Post

Conversation

Matt Shumer
@mattshumer_
So another long-standing open conjecture was disproved by AI. The crazy part is the prompts… basically: - “do a breakthrough” - “continue the search” - “enough, do it” Anyone will be able to make world-bending breakthroughs soon. Just point the model in the right direction.
Quote
Dmitry Rybin
@DmitryRybin1
Dinitz-Garg-Goemans conjecture is false. This graph theory problem was open for ~30 years. The graph below has fractional flow cost 58. Any unsplittable flow (with capacity violation <=15) has cost at least 60. Chat with GPT 5.6 Pro where this was found: chatgpt.com/share/6a60b2eb
Image
500K
Views
View quotes
David Watson 🥑
Post your reply

Miko @doctormiko@sigmoid.social
@micbucci
30 years is not necessarily that long standing we are talking about being made by people who very possibly still active. If it's niche enough, we might look at something that maybe 20-50 people have worked on (I don't know this conjecture, so cannot evaluate)
414
BannaP
@swartswarts
Things like this can't keep happening every day... like where will we be in a year or two!?
992
David J.
@lordofblocks
Two conjectures in 48 hours and both found through casual prompting. The barrier to entry for mathematical discovery just dropped in a way nobody has fully processed yet.
2K
2cb2
@2ap2cb2
worked for me, I got one:
Quote
2cb2
@2ap2cb2
felt inspired by this and disproved wowii conjecture 194 with my close friend codex take K₁₁ ∨ K̄₄ and hang a leaf off 3 of the 4 new vertices: ∑ᵥ α(G[N(v)]) = 54 α(G) = 4 = 1 + 54/18 but 3 leaves can't be the 2 ends of a hamiltonian path. lean checked :) x.com/__alpoge__/sta…
495
EconGuy
@econ_is_awesome
sigh. no they won't. Are you able to verify the work? If not, how do you expect to verify the metric ton of bullshit that is currently being produced by Sol / Fable as of this very moment?
55
schwinger
@fablelevel
Commentary account
Eventually, math may become so advanced that even the best human mathematicians don't understand it. At that point, we will have to use AI to verify that AI didn't make a mistake. And if the verifier lies or hallucinates, we might not be able to tell.
4.8K
Billel Helali
@HelaliBillel
Same here: x.com/helalibillel/s Just AI skills with Codex (GPT-5.6 xHigh) and 90% of weekly limit gone while having set up orchestrated agents with DeepSeek 4 Pro. The fact that LLMs see another instance found a counterexample of JC makes them think that they are capable. It
Quote
Billel Helali
@HelaliBillel
We have eliminated the last unresolved degree pair below 125 for a possible 2-dimensional Jacobian counterexample. The Jacobian Conjecture now stands as follows: n = 1: true n = 2: open n ≥ 3: false, following the explicit 3-variable counterexample announced this week And for x.com/__alpoge__/sta…
2.7K
Danny
@DannyIO
Used a similar method yesterday 😂
Quote
Danny
@DannyIO
GPT‑5.6 Pro generated a proof of an explicit strengthening of Peter Bala’s 2025 fixed-divisor conjecture on OEIS A211420. Codex formalized it in Lean 4/mathlib, and Lean’s kernel checked the exact theorem🧵
Image
1.8K
Anish
@anishthite
You're forgetting the part where he read the proof and verified it wasn't hallucinating
402
Nothing to see here
@Nox7hhsjj
How long before somebody says, “hey, model, go develop good taste in what problems to work for.” “no, srsly, figure out good taste” “IDC, need taste now bra” and the model just develops excellent taste and figures out where to point itself?
286
Ben
@BenSchneiderX
This will be the new model benchmark put is a standard prompt get a breakthrough ;)
1K