that the AI can’t solve it. Supposing you could, are

In the near future it may be that when you ask one of the big AI systems to solve significant problems they may think for a few minutes and then give you a cost estimate for solving that problem. You will have to decide how much that labor/inference is worth. Sure, I can write

The moment you realize that you can’t do some part of your job without AI, you should start asking what it would take for AI to do that part of your job without you.

At this stage, the adoption and integration of AI is more of a bottleneck for leadership than model intelligence. The nations and organizations that recomp around AI fastest can become de facto leaders in AI, even if they don’t develop the models themselves. Org charts, physical

I had a lovely visit to Argonne CSE division today to give a seminar on new flavors of machine learning for chemical problems. Huge thanks to Stephen Klippenstein for hosting me. It was both humbling and exciting to learn how much that I didn’t know about gas-phase chemical

We already have much more intelligence available than we are utilizing.

There will always be a timescale mismatch between the speed of thought and the speed of an experiment, but great things happen when this gap shrinks. How likely are you to pursue a potentially high impact research idea that will take 2 decades to discover is possibly misbegotten?

I had a delightful visit to the University of Chicago yesterday to discuss how machine learning is pushing the prediction frontier in chemistry. Thank you to the inspiring Andrew Ferguson for hosting me and the materials group within PME for teaching me so much. It’s high praise

I’m deeply honored to receive the 2025 JPCA Lectureship Award. Some of the first academic papers that I ever read were from JPC. Thank you to my mentors, research group, and those that support our work: axial.acs.org/physical-chemi…

The best equilibrium for a human is non-equilibrium. #equilibriumisdeath #statmechwisdom

Tip: Close all applications when you are done with work. You can reopen what you need when you are ready to work again. This includes your browser and email client. It will do wonders for your mental hygiene when you are moving between tasks or trying to start your day.

There are many problems that are easy to check but hard to solve. Many of the hopes for AI agents depend on whether alignment is one of those problems.

Generalization is very hard without mechanisms. Mechanisms provide the ability to predict the full scope of downstream behaviors, but observations only allow you to weakly interpolate possible behaviors. Compare medicine pre and post germ theory. Compare foundation AI pre and

Right-sizing the time per task…I’m jealous of these AIs.

I’ve been taking a break from social media until I catch up on some overdue writing, but I have to break this self-imposed restriction to highlight some of our collaborative work with the Letian Dou group that just appeared in Nature Chemistry. Check out the high MW