That same sentence made me stop reading. It's perhaps not an obvious point, but there are several dimensions that require excellence to succeed in entrepreneurship, and failing at passing the bar on any one of those dimensions make the probability of success drop dramatically.
Net win is the key word here. These systems are designed to improve safety, which is a pretty low bar, and should not be marketed to the contrary. NCAP standards read like they were written by the manufacturers themselves, to be able to pass in easy conditions.
Agreed. There's a huge difference between mathematical beauty in theoretical physics and in economics, the latter not really being on the radar in modern math. The author is conflating mathematical beauty with the "fancy math" effect of being able to easily publish papers in fields that aren't very mathematical just by having equations as part of the analysis. It's not like economists have found some amazingly beautiful theory (by modern math standards), they just use math to promote/legitimize their work.
"LASER achieves these results by embedding all languages jointly in a single shared space (rather than having a separate model for each)". There could be a good reason for why the mutual embedding of several languages works better than individual, beyond the extra data. If human languages share some minimal representation (universality so to say), training on multiple languages may be required to extract it with today's techniques, since training on just one language is bound to overfit to its particulars.
I don't at all recommend limiting yourself to self-study when you start this endeavor, it's too daunting given all the unknown unknowns. To start with it helps a lot to take courses, even if it means sitting in/enrolling in night courses at the local community college. It provides a reference point to what others understand, allows real-time querying with an instructor, and is more motivating than self-study. Once you have completed and fully grasped a difficult proof-based course (real-analysis, algebra, etc), then you can possibly embark on a successful self-study. I'm not saying it can't be done, but you will get to mathematical maturity faster this way and there will be less risk of early abandonment.
These statements are consistent modulo zero measure (probability) sets. For instance: there are infinitely many rational numbers, but the probability of picking a rational number at random from the interval [0,1] is zero.
Helm.ai | Onsite | Bay Area | Research Engineer, SWE, SRE
Helm.ai is building the next generation of AI technology for autonomous driving. On the road toward full stack L4/L5 AI software, we are productizing in the ADAS space with a superior L2 offering.
"But then it seems that inevitably you get to the point where your real-time classifiers are basically running the show anyway so are you back to - what good is a Map anyway?"
Yep, pretty much. Robust real-time perception, for a wide variety of tasks that happen very rarely is one of the key bottlenecks.
Hey, I'm not against recreational activities, everyone needs them to stay sane. It becomes different when an accidental bitcoin millionaire decides to put real money behind a meaningless task to watch someone else dance. At that point it stops being recreational for the problem solver, but rather a form of work.
Cogently dismissing other pursuits is an incredibly effective form of recruitment, so there's nothing inherently academic about it.
"Do you want to sell sugar water for the rest of your life, or do you want to come with me and change the world?”
"My point was perhaps too broad - but deigning to tell someone their pursuit isn't as worthy as one you think they should be pursuing is, frankly, pointless."
That seems to assume immutable human values, and/or a postmodern view of the world. Promoting thoughtfulness about the meaning of human activity seems pretty valuable, regardless of the conclusion.
Actually it's probably one of the most worthwhile questions, philosophically and practically. If this question doesn't get properly considered, we end up with blind glorification of activities whose main selling point is revenue generation and shitty feedback loops in society over generations.
I actually have nothing against frivolous spending, since buying a yacht doesn't subtract from the high end talent pool. If you're going to put up a few mil to solve a math problem, why not something like P v. NP, or the Riemann hypothesis, or countless applied math problems. Instead it was spent on some kind of spectacle of arbitrary mental effort.
One can think of many better ways to harness human intelligence than work on an arbitrary puzzle... unless this is an interview for something much bigger.
"Bringing Robotaxis to market is very capital intensive". With the front and center approaches rooted in the DARPA Grand Challenges yes, with much of that money being spent on propping up simplifying assumptions that break at the tail end and on technologies that will be obsolete in a couple years. HD mapping curated by humans, human annotation for training/testing data, Lidars, and all that superfluous real world testing, will soon be replaced by approaches that are orders of magnitude more capital efficient. The less efficient approaches that currently dominate will sputter at 1% market penetration at best.
It seems like a stretch to call this a big discovery in the math world. Even if it had been published in a top journal etc, it would be unlikely to be heralded as a major achievement.
