Jonathan Oppenheim

I warned my students that learning Bell's theorem could be a mentally destabilising experience, but they didn't listen. Now who will be there to pick up the pieces?

Video of the talk at superposer.substack.com/p/large-langua… along with a link to Tobias's github, a demo, and a wild conversation I had with Claude where it revealed its system prompt (the hidden instructions that tell the model how to behave, before you even type anything).

In late 2025, tools Like Claude Code crossed a threshold. Like hyperactive master's students, their physics calculations can now be made more reliable with domain expertise and some scaffolding. Check out Tobias Osborne's talk on using a swarm of agentic verifers to do physics.

@maria__violaris They caused me to shift from research in quantum gravity. I felt that we needed a deeper understanding of quantum theory to make progress, and quantum information theory gave us the tools to go there.

@postquantum What did his work cause you to shift fields from?

Congratulations to Charlie Bennett and Gilles Brassard for winning the 2025 Turing Award (akin to the Nobel Prize for Computer Science)! They invented quantum cryptography and quantum teleportation. 1/

It was reading his early works, that got me interested in quantum information theory and caused me to shift fields, so I owe a lot to him. This paper on the thermodynamics of computation is a classic: sites.cc.gatech.edu/computing/nano… 3/3

Charlie was responsible for our understanding of entanglement (as a resource) and was a founder of quantum information theory. He used info theory to exorcise Maxwell's demon (the demon's brain has to be reset like a memory tape and this costs work). awards.acm.org/turing 2/

@KSHartnett x.com/i/status/20341…

As recently as a few years ago, my impression is, knowledgeable people thought AI would be better at writing than coding. Obviously hasn't turned out that way. I think it all comes down to surprise. Great writing constantly surprises you. Within a single sentence it takes turns you didn't expect. AI is built not to surprise. (This is why, of all forms of writing, it may be worst at humor.)

They're fast at coding, which is a great advantage. Being fast at writing mediocre poetry is not. Their math skills are also pretty unreliable. They sometimes do a PhD level math calculation while sometimes outputting nonsense.

LLM's are probably just as good at poetry as they are at coding, but code just has to compile and implement the function. Poems need to be beautiful.

I guess the DoW shouldn't have switched to ChatGPT to make these decisions....

Where are you @demishassabis @sama @sundarpichai ?

I would guess the opposite will happen. Bespoke software is now very cheap, but we will need a horde of nerds to arise and integrate these systems into current infrastructure, maintain the vibe-code, patch the security holes etc.

@AlessandroStru4 Partly it was the scaffolding that was helpful (skill files etc) but I found Opus 4.5 to be better at calculating, while Gemini 3 I found better at research taste. It may be domain specific.

@postquantum Is 'Claude' better for theory than other AI?

"Good luck to future me." That was Claude's farewell before I wiped its memory for the fourth time. I finally got an AI to do a research-level physics calculation correctly via a Groundhog Day loop. Details below and in my post 🧵1/7

@arielchouminov @PTrubey Well, it's taken awhile for it to become reliable enough that it saves more time than it wastes.

@PTrubey @postquantum It's incredible to think that research still hasn't fully adopted AI. I'm guessing an exponential amount of innovation to come.

Must read for everyone interested in how fast AI is evolving and how to use it to solve the most complex problems we have. @postquantum is a professor of Quantum Theory at University College, London. He is currently pursuing a new theory that would unite quantum mechanics and general relativity, a problem that has bedeviled theoretical physics for 100 years. Needless to say the mathematics involved is daunting. Recently he's had a go at using the latest AI tools to see if they could replicate calculations that his team takes weeks to figure out. To make this happen, he invented various techniques described in this very readable blog post. I was struck by how these techniques could relatively easily be incorporated into AIs we currently have through scaffolding to make them geniuses out of the box. As his conclusion shows, the best human researchers are still better, but the AIs are in hot pursuit. Enjoy.substack.com/home/post/p-18…

The science slop is coming and the tools are not reliable...yet. But with the right setup, they will be able to significantly enable research-level work. Full post: superposer.substack.com/p/teaching-an-… 7/7

After several rounds, the skill file contained hard-won wisdom from iterations it couldn't remember. Fresh Claude, no memory of any of this, reads the skill file, does the calculation. Five minutes. Right answer. 6/