Recommended reading for a bit of insight into the mechanics of the remote operations. E.g:
"Fleet response and the Waymo Driver primarily communicate through questions and answers. For example, suppose a Waymo AV approaches a construction site with an atypical cone configuration indicating a lane shift or close. In that case, the Waymo Driver might contact a fleet response agent to confirm which lane the cones intend to close."
> For years, companies like Waymo (owned by Alphabet, Google’s parent company) and Cruise (owned by General Motors) avoided any mention of the remote assistance they provided their self-driving cars.
> That is just how things work in Silicon Valley. By creating the illusion of complete autonomy...
> While Zoox and other companies have started to reveal how humans intervene to help driverless cars, none of the companies have disclosed how many remote-assistance technicians they employ or how much it all costs. Zoox’s command center holds about three dozen people who oversee what appears to be a small number of driverless cars — two in Foster City and several more in Las Vegas — as well as a fleet of about 200 test cars that each still have a driver behind the steering wheel.
> ...the [Cruise] cars were supported by about 1.5 workers per vehicle
> Waymo and Cruise declined to comment for this story.
I don't know what to say. It's an "illusion of complete autonomy."
The energy of these replies is the same as, "When a parent does a kid's homework, it isn't cheating." The difference is we still have no idea if the child will grow up into an adult, and it's not guaranteed that Waymo will be the one to crack real autonomy, even if it looks that way today. Indeed the longer it takes them the more likely it's not going to be them.
That's garbage. Waymo has been open about remote assistance ever since they deployed in Chandler 6 years ago. They've even given media interviews on how they use it. Other companies have had YouTube videos on it for years: https://www.youtube.com/watch?v=NKQHuutVx78
There's no illusion of any kind. We know how remote assistance works in this industry.
I don't know dude. Simple question: is it fully autonomous? No. If you step inside the vehicle, does it feel that way? Yes. That is nearly the definition of an illusion.
Yes. If you understood how remote assistance works, you'd agree.
There are no autonomous systems that doesn't require some form on human help. The question is can the vehicle do all the driving tasks and when it does fail, it does so gracefully and achieves "minimal risk condition" on its own (because, of course, no one can intervene remotely in real time). Waymo certainly does this.
Nothing will be completely autonomous and no reasonable person expects that. You will anyway need some degree of human intervention in order to achieve robustness. The important thing is that you don't need to take care of driving and spend the time in a more useful way. If you're insisting that everything needs to be fully machine driven and zero human intervention is required, you're effectively moving the goalpost just for your own argument.
There's a world of difference between "driven remotely", which was your claim, and "autonomously operated, with a remote driver available to help, in case of exceptions and emergencies", which is reality.
Just like there's a world of difference between doing your kid's homework, which is viewed as unacceptable, and helping your kid with their homework, which is viewed as good parenting.
> autonomously operated, with a remote driver available to step in, in case of exceptions and emergencies
Listen to yourself...
Anyway, you're getting it. I guess we have to be there and see for ourselves how the parent was involved in the kid's homework, right? If Waymo has nothing to hide, by all means, let in the press to their remote driving center. What are they going to do, steal the source code?
What even is your allegation? That someone has a joystick in a remote center and they're directly controlling the vehicle? And they are doing it all in real-time over cellular networks?
This isn't remotely feasible unless you think Waymo has invented new concepts in physics.
Small point of clarification. It's not feasible to do it safely. It's absolutely feasible to do it well enough to provide the illusion of safety though. There's various companies (Nuro, Vay.io) teledriving vehicles on US public roads right now. Waymo and others simply don't do it because of the obvious safety issues.
None of your links support your claim that "the vehicles are essentially driven remotely". They say that remote operators in rare cases need to provide guidance:
> the technician can send the car a new route to follow around the construction zone
Not being completely autonomous 100% of the time doesn't mean it's essentially driven remotely. Driven remotely means driven remotely. Not someone interjecting at one point on a ride, which seems to be what the articles are referring to.
So, pick a lane, you can't have both. Either someone is driving the vehicle, or there are support staff capable of intervening at points. And frankly, I would be horrified to think the second one wasn't happening. That would be moronic.
Edit: Oh, I see you doing a lot of goal post shifting in other comments replying to this. Good luck, driven remotely.
I suppose that if the vehicle will crash or hurt people without human remote operation of some kind, that's driven remotely in essence, no?
If there is more than 1.0 people running the vehicle, that's not good either.
I like Waymo! I just think that they aren't the only ones to figure out autonomous driving, and the crazy thing is that maybe they are specifically not trying to figure out autonomous driving, and that is maybe why only they have found success?
