[matmatmatmat]

This isn't really here or there, but I've recently been going through the deeplearning.ai course by Andrew Ng and friends, and at the end of each week, there is an interview with a luminary in deep learning.

A couple of weeks ago it was Andrej Karpathy. I got about three sentences in when I realized this guy is really, really smart. The way he spoke about neural nets and the problems he was working on suggested to me a deep and nuanced understanding, and a way of thinking that always tries to expand that depth and breadth.

Anyway, I figure if a guy like that couldn't make it work after so many years, even with a team that surely has other strong players, then it's just out of reach for the time being, with the hardware they're constrained to. It's even possible that deep neural nets will just never be able to do FSD at a level that will gain broad acceptance and some new architecture will be necessary.

[noizejoy]

> I got about three sentences in when I realized this guy is really, really smart. The way he spoke about …

As a non-expert, how would you be able to judge?

ChatGPT also sounds really smart while giving brutally false answers.

[matmatmatmat]

Well, if you'd permit not getting into details, I'm not exactly a total non-expert.

[manmal]

Ah, why so gloomy. Solving this probably comes down to 1) sensors, and 2) computational power available in a car. The sensors used in Teslas were a joke last time I looked (low res, bad low light performance, not enough cameras probably), certainly we can do better? And Moore‘s law is still alive in a way, allowing stunning progress like demonstrated in Chat-GPT. Ever-growing car batteries will also allow for way more power draw for computing. I expect vast improvements in the next few years. Maybe we can get from „drives like a drunk teenager“ to „drives like an overly cautious grandparent“ at least.

[sofixa]

> 1) sensors, and 2) computational power available in a car.

Waymo and Cruise use dedicated hardware and aren't artificially constrained either by preexisting sensors nor compute. Yet they haven't fully solved the problem yet, with Waymo still struggling with unexpected but really should be expected stuff such as road construction, while being available only in specific geographic areas and with quite expensive sensors and years of work.

[lostmsu]

Care to link the interview? I love Karpathy's work.

[matmatmatmat]

https://www.youtube.com/watch?v=xxu4IqwKw0w

[lostmsu]

Dang, I did not realize it is from 2015.