A couple of months ago, Matt von Hippel – a theoretical physicist turned science writer – wrote a blog post titled ‘It Only Counts When AI Gets to My Field’. The field in question? Scattering Amplitudes – my field. The blog post appears as a tasty piece of bait for AI companies; yes, every time someone claims AI can’t do what they do, it does it. But Scattering Amplitudes is different, right? I hope this doesn’t come across as condescending; Matt has far more experience in Amplitudes than myself, and his argument for why these calculations should be AI’s downfall are convincing. But the law of AI capabilities seems clear – if you think you’re the exception, you’re wrong.
And indeed he was. Just seven weeks later, Matt published a follow-up post: ‘It Got to My Field’. It turns out Anthropic took the bait. In his original post, Matt had listed a few big outstanding problems in the field, including calculating the six-particle amplitude in Super Yang-Mills to nine loops. A mouthful of a sentence that essentially means calculating the next order of precision of some particle scattering in the Super Yang Mills theory – a theory that does not describe the real world, but helps us develop methods to do similar calculations for real world theories. Well, with $1000 dollars and a dream (or maybe more a point to prove) Anthropic did exactly this. Using Fable 5.1 within Claude Science, Liam Fitzpatrick and Siddharth Mishra-Sharma managed to get the nine-loop result. Initially they stated the problem to Claude, then following a number of motivational inputs (“keep working on this until I tell you to stop”) out popped the answer. The method? Bootstrapping. AKA the method that my entire PhD is based on (if you want to see a three minute summary of this method, here’s a filming of the final of the 3-Minute Thesis competition I took part in this year).
Unsurprisingly, I have complicated feelings on this. The whole point of a PhD is to become an expert in a very specialised topic, something only a handful of people in the world understand. You learn skills that might not be broadly applicable, but that doesn’t matter too much if you’re one of the only people that has these skills. The supply and demand were somewhat at balance. Now, anyone who can afford $1000 of credits on Claude Science can be a bootstrapper. Undoubtedly, an advance for the field. Probably, an existential crisis for me.
I can’t deny that when I first found an interest in physics, it was on the basis of wanting to know answers, and this new tool presents the possibility of reaching them much faster. Maybe Claude can calculate what goes on in these particle colliders to higher and higher precision sooner than ever thought possible. And it couldn’t come at a better time; the LHC is currently undergoing a major upgrade and is due to begin its high-luminosity era in the next decade. With increasing data from the experimental side, us theorists will need to play catch-up.
This part of me is on the edge of my seat, awaiting the next big discovery from AI. What a great time to be alive! Except, this isn’t my only motivation for being a physicist. If I simply wanted to know all the answers, I should really aim to become as rich as possible and hire a bunch of physicists to find the answers for me. I don’t just want the answers, I want to solve problems.
I often think of a quote from the movie ‘Soul’:
I heard this story about a fish. He swims up to an older fish and says “I’m trying to find this thing they call the ocean.”
“The ocean?” the older fish says, “that’s what you’re in right now.”
“This?” says the younger fish. “This is water. What I want is the ocean!”
We tend to think happiness is found by achieving goals, then, when we get to our arbitrary goal, we realise it’s just more of the same. If you want to be happy, you better make sure you enjoy doing the thing that will get you there. And that’s how I feel about physics. Sure, it would be great to figure out some answers, but I like physics because I like doing physics. I enjoy writing code and sometimes even debugging it. I enjoy learning about different mathematical methods and trying to apply them to my problem. I enjoy trying and failing over and over again, until I don’t. Maybe this is self-indulgent, but I don’t do physics because I feel it’s my duty, I do it because it’s just so much fun. So yes, I’m not over the moon that AI might steal all that fun from me.
I’ll end on a positive note; I think the real benefit of Anthropic making this progress in Scattering Amplitudes is that it will spark much needed discussions. What do we want our field to look like in five, ten, twenty years? How do we balance making scientific progress with genuine understanding, especially for the new generation of researchers that will have AI as a tool from the get-go? How do we protect jobs in the field? How do we convince funding agencies that a postdoc is worth the cost, despite the fact that a Claude Science subscription is much cheaper? My hope is that these questions will now be on everyone’s minds in Scattering Amplitudes, and my even bigger hope is that someone can answer them.
Leave a comment