There are some interesting connections between GPs and neural networks. Deep neural networks with random weights concentrate about gaussian processes, so the two may not be so far apart in practice.
Exactly. But hey cool concept cars make for pretty pitchdecks. Technology will solve itself if you just throw money at it, right? Next up, we'll have startups whose main differentiation is designing the interiors of time machines.
The reason I'd want to see a vision-only mappless demo is because Lidar and HD maps are a crutch, which make it easy to create a 90% functional demo, yet are insufficient to cover the long tail end, which is inherently a computer vision problem. There is not enough information density or semantics in Lidar data to make high level judgements, which makes it useless for many relevant corner cases. Meanwhile, vision has all the information but it's a hard yet necessary AI problem to extract it.
The goal is a working product indeed, but if it involves teleoperation, human annotation for maps, and over-relies on Lidar, it will be a slow and unreliable AI granny driver, which will penetrate at most 1% of the market. There is a whole lot of unscalable human effort being swept under the rug in most autonomous demos, so don't believe everything you read in the media.
Yeah. I'm saying there are plenty of companies capable of this particular technical feat given the offline HD mapping + Lidar, so it doesn't justify the amount invested in and valuation of Zoox. This technology was available years ago. If say, they did this with vision only and no HD maps, maybe then I'd be impressed.
Meh. It's a demo. Doesn't prove reliability. With an HD map, slow speed and plenty of Lidar sensors, it's like a car on rails. Doesn't tell you what happens at the very long tail end, for instance when that HD map is not correct, making eye contact with a pedestrian is required to infer intent, or distinguishing a tumbleweed from a rock is required to make a driving decision at high speed. Fully autonomous driving may be AI-complete for all we know, and these one off demos are child's play in comparison.
Indeed. They're able to raise lots of cash with presentation skills, but in reality their business plan is completely bottlenecked on super reliable perception/intent prediction AI, on which they aren't showing evidence of an advantage.
How do we test human readiness for driving? We give them a driving test, which takes all of 40 minutes. This is possible only because we know a lot about what it means to be a human. To test self-driving technology we need to understand a lot about the AI that is behind it, which is an active area of research with much progress. Moreover, there is adversarial testing in simulation and on raw sensor data, which can tell you a lot about your system before any large scale real world testing begins, including when it might be safe to start real-world testing. Real-world testing at such large scales is neither a practical nor a necessary component to validating self-driving systems.
I think the counterintuitive aspect of problem solving at that level is how nondeterministic it is. It is no doubt that learning everything you can about a problem is critical to setting yourself up for being able to solve it, but this can happen a lot earlier than when someone burns out on a problem.
Are you confident the majority of the gains weren't made in the first year? It's exactly my point that knowing when to give it a rest is hard to estimate, and those who work on these problems professionally are likely to just keep pushing, potentially just spinning their wheels. Interestingly, the brain keeps working on these problems subconsciously, so you can actually get more done if you work on more things and tend to give up a bit earlier. I've solved hard problems before in an afternoon, a year after being hopelessly stuck on them, and my sincere estimate is that I would have made exactly zero progress had I kept pushing.
Is that a generic claim? I'm making a specific, counterintuitive, and empirical claim about productivity and work-style of mathematicians, having been on both sides. Even for professional mathematicians, the solution to a problem can come all at once, while one is not even actively pursuing it, which begs the question of the purpose of beating your head against the wall when doing math. For just one example amongst many, see the story of Yitang Zhang, who made his breakthrough in establishing the first finite bound on gaps between prime numbers, while out at a barbeque and having largely given up on the problem after two years of effort.
In my experience, being an amateur mathematician is more fun that doing it professionally, and potentially just as productive. You don't get nearly as much time for math, so you make the most of it and work only on the most interesting problems as opposed to just writing another paper. There's no time for beating your head against the wall, so you just do it when you're inspired, which is how problems get solved anyway.
relatively mundane when subjected to appropriate representations.
You omitted the important part of my statement. And the difficulty humans encounter in reasoning about the nature of abstract thought itself is something I also pointed out. My point is that this is not necessarily an indication of inherent difficulty. The ability to model and explain the nature of a self-reflecting mechanism is not a guaranteed natural ability of that mechanism itself.