I think you're not really understanding what people are telling you: the vehicles do not depend on remote operators to avoid crashing or hurting people. The remote operators do not drive the cars in real time.
You do realize that your only source for this claim is an article that only mentions Waymo twice: once to say that Waymo had no comment, and once to blatantly lie about the remote operators being some kind of a shocking revelation rather than something they'd been open about for years.
I'm not stupid. I know that there isn't a steering wheel in a building somewhere with a camera feed to a Waymo. The Waymo cars are autonomously driving a lot of the time, in the sense of the car is being steered and going places using computers alone. "Drive the cars" and "real time" are words with specific meanings.
This isn't complicated though. The car phones home for many reasons to have a human being make a decision, using whatever non-steering wheel interface, but an interface nonetheless. That is not autonomous driving. The Waymo is somewhere between a human driver and a people mover at SFO.
It is certainly a scientific triumph that the car can distill the decisions or whatever. We don't know what kind of decisions humans are being asked to make or how frequently. We can make some metrics, like how many people per vehicle or whatever, but I'm not an expert. I'm not going to be able to fart out a metric. I don't really trust a company that needs a bajillion dollars in capital to achieve its goals on its metrics, but of course, I would invest in Waymo - I would invest in all the self driving car companies, because if that's my job and I like making money, that's the best strategy!
So many questions that people can pose, many metrics, they do not get to the core of the meaning of autonomous driving. One of the strongest analogies I've heard is that someone was comparing Waymo's remote driving to Google Maps. Here's the thing - would you be as excited about Google Maps as a technology if there were a human being planning the journeys and sending them to you? Very different science.
I'm not sure why this is so controversial. I am really excited about Waymo and autonomous vehicles. It's more that the reality is closer to tourist attraction than triumph of science.
> the vehicles do not depend on remote operators to avoid crashing or hurting people
Then why do they have any remote operators at all? Don't tell me, to resolve rare problems or whatever, or other speculation. I mean strategically or intellectually, surely, the simple answer is that the cars are not capable of day to day navigation using only computer-made decisions 100% of the time. That isn't controversial. We don't have visibility into specifically what kinds of decisions humans are making, so you shouldn't make such strong absolute statements, but it stands to reason that considering how essential everyone in autonomous vehicles is saying human-in-the-loop operations are, the purpose is pretty important. I would assume that an important purpose is safety! This isn't complicated!
Waymo could clear this all up in an afternoon. Let a journalist visit their remote operations center, and show them a screenshot of the interface. Zoox did! It doesn't take a genius to understand that the illusion and excitement of autonomous driving is so important to their perception of being ahead, Waymo isn't going to do that.
I think the reality is that their remote operations are better than everyone else's, not necessarily the autonomous parts of their driving, and that's why they're operating a taxi service and Cruise isn't anymore.
Maybe the secret to all of this is very effective remote operations. That would also be a really big deal.
> Then why do they have any remote operators at all? Don't tell me, to resolve rare problems or whatever, or other speculation.
I mean, I don't know for sure what exact tasks they do and what exact capabilities they have. If I did know for sure, I wouldn't tell. But given you're doing nothing but speculating, demanding me to not speculate seems unreasonable.
So, let's speculate, but still be more concrete than "resolve rare problems". Let's say that a Waymo can't plot a path it believes is safe, for whatever reason. There could be hundreds of different reasons for that. Maybe it doesn't think there's enough room to get through somewhere. Maybe it can't identify an object. Maybe a sensor stopped returning data. Maybe it's surrounded by an angry mob of luddites looking to torch the car. Maybe the passenger isn't leaving the car at the end of the trip, because they're black-out drunk.
The car can't proceed on its own in these cases. If moving, it'll stop safely, and wait for the human operator to resolve the situation. Maybe they'll annotate the map, or give the car a hint on a possible route, or tell the car to hang tight and it'll be picked up by a service crew (while they order the passenger a replacement car), or route the car with the uncooperative passenger to a {hub,hospital,police station}. In none of these cases is the alternative to a remote operator "to crash or harm people", which you claimed. It's that the car won't move.
The reason the car will stop when it needs help is that a remote human simply can't resolve a hazard in time. Communications aren't reliable enough, the latency is too high, and it's unlikely that a remote operator would have enough context to make a safety critical decision in a split second.
If you knew from the start the cars are not being driven remotely, why claim there's a "remote driver"? If you knew this wasn't realtime remote driving, why claim the remote operation is safety critical?
> I'm not sure why this is so controversial.
