Its true they have shown the most public progress for RL in general, but Mobileye has shown the most public progress of RL for self driving.
Okay, this is a helpful clarification. I think this gets back to the recurring debate on this forum about demos — whether we should treat a demo as meaningful evidence of real technological progress, or whether we should doubt that a demo tells us anything meaningful about the commercial viability of the underlying technology.
That's like asking does Microsoft have nearly as much compute as Amazon.
They both have enough that its mean-less to ask that question.
9600 TPUs were used for AlphaStar's 600 agents for example, this is nothing for a big company.
That’s a fair point. While we don’t know how much computation will be required to train true self-driving agents, if we assume the required compute will be similar to AlphaStar or OpenAI Five, then even a well-funded startup like Zoox wouldn’t have a problem getting enough compute. A few million dollars a quarter for an AWS/Azure/Google Cloud bill is no big deal.
However it is your view that anything that can be considered as an advantage be portrayed as "Insurmountable, Immense, Moat"
Nope. I’ve been clear that I think some advantages are surmountable. You keep misrepresenting my views. I feel like you aren’t interested in understanding what my views really are.
Some of my views have changed over the last 2 years as I’ve learned new things — I think that’s good. I try not to be too personally invested in being right. I don’t think it matters ultimately. What matters is learning, error correction, and considering alternative ideas — all things that push us in the direction of truth.
If you’re invested in being right, that means you’re invested in whatever you happened to believe at some point in the past being true. So you will look for evidence to confirm that belief. That’s the cart leading the horse. If you’re not invested in being right, that means you are more open to evidence disconfirming that belief. That’s why I think it’s hazardous to get invested in being right. It’s antithetical to finding truth.
I could go more into detail about how and why my views have changed, but you’ve misunderstood both what my views used to be and what my views are now. I don’t see how it’s helpful to continue along that line of discussion if basic communication is failing.
based on a single Elon tweet.
This is something you can only say if you didn't read
the article, or if you don't remember what it said at all. The article cites an in-depth conference call transcript, a Tesla blog post, and a paper on radar-based pedestrian detection from the journal Advances in Radio Science — not just “a single Elon tweet.”
I don’t think radar is nearly as important as I used to think it was, and today I wouldn’t say radar with fleeting learning is better than lidar without fleet learning — they are just two different sensor modalities with their own strengths and weaknesses. But it’s false and also insulting to say the thesis, even if it was wrong, was just based on a tweet. This is the kind of thing that erodes trust that what you say is true, and erodes civil discourse because it feels defamatory — something that is provably false, but would be bad if it were true.
Remember before Elon extended his timeline to end of 2020, you provided a detailed explanation how Tesla could solve self driving by the end of 2019.
I said that to say that you believe that Tesla has "immense lead" and still do, even though all three of the tenants of your thesis has been dis-proven. You even changed your thesis on Mapping and raw data, but you never changed your date or the fact that Tesla has "immense lead". You still believe and promoted the 2019 date right until Elon delayed it again to the end of 2020.
12 months from now? The conclusion i come to is that you will forever see "Tesla" as having "immense lead" no matter what, right up to the very end and that nothing is stopping you from moving the goal post yet again in 2020 and coming up with another reason/thesis why Elon's new timeline of end of 2021 is absolutely correct.
This is completely false and misleading. As I’ve explained to you before, I’ve never made any firm prediction on when self-driving would be solved. I didn’t say it would happen by the end of 2019. The point I’ve made is that sometimes AI progress is not gradual. Or more precisely: the obvious, direct, publicly visible part of progress on specific applications of AI is not gradual.
Some people look at progress on autonomous driving over the last 5-15 years and it feels gradual. In 2007, autonomous cars could pass a minimal urban driving challenge. In 2019, cars can do more complex urban driving with a high disengagement rate and a lot of problems. Waymo’s safety-critical disengagement rate is every 11,000 miles, but its total disengagement rate might be more than once every 100 miles based on rider anecdotes. Self-driving cars are much better than in 2007, but still janky and unreliable, so some people conclude it will be many years or even decades before self-driving is even close to being solved. They would scoff at Amnon’s 2021 target for being feature complete almost as much as Elon’s end of 2019 target. The difference between 1 year or 3 years is not the debate; it’s more like the difference between 1-3 years and 10-30 years.
My point about non-gradual AI progress is to challenge the idea that we are many years, if not decades, away from true self-driving. As an analogy, DeepMind and OpenAI have made non-gradual progress on games like Go, Dota, StarCraft, and Montezuma’s Revenge in a few of years of secretive development. With these games, we didn’t slowly inch toward a solution a little bit more every year, so that it was obvious far ahead of time when we were close and when it would be solved. The solutions came as a surprise to many people.
This argument applies equally to Mobileye launching a full self-driving product in 2021 or 2022 as to Tesla launching one in 2019 or 2020 — or in 2021 or 2022. It’s not a Tesla-specific argument at all, or a year-specific argument. It’s about whether self-driving can be solved within a handful of years, rather than in a decade or multiple decades. Whether it can be solved surprisingly fast, or it will be solved predictably slow.
I don’t feel any constructive discussion is possible if you can’t/won’t try to understand my views, and keep making
strawman arguments.