Here’s an important little topic.

Here’s an important little topic. OpenAI said
they made an actual discovery in theoretical physics.
Which kinda, you know, hints like “welp, AGI is here.”
I couldn’t walk past it (I love physics and even messed with it a bit at total noob level). Decided to read more closely what’s going on there.
We’re not gonna dive into what MHV amplitudes actually are. But let’s note a few things:
- most of the work on this topic came out in the 80s and a lot of results were already obtained
- GPT 5.2 was “cooked” in a special way to help solve this task and it thought for 12 hours
- the net really did propose an original simplification of the formula based on other results
- the net was prompted by people, and not just people, but very big-name scientists in this field
- it’s important to note that specific formulas were derived earlier, they needed to be generalized, and then humans also checked and proved the solution found was correct
- knowing what can be proven, you can guess how they pulled it off via the net, through RLVR or similar mechanisms like
RLM
(even without knowing the result you can check whether you got something correct and reward the steps that lead to something you can check), I’m not a specialist here, but maybe someone who knows their stuff can say if I’m thinking right
- there was no review of the work

You can read the comments and even find comments from the author of the work
here

What’s notable from my little bell tower:
- it feels more like they sicced brute force on a specific crisp task, i.e. the task formulation was clearly harder than its solution (in the sense that solving is extremely hard, but possible since you can verify)

A very important comment from the paper’s author
Hi I am an author of the paper. We believed that a simple formula should exist but had not been able to find it despite significant effort. It was a collaborative effort but GPT definitely solved the problem for us.
And then this one
There are closely related "MHV amplitudes" which naively obey a really complicated formula, but for which there famously also exists a much simpler "Parke-Taylor formula". Alfredo had derived a complicated expression for these new "single-minus amplitudes" and we were hoping we could find an analogue of the simpler "Parke-Taylor formula" for them.

So for now it’s cool, but not scary yet, still. Reminds me of the invention of packages like Mathematica and how they got used in the scientific community. Of course in this case AI pushed the bar to a new level.
Well then, are we waiting for a warp drive?

Links from the original post:

• https://openai.com/index/new-result-theoretical-physics/
• https://www.primeintellect.ai/blog/rlm
• https://news.ycombinator.com/item?id=47006594