Researchers clash over Sakana AI's bioplausible learning claim: MNIST results don't count
aran_nayebi · x · 2026-09-16
A dispute has erupted over Sakana AI's bioplausible learning research. Researcher Aran Nayebi pointed out the results are only on MNIST — a benchmark where even trivial methods like Feedback Alignment work — and noted his team demonstrated bioplausible learning that scales with backprop on ImageNet six years ago.
Reposting the exchange, dyamins quipped "ImageNet or it didn't happen," guessing the true state of the art in bioplausible learning surpassed this years ago.
More from Fun
- Sarcastic job ad mocks self-proclaimed "AI Safety Experts" bankrolled by billionaires — WolframRvnwlf · 2026-09-16
- Astra nails the 3D candle then drowns it in six lines of pretentious copy — chaseleantj · 2026-09-16
- Gary Marcus jabs Sam Altman: 'What Sam really cares about? Hint: it's not safety' — GaryMarcus · 2026-09-16
- srush clown-moments himself: the 'System 2' joke was literally how OpenAI built thinking models — srush_nlp · 2026-09-16
- Friends vibe code a personal chat app with retro Windows-style UI, planning a shareable release — floguo · 2026-09-16
- Chinese company demos solar panel-cleaning drone by washing a dog on the street — jjvincent · 2026-09-16