Recently, i've come to realize one real use of those clouds was to provide a good US-EU network connection. If you want to provide both continent users with correct bandwidth to your service, you have no choice but to have them connect to a datacenter on their own continent. Public data transit across the atlantic is simply miserable.
Then, because they probably have private atlantic cables, you can replicate at good reliable speed.
You don't have to be a researcher to do original work. Out of ten people on my team one has a PhD, but we're all doing more-or-less unique work a large fraction of the time.
Anyone knows whether there are fundamental differences in the approach waymo took vs tesla, or are they following the same algorithms and techniques, with more polishing or experience in the "small details that makes all the difference" ?
This an explanation from a layman on the auto side.
Tesla has focused on attempting to solve all cases using vision and generalize it as much as possible. The pro is you would be able to drop FSD capabilities anywhere in the US and it will work. The con is the tail of edge cases take significantly more effort then the first 80%.
Waymo instead of being generalized, works off of a detailed map for each region they drive in. They have a complete expectation of every detail on those roads and so then only need to account for the dynamic unknowns on the road. The pro is that they are getting to hands off driving sooner. The con is that anywhere its deployed needs detailed mapping and for that mapping to be kept up to date.
> Waymo instead of being generalized, works off of a detailed map for each region they drive in. They have a complete expectation of every detail on those roads and so then only need to account for the dynamic unknowns on the road.
This isn't entirely true. Everything at Waymo is built to be generalizable. Engineers and execs at Waymo have said to multiple times, as recently as yesterday [1].
They are able to drive just fine without up-to-date maps as well. It's an assumption built in to the system. The cars are also self-mapping. Ultimately, maps are just another input to the general driving software. The same driving software is deployed across all their cities.
A cache is exactly how I explain maps to people. They are saving precious compute cycles not computing the same static objects over and over again, by N number of vehicles.
A con does not have to be a flaw. But I would see it as a tradeoff. With all of these companies its hard to sometimes see beneath the curtain of what they promote.
They test in many cities (Buffalo, Tahoe, Bellevue, NYC) . But they won't deploy to any cities which they haven't mapped because they consider it critical to safety.
So the answer is no and that's why I outline it as a potential con. They are unable to just drop deployment into a new city, they as a matter of history, map out the entire city and keep an active map of the driving territory. I have no idea who wins in this dog and pony show but I think its a valid potential con and a easy way to see some of the decision differences in companies.
Does Google Earth/Maps/StreetView not count as mapping? I was feeling certain that all that infrastructure and data was invested into Waymo. It would seem that Google had a huge advantage from running mapping/camera cars for years and years, before SDC were even a glimmer in their eyes.
The mapping that waymo uses is much more detailed than streetview, the biggest difference being that it includes lidar data. I wouldn't be surprised if they combined the two for some of their latest data updates, but the pre-existing streetview data is not obviously enough to give an advantage.
Obtaining a detailed map doesn't seem difficult. This is the same company that created Street View. E.g. SF has ~1000 miles of streets [1], that would only take 50 hours of paying for manual driving to map all of SF at 20mph.
Never said it was difficult but that their driving fleet relies on a lidar map of the region it serves. Its a definite trade-off so I would list it as a con but it does not mean one is better than the other.
And probably because of this strategy, by the time their system is successfully deployed in several cities, the technology to make fully self driving has become much more accessible and they can relatively easily add that incrementally.
The other pro is cost: cameras are much, much cheaper than LiDAR. Tesla is making the bet that they can close the performance gap between cameras and LiDAR faster than the cost of LiDAR will come down.
I wonder if this will ever be a significant factor. How much can a LiDAR setup cost? 4k? Maybe 2k if built in-house at car manufacturer scale in the long term? Is that significant when buying a whole car?
I’d advise you to ignore lay explanations of the space - outside of the industry most of the discourse about self driving cars is poisoned by Elon’s deceptive presentations and his followers who parrot what he says.
If you want a grounded explanation of how Tesla’s stack works, follow @greentheonly on twitter. He’s a Tesla reverse engineer who regularly posts about the software that’s actually running on the car.
If you want an explanation about how real AV companies stacks work, I’d read Sebastian Thrun’s robotics textbook - then imagine what’s outlined in that book but with ML plugged in to a ton of spaces throughout the stack. This is also similar to how Tesla’s stack works, btw - greens just good to follow because a lot of people refuse to believe Tesla isn’t running some kind of “LLM but for driving” fully end to end black box model.
Tesla won’t launch a robotaxi anytime soon because they can’t use remote support or HD maps - although I think they’ve been stepping up their mapping efforts. Even the demo at Universal studios a few weeks ago was HD mapped - per @greentheonlys twitter.
I worked in the space for years and have seen the internal of both a traditional robotaxi company’s stack and Tesla’s.
For reference, Sebastian Thrun led the Stanford team that won the Darpa (self driving car) Grand Challenge in 2005, and then joined Google to lead Waymo (then called the Google Self-Driving Car Project), among other accomplishments.
Tesla chose to recklessly endanger lives, killing at least 50 people by releasing an unsafe beta test. Their goal was to capture market share from Waymo.
This is the thing that drives me nuts: wantonly ignoring regulations designed to protect consumers, and claiming that it was going to save more people than it hurt, in the long run. Tesla's participation in the market has not and will not make self driving arrive earlier than if they had refrained from participating. All it has done is given them an opportunity to compete for market share, and engage in pump and dump scams by misrepresenting their progress to naive investors.
People will tell you they're fundamentally different, but they are in fact the same. First, there are two different, independent aspects: Perception and Action. Seeing the world, vs taking action in that world.
