Hacker News Reader

Comment by javierluraschi | Hacker News Reader

Historically, there have been significant bottlenecks on algorithms, I'd argue software plays a bigger role than hardware; especially during inference.

Going back to AlexNet, while GPUs helped the training process, it included significant software improvements that made a huge difference: training with dropout, convolutional neural networks, using rectified linear units, etc.

Also, training hardware is much more relevant than hardware for inference. Tesla's cars do not include hardware to train the neural network but just for inference. The training hardware is actually a server cluster that Tesla owns. So if hardware is insufficient, Tesla might have to buy a larger training cluster, but not necessarily upgrade the inference hardware.

That say, I do agree that there is some risk in requiring a hardware upgrade for inference, but to me that seems very unlikely.

View on HN

so, you're saying it doesn't work then?

Correct! I'm saying FSD only works on 90%-99% of cases, which is currently insufficient to be actually considered FSD. Keyword, currently.

However, I'm NOT stating that this won't ever work. I do believe that FSD will work in the future. Are there risks? For sure. But I'm very optimistic.

I'm trying to communicate that, even with current in-car hardware, this should be sufficient. Why? Cause driving the car with a neural-network is actually quite trivial. Now, training the neural-model is actually REALLY hard. But the training does not happen in the car, car just collects data and sends it to a data center with GPUs that trains the network. If any hardware will need upgrade, is the data center were training happens, not the actual in-car hardware.

Are you really saying that FSD will never work? Cause that's a pretty bold claim.

I'd add, we can be more specific about what a "case" is. I'd say, a trip might describe this better. As in, 90-99% of trips FSD Beta 9 will take you there without interventions. That's still lame, but still quite impressive. Achieving 90-99% accuracy on trips still seemed impossible one year ago, I don't think that's the case anymore. The remaining risk is, how fast can this improve between 99% to 99.9% and then to 99.99%, etc.

The reasonable expectation is that cars will not be able to achieve actual self-driving without either AGI or more specialized hardware than a camera. AGI is almost certainly a 'next century' kind of technology. Adding sensors other than cameras (lidar, others) to a Tesla means that they don't have the hardware.

This claim is quite obviously true, because despite the claims of Tesla's people, we don't know of any vision system that works without relying on general intelligence. Human/animal vision in particular is only accurate because we inherently use our knowledge of the world and physical intuition to resolve ambiguity. Put a human (or other mammal or a bird) in a fully artifical environment, without recognizable objects and without shadows, and you'll see that they are about as bad at vision and navigation tasks as many current algorithms. The effect will be even worse if you recreate realistic objects but with unrealistic proportions and behaviors.

Since we can't hope that computers will gain animal level intelligence in understanding the world (no one is even really working on the necessary problems), the only path to self driving is through more advanced sensors, as everyone but Tesla has realized for a long time. This is why Waymo has self-driving cars on the road today (in perfect ideal conditions), and Teslas can sometimes barely navigate an empty intersection without a human driver intervening.

> I'm saying FSD only works on 90%-99% of cases

This is an extremely bold claim. You're also ignoring the fact that driving is an extremely heavy tailed distribution. We know that both processing heavy tailed distributions is difficult (not training, but pre-processing) as well as we see a strong correlation between larger networks and performance on heavy tailed distributions (more normal like distributions can get away with substantially smaller networks). But larger networks means more MACs. More MACs means slower inference. There's still a tradeoff between accuracy and speed. In driving you need both.

> Are you really saying that FSD will never work?

No one here is saying that. Everyone here is calling you out on your BS that we _know_ that FSD will work on _current_ hardware. We don't _know_ that. Maybe it will, but it's not a pony I'm willing to put money on.

Reply on news.ycombinator.com