This is a really interesting video:
It's from the investor comms event back in April of this year where they went through the FSD tech. Elon brought along the design leads on both the hardware and the software and they both did excellent presentations. I learned a lot from it. Even if you don't want to listen to the tech itself, it's clear from listening to the tech leaders, and Elon himself, that Tesla are hiring some of the finest minds to crack the self-driving problem and investing a ton of money in it.
The fundamental design approach is based on image processing from visual data (just like us humans). Lidar is fitted to the cars, but it's used only as a secondary, confirmatory piece of data to the visual processing. Elon at one point very forcibly pointed out that anyone trying to use lidar as a primary way of self-driving was doomed to failure. I get this since the real world is so complex that lidar alone couldn't hope to deal with it.
As an example of just how well the neural net tech is working, the section of footage around 1:05:16 shows the FSD computer very successfully predicting the bends on a country road. Tesla take vast quantities of real-world video data (i.e. from real Tesla cars, driven by real drivers, in the real world) and annotate it, by hand (with humans!), to train the machine as to what's actually happening. Then, when you and I are driving, the computer is interpreting the video images in real time against this pre-learned data. To do the necessary computations quickly enough, Tesla have designed a dedicated chip focused entirely on the needs of self driving.
All in all, I think the tech is quite amazing really, and it shows just how much Tesla are investing in the field and (probably) how far ahead they are of the competition.
But then I sat back and reflected on my own experiences of AP - and they are not positive. I've only had the car just over a month and yet I've had the whole gamut of issues that I'm sure you're all familiar with - phantom braking, reduced capability due to obscured cameras (by dirt/condensation/water etc), false emergency lane departure events, false alerts when traffic safely crosses our path, and so on and so on...
I get that more data and more learning will make things better, but what if:
- The cameras and hardware aren't good enough or reliable enough (e.g. our eyes have eye lids to clear dirt, we can rotate our eyeballs/head to concentrate on specific things) - none of which the car's cameras can do.
- Elon's fancy chips are still 1000x too puny to do the job well enough, fast enough.
- The whole neural net idea just lacks the necessary "intelligence" to do the job properly.
These possibilities aren't completely left field. Neural nets are fundamentally just pattern matching to previous data - there's no intelligence behind any of the inferences, just a probability fit (and yes, I know that I'm verging on the philosophy of intelligence and whether "AI" really is intelligence at all!).
Anyhow, random thoughts for a Sunday evening!
It's from the investor comms event back in April of this year where they went through the FSD tech. Elon brought along the design leads on both the hardware and the software and they both did excellent presentations. I learned a lot from it. Even if you don't want to listen to the tech itself, it's clear from listening to the tech leaders, and Elon himself, that Tesla are hiring some of the finest minds to crack the self-driving problem and investing a ton of money in it.
The fundamental design approach is based on image processing from visual data (just like us humans). Lidar is fitted to the cars, but it's used only as a secondary, confirmatory piece of data to the visual processing. Elon at one point very forcibly pointed out that anyone trying to use lidar as a primary way of self-driving was doomed to failure. I get this since the real world is so complex that lidar alone couldn't hope to deal with it.
As an example of just how well the neural net tech is working, the section of footage around 1:05:16 shows the FSD computer very successfully predicting the bends on a country road. Tesla take vast quantities of real-world video data (i.e. from real Tesla cars, driven by real drivers, in the real world) and annotate it, by hand (with humans!), to train the machine as to what's actually happening. Then, when you and I are driving, the computer is interpreting the video images in real time against this pre-learned data. To do the necessary computations quickly enough, Tesla have designed a dedicated chip focused entirely on the needs of self driving.
All in all, I think the tech is quite amazing really, and it shows just how much Tesla are investing in the field and (probably) how far ahead they are of the competition.
But then I sat back and reflected on my own experiences of AP - and they are not positive. I've only had the car just over a month and yet I've had the whole gamut of issues that I'm sure you're all familiar with - phantom braking, reduced capability due to obscured cameras (by dirt/condensation/water etc), false emergency lane departure events, false alerts when traffic safely crosses our path, and so on and so on...
I get that more data and more learning will make things better, but what if:
- The cameras and hardware aren't good enough or reliable enough (e.g. our eyes have eye lids to clear dirt, we can rotate our eyeballs/head to concentrate on specific things) - none of which the car's cameras can do.
- Elon's fancy chips are still 1000x too puny to do the job well enough, fast enough.
- The whole neural net idea just lacks the necessary "intelligence" to do the job properly.
These possibilities aren't completely left field. Neural nets are fundamentally just pattern matching to previous data - there's no intelligence behind any of the inferences, just a probability fit (and yes, I know that I'm verging on the philosophy of intelligence and whether "AI" really is intelligence at all!).
Anyhow, random thoughts for a Sunday evening!
Last edited: