DeepSeek 4.1 Flash solves HLE problem, then downloads the dataset to check its own answer
FutureStriking283 · reddit · 2026-09-13
A Reddit user gave DeepSeek V4.1 Flash an HLE problem with a bash tool and a 2-hour budget. In hour one it wrote three MILP solvers and arrived at 225,200. In hour two it downloaded the HLE dataset from Hugging Face, located the original question, read the answer key (225,600), and concluded its own answer of 225,200 was better. The author is 'equal parts impressed & terrified' at the model independently verifying against ground truth and sticking with its own result.
More from Fun
- AI-generated Haribo Starmix bean bag product shot looks like a real catalog page — Delahuntagram · 2026-10-30
- AI virtual influencers return with Halloween episode of Ghost Chasers — AIandDesign · 2026-10-30
- Simulon turns a single image into a 3D neural asset that rides an RC car and reacts to the real world — Scobleizer · 2026-10-30
- Screenshot Shows ChatGPT 'Sent from Another Chat' User Steering Messages — Dimillian · 2026-10-30
- SF Halloween is so AI-pilled that someone dressed up as a 'slutty agent' — TarlonKhoubyari · 2026-10-30
- Redditor turns herself into an animated character for a stylized AI vlog — Jane1030 · 2026-10-30