4 ms·

Waymo is doing actual for customer self driving right now.They need to scale it up a lot, but 5-10 years is a lot of time in AI research. Deep Learning did not

Waymo is doing actual for customer self driving right now.They need to scale it up a lot, but 5-10 years is a lot of time in AI research. Deep Learning did not exist 10 years ago. Tesla is trying hard too, but it goes with a big dose of bullshiting from Musk. They are a lot more dependent on AI break troughs in DL due to their vision only philosophy. OpenAI research around scaling laws in DL suggest that naive 1000 increase in training compute over CLIP should get us to state of the art vision in zero shot settings i.e. without task specific training data. But they probably can cover 1/3 of that with algorithms improvements. Beyond that machines should really be above human vision across the board besides adversarial settings. Quote https://arxiv.org/abs/2103.00020 https://arxiv.org/abs/2103.00020 : > While scaling has so far steadily improved performance and suggests a route for continued improvement, we estimate around a 1000x increase in compute is required for zero-shot CLIP to reach overall state-of-the-art performance. There is also the question of robust car control, but there should be good progress on it too due to RL research.

gizmo 6y ago
Waymo has cars outfitted with 100k in sensors, HD maps, and their cars are geofenced to a few neighborhoods with excellent visibility. What Waymo is doing is way easier than solving the general problem of self-driving, where cars have to figure out safe solutions to not-seen-before problems. I think it will take years before Tesla FSD can drive from A to B without making catastrophic errors, but at least they're trying to solve the correct problem. And I think it's important to remember that it's OK if self-driving cars behave stupidly in ways that don't kill people, because the status quo has people dying every day because they're drunk/tired/texting, and self-driving only has to outperform the worst drivers in order to be a big leap forward.
- threeseed 6y ago
  Tesla is not solving the correct problem i.e. “how to build the best FSD car with the best technology” They are trying to solve the “how to build FSD that works on existing Tesla cars because Musk has tricked a whole bunch of buyers into handing over money” problem. Our heads constantly move around and have the worlds most advanced neural network trained over decades to understand objects and how they behave. So for us vision works. But because Tesla isn’t solving the right problem they insist on going down this road instead of LiDAR which is safe and works.
  - tibbydudeza 6y ago
    Evolution took only 6 million years to get the wetware sorted that is then trained daily for years to get it in a state to make it able to drive with approx. 4 accidents occurring during it's entire lifetime.
- ahartmetz 6y ago
  I think Waymo is going about it the right way, given the reliability required: Solve 100% of the problems in a restricted environment, then gradually remove restrictions on the environment and solve 100% of the problems after every step. I think this offers some divide and conquer benefits and guards against introducing sloppy hacks that are hard to disentangle from the system later. Sounds logical, doesn't necessarily mean it's true. I'm not an expert.
gambiting 6y ago
Sure, but I don't believe Waymo will be able to sell this tech in a regular vehicle that any customer can buy. They will run a fleet of their own taxis where they accept full liability for the system.
- exhilaration 6y ago
  If that fleet is big enough, none of us will need to buy cars. That's the future of self-driving cars.
  - gambiting 6y ago
    So I keep hearing this argument and I just 100% don't believe that. It would already be cheaper for me to sell my car and just take an Uber wherever I need to go. Or get rid of my SUV and get a tiny car for 1/2 monthly cost and just rent something larger when I actually need it. So....I know this, and yet I still prefer to have my own car that I drive myself. What do self-driving cars solve here? A self driving car is literally like a taxi, minus a driver. Yet even if taxis were quarter of their price, I would still like to have my own car. But sure, maybe it's just me. But I think a lot of people prefer to have their own vehicle rather than take taxis everywhere, even if it's logically cheaper thing to do.
etimberg 6y ago
Deep learning took off in the last 10 years, but the idea has been around for more than 50 years. https://en.wikipedia.org/wiki/Deep_learning#History https://en.wikipedia.org/wiki/Deep_learning#History
- Isinlor 6y ago
  Good catch. Still, I would say that it is the 10 last years that matter, because that's when people started to take it really seriously. It's also interesting to think that we may have a lot of ideas like that. They are just waiting patiently for compute in some small communities. And the compute quite consistently grows exponentially. We will make bigger absolute increases in compute in next 5 years than we did ever before.
Bancakes 6y ago
Yeah.. I'm not touching self-driving until all cops and judges make it clear who's responsible for crashing/speeding/etc.

gizmo 6y ago

Waymo has cars outfitted with 100k in sensors, HD maps, and their cars are geofenced to a few neighborhoods with excellent visibility. What Waymo is doing is way easier than solving the general problem of self-driving, where cars have to figure out safe solutions to not-seen-before problems. I think it will take years before Tesla FSD can drive from A to B without making catastrophic errors, but at least they're trying to solve the correct problem. And I think it's important to remember that it's OK if self-driving cars behave stupidly in ways that don't kill people, because the status quo has people dying every day because they're drunk/tired/texting, and self-driving only has to outperform the worst drivers in order to be a big leap forward.

threeseed 6y ago
Tesla is not solving the correct problem i.e. “how to build the best FSD car with the best technology” They are trying to solve the “how to build FSD that works on existing Tesla cars because Musk has tricked a whole bunch of buyers into handing over money” problem. Our heads constantly move around and have the worlds most advanced neural network trained over decades to understand objects and how they behave. So for us vision works. But because Tesla isn’t solving the right problem they insist on going down this road instead of LiDAR which is safe and works.
- tibbydudeza 6y ago
  Evolution took only 6 million years to get the wetware sorted that is then trained daily for years to get it in a state to make it able to drive with approx. 4 accidents occurring during it's entire lifetime.
ahartmetz 6y ago
I think Waymo is going about it the right way, given the reliability required: Solve 100% of the problems in a restricted environment, then gradually remove restrictions on the environment and solve 100% of the problems after every step. I think this offers some divide and conquer benefits and guards against introducing sloppy hacks that are hard to disentangle from the system later. Sounds logical, doesn't necessarily mean it's true. I'm not an expert.

gambiting 6y ago

Sure, but I don't believe Waymo will be able to sell this tech in a regular vehicle that any customer can buy. They will run a fleet of their own taxis where they accept full liability for the system.

exhilaration 6y ago
If that fleet is big enough, none of us will need to buy cars. That's the future of self-driving cars.
- gambiting 6y ago
  So I keep hearing this argument and I just 100% don't believe that. It would already be cheaper for me to sell my car and just take an Uber wherever I need to go. Or get rid of my SUV and get a tiny car for 1/2 monthly cost and just rent something larger when I actually need it. So....I know this, and yet I still prefer to have my own car that I drive myself. What do self-driving cars solve here? A self driving car is literally like a taxi, minus a driver. Yet even if taxis were quarter of their price, I would still like to have my own car. But sure, maybe it's just me. But I think a lot of people prefer to have their own vehicle rather than take taxis everywhere, even if it's logically cheaper thing to do.

etimberg 6y ago

Deep learning took off in the last 10 years, but the idea has been around for more than 50 years. https://en.wikipedia.org/wiki/Deep_learning#History https://en.wikipedia.org/wiki/Deep_learning#History

Isinlor 6y ago
Good catch. Still, I would say that it is the 10 last years that matter, because that's when people started to take it really seriously. It's also interesting to think that we may have a lot of ideas like that. They are just waiting patiently for compute in some small communities. And the compute quite consistently grows exponentially. We will make bigger absolute increases in compute in next 5 years than we did ever before.

Bancakes 6y ago

Yeah.. I'm not touching self-driving until all cops and judges make it clear who's responsible for crashing/speeding/etc.