GPT-5.6 Sol helps disprove Maxwell's 150-year-old conjecture
The same model you code with proposed the counterexample that broke a bound standing since the 1870s. How to aim Sol at hard problems, and the caveats.
Copy markdown
The counterexample, in one picture
Put three equal charges on a triangle and two tiny ones just above and below the center: the central balance point splits into 21, giving 24 equilibria from 5 charges — past the (n−1)²=16 maximum assumed since Maxwell's 1870s work. GPT-5.6 Sol proposed the construction; mathematicians at Babson, Missouri and UMBC verified it in a four-page arXiv proof.
It's the model you already code with
This is Sol — OpenAI's frontier GPT-5.6 tier, live in ChatGPT and the API. The signal for builders: it's now credible enough to point at hard reasoning, proof-search and ideation, not just code. Sol suggested the key construction; the humans did the rigorous checking.
Read the asterisk before you cite it
It's an arXiv preprint, not yet peer-reviewed, and Sol suggested rather than proved the result — every step was human-verified. The durable workflow: AI proposes, you verify. Treat model output as a lead, not a citation.