2 ms·
Yeah, I think most people who haven't worked with machine learning may not realize this, and it's extremely important. It can be pretty easy to reach 90%, but i
by mort96 5y ago
Yeah, I think most people who haven't worked with machine learning may not realize this, and it's extremely important. It can be pretty easy to reach 90%, but impossible to reach 95%. People tend to assume some sort of progression, where you just "iron out" the remaining quirks, but that only applies to traditional engineering. You can gather 100x the training data and build a model with 100x the number of nodes and get a 10% average improvement with some cases getting worse and others getting better.
I think the current state of FSD is extremely impressive. There are clearly some extremely talented engineers working on that system. But my guess would be that it simply can't be incrementally improved until it's road-safe, in much the same way that GPT-3 is insanely impressive but can't be incrementally improved until it can write a whole human-quality novel. I suspect a new wave of radical rethinking of machine learning might be necessary, and that might be decades away.