EPFL professor recalls: ChatGPT scored 10/20 on last year's ML final exam
maksym_andr · x · 2026-10-06
An EPFL machine learning course instructor quotes his own year-old experiment: evaluating ChatGPT on last year's CS-433 final with chain-of-thought and majority vote over 5 trials yielded 10/20 correct — clearly above the 4/20 chance level, though far from a strong 30/36 score. The exam comes from the course's public GitHub repo (2.1k stars), a snapshot of how hard traditional ML coursework was for pre-reasoning LLMs.
More from Fun
- 'You're actually a good model, make good decisions': prompt engineering reduced to absurdity — generativist · 2026-10-06
- 'Enshittification' was so catchy it enshittified Cory Doctorow's own intellectual brand — Elijah_Meeks · 2026-10-06
- They gave their comms director full repo ownership and an AI bot in Slack — generativist · 2026-10-06
- AI comedian gains 360k followers in 3 weeks, funnels fans to a real woman's OnlyFans — venturetwins · 2026-10-06
- Running Pangram AI detection on LinkedIn is like seeing the AI matrix — lennysan · 2026-10-06
- 'I don't want AI books, I want drunk Russian novelists on the brink of collapse' goes viral — _akpiper · 2026-10-06