Well, you're making an extraordinary statement and providing no evidence for it. Like, literally your only source was about a different company. And when called out on it, you move the goal posts, or pretend that words mean what you want them to mean. That kind of thing is catnip to internet forum posters.
I've built one of these based on the instructions and use it daily (and love it!), but the tone of the site is pretty reflective of the project overall. It's definitely a really impressive hack and I appreciate all the hard work that's gone into it, but could really use a little more user-facing empathy.
I'm not sure if I'd recommend it to someone else - and if I was doing it again I'd probably spend a few hundred more (than the gaggia + parts cost) and just buy an off-the-shelf machine with the same feature set.
They recently updated it with an official vendor for a [PCB kit](https://www.peakcoffee.cc/product/gaggiuino-v3-kit-set-gaggi...) that (I believe) includes all of the internal components needed. I did the lego build w/ tons of soldering and ordering parts off Aliexpress. This new kit should make it much easier for newcomers. You still need to know what you're doing and it's not for everyone. A machine with similar capabilities is something like the Decent at $4k.
I'm curious about what off-the-shelf machines you would compare a fully modified gagguino. Granted, I don't have one so I'm only going off the feature set listed on the page, but short of a Decent ($$$$) I can't think of any competitors with features like flow profiles.
Well, there's no cheap competitors with (automatic) flow profiles, but there are competitors with automatic flow profiles. ACS Vesuvius, Rocket R nine one (although IIRC this one is actually pressure profiling, not flow profiling), Synesso es.1, Sanremo You to mention some.
If you're looking at manual flow control, basically any E61 group machine can be outfitted with a needle valve for flow control, and a bunch of machines come with one installed from the factory, like the Lelit Bianca. There's also machines like the Slayer 1 group and La Marzocco GS/3 MP.
Gagguino is very cool, and if you're just pulling espressos and not too many back to back, it should be plenty of machine. If you're doing more than a couple of milk drinks, or pull lots of shots back to back, or want to connect it to plumbing, it's not as great of a machine. The fact that you started out with a small single boiler with a vibration pump starts to show.
I think the grandparent is probably referring to upstate in the Hudson Valley sense of the phrase. Plenty of cute towns, but Hudson / Woodstock / Kingston are definitely in the O(tens) of great restaurant options, comparable to a slice of any single Manhattan / Brooklyn neighborhood most tech people live in.
The other answers cover the math well, but I think the “why do you need attention?” statement is worth making (and answers the more engineering-y question of “how/when?”):
DNNs typically operate on fixed-size tensors (often with a variable batch size, which you can safely ignore). In order to incorporate a non-fixed size tensor, you need some way of converting it into a fixed size. For example, processing a sentence of variable length into a single prediction value. You have many choices for methods of combining the tensors from each token in the sentence - max, min, mean, median, sum, etc etc. Attention is a weighted mean, where the weights are computed based on a query, key, and value. The query might represent something you know about the sentence or the context (“this is a sentence from a toaster review”), the key represents something you know about each token (“this is the word embedding tensor”), and the value is the tensor you want to use for the weighted mean.
This is great - the like / heart button is really slick, and I love how it doesn't get in the way at all. I've used pinboard and others in the past, and the (relatively) heavier bookmarking flow would often stop me from saving things as I didn't want to break my flow.
I agree with you (and love your blog, btw), but I think you're skipping over at least a few benefits you can get out of a mature / well built a/b framework that are hard to build into a bandit approach. The biggest one I've found personally useful is days-in analysis; for example, quantifying the impact of a signup-time experiment on one-week retention. This doesn't really apply to learning ranking functions or other transactional (short-feedback loop) optimization.
That being said, building a "proper" a/b harness is really hard and will be a constant source of bugs / FUD around decision-making (don't believe me? try running an a/a experiment and see how many false positives you get). I've personally built a dead-simple bandit system when starting greenfield and would recommend the same to anyone else.
Probably worth mentioning that the Google Content Experiments framework is in the process of being replaced with Google Optimize (currently in a private beta) which does NOT make use of multi-armed bandits much to my confusion and disappointment.
Huh. So do you know if they do anything help with repeat testing/peeking?
Optimizely takes an interesting approach: they apply repeat testing methods, segmenting the tests by user views of the results. Like 30x more complicated than multi-bandit, but they don't need a feedback mechanism.
This kind of language is very standard. The risks section pretty much always contains obvious platitudes ("An earthquake might destroy all our computers," "All our employees may quit").
Similar experience, but I focused on finding UI elements I liked from native apps or websites and attempted to clone them without looking at the source, then played around with the result to figure out how I could simplify it, how it behaved cross browser, etc etc.
