Tesla is ‘no longer compute constrained’, running out of excuses for self-driving Leave a comment

Elon Musk says that Tesla is “no longer compute constrained” in its AI training – meaning that it is starting to run out of excuses for not achieving its promised self-driving.

Self-driving is an incredibly difficult and important problem to solve.

Improvements in safety over human drivers could save tens of thousands of lives per year, and cutting down on drive time could free up billions of hours per year – resulting in incredible value creation.

Normally, we would give a lot of slack to someone trying to solve this difficult issue, but Tesla has shortened its own slack by selling its “Full Self-Driving” package to customers since 2016 and promising to achieve self-driving every year since 2019.

Despite yearly guidance that it is about to achieve full self-driving, Tesla has had some arguably good excuses for not being able to achieve its goal.

For example, neural nets didn’t power vehicle controls. Early Tesla comments about drivers “teaching” its self-driving system pointed toward this always being the case, but Tesla didn’t introduce neural net-based vehicle controls until FSD beta v12, which is being rolled out right now.

Another excuse has been compute power limitations creating a bottleneck in its neural net training.

Tesla’s Dojo supercomputer program was meant to fix that, but the automaker has seen significant delays in that program.

To compensate, Tesla has been investing billions of dollars in Nvidia computing power to create new supercomputer clusters to train its neural nets.

Now, CEO Elon Musk says that Tesl is “no longer AI training compute constrained”:

While Musk didn’t elaborate on the situation, it sounds like Tesla brought more compute power online recently because the CEO was complaining about compute power being a bottleneck just a few months ago.

Electrek’s Take

This is the combination, training compute and end-to-end neural nets, that FSD fans have been hailing as the last piece of the puzzle for a while now.

In theory, it makes sense to me. If Tesla’s approach is the right one, now that everything is powered with neural nets and now that it has the computing power to train the neural nets, we should see a massive increase in the rate of improvement.

But we need to see that happen now because there are no more excuses after this.

Tesla needs to start releasing clear data that shows a path to FSD becoming a true unsupervised self-driving system as promised. Otherwise, it’s going to have to admit that it can’t achieve level 4 or 5 autonomy.