i think people in ai underestimate how hard it is to do very good science
no bench hacking, no overfitting to tiny gains, no p-hacking, measuring all the important things (scaling, generalization, performance tradeoffs), tuning every baseline optimally
i dont think there are any papers that have done close to perfect science, in many ways, perfect science is impossible (hparam space is too large to search exhaustively)
“With AI we can find new blockbuster drug targets that have escaped everybody’s notice, using public data. It's like getting something for nothing"
Okay, what do you need to get started?
"200 million dollars”
Gemini is still the clear king for useful vision tasks (bounding boxes, object coordinates, etc)
The gap is so large it's honestly strange how far behind the other models are.