For perception Waymo uses more sensors than Tesla. It uses lidar to construct a 3D scene of what the car is seeing, while Tesla uses SLAM-like techniques with their cameras. What people are missing is that these are only a small part of the perception problem - lidar returns a monochromatic 3D scene so it cannot see labels, markings, read signs, lights etc. LIDAR simply doesn't carry the necessary information needed to navigate the world, thus it is a secondary source of info. Cameras in motion do carry all the information needed, so the big difference between them is only one part of the overall perception stack.
Once you have constructed a labeled, accurate 3D scene (whether by lidar or SLAM), "action" is the same and there is no difference between Tesla and Waymo here. They both have to learn how to drive safely using the same information, so it's going to be a lot like LLMs where there's difference between LLAMA/Claude/GPT, but they're also all kinda the same thing.
The fact that you don’t know the difference in capabilities between cameras and LiDAR means you shouldn’t really be commenting. LiDAR allows you to also ‘see’ things in conditions where weather is bad visibility is limited. Cameras cannot allow you to do this.
Going through this thread, it is mind blowing how people let their fanboyism for others talk nonsense. So many Elon lovers here. Can all those Elon lovers just jump in the backseat of a Tesla and turn on FSD (which is an oxymoron) and jump on a high way or something? The world would be a better place. I can’t wait to see their faces and a the drivel they will come up with next year when Elon announces delays and other bull**t fit why he can’t launch what he promised. It’s going to be hilarious.
Tesla is taking a minimalist approach to every important factor: Specialized vs unmodified production cars; back seat in-vehicle UI app vs app-only for sitting in back; multiple sensor types vs camera-only; intensive mapping of a service area vs crowdsourced; only Waymo-owned vehicles vs private cars seconded to the service.
The biggest small detail is that Waymo's expansion is gated on achieving performance goals like very little need for remote supervision because Google won't hire a building full of remote monitors. Tesla claims this will be possible "by next year."
Fundamentally different - people will flock to LiDAR vs Vision when debating this, but the more fundamental difference is that Waymo uses HD mapping remote support.
Thanks, i was expecting something beyond just the lidar thing.
Can you elaborate more about mapping remote ? I've never tested neither a tesla autopilot nor a waymo (european here).
edit : also, are there differences in the core algorithms ? Tesla seems to be full AI / ML. Is waymo the same ? ( as the company is older i wonder if they haven't built more things manually)
Neither of them are “full AI/ML”, they’re both traditional robotics systems with ML used for detection/prediction/planning at certain steps. Elon will sometimes say something about moving to a “new ML stack”, but Tesla reverse engineers regularly look inside of what’s running in the cars and that’s not the case at all.
Contrary to what other people in this thread are saying, the remote support isn’t remote direct driving of the car - essentially what will happen is that if the car finds itself in a situation where it’s unsure of how to proceed and it’s safe to stop, it will pause for a few seconds and wait for a remote operator to clarify a situation for it.
A good example might be road construction - if the car detects new road construction work that doesn’t match its map of the area, and its onboard systems determine that it’s not sure how to proceed through the construction with confidence, it will send what it thinks the top five likeliest ways to proceed to a remote operator. The operator then selects the proper path (or says that none of them are proper). The car will then follow the path presented by the operator, but actual driving behavior /collision detection / pathfinding is still determined locally. Think of it like ordering a unit around in StarCraft.
You can actually see this behavior in the car when it runs into a difficult situation. It tells you it's asking for assistance or something similar, and pauses for a few seconds.
I also made the mistake of assuming the remote operator drives the car but if you watch Waymo's technical videos, it's clear that the AI is in control of the car all all times and the remote operator is just doing near real time labelling of what the car is seeing.
Same way reverse engineer get access to internals of other devices - a bunch of tricks. :^)
In one case, greentheonly realized some fraction of Tesla’s cars are shipped out of the factory still in dev mode, with debug mode enabled and increased privileges. He found someone with a car like this who was down to helped and swapped part of his cars hardware with their car, and from then on was able to get a much better view of what was running on his car.
Unfortunately twitter is awful to search and a lot of his info is buried deep in old threads, but a few (old) examples to illustrate that he regularly does this.
It's a fallacy to really compare them directly. Waymo has active cars in service that are available to the general public. Tesla right now just has promises that they have been unable to deliver on for almost a decade now, as their approach is pretty radical.
It's a mile driven by the car, not by a human. An easy concept to understand, unless you don't want to understand it.
If you want to play the "ultimate goal" game, the ultimate goal is to do it all profitably at scale -- and Tesla is way ahead on that front, which is why their fleet self-drives almost as much per day as Waymo's fleet has ever driven.
Time will tell, but anyone who can't make a case for both "Tesla wins" and "Waymo wins" scenarios is a fanboy with deeply compromised thought processes.
If miles driven by cars under human supervision counted, Toyota cruise control would clock the highest. Autonomy is a binary: it’s either driverless or it’s not. There’s no need to invent terms such as “self driving mile”.
Tesla may be profitable, but nowhere close to a working solution. So how far ahead are they really?
Tesla's system is so fragile it needs a human ready to take over at any second to prevent a crash. Compared to Waymo Tesla FSD is like a kid using training wheels.
Touched on it to your original question but they require cars to drive through their service regions on a regular basis to keep a very detailed map of the environment. Not just the 2d map of a representation of the entire environment. On the remote part, Waymo will encounter trouble and they have support to take control of the vehicle remotely.
People get hung up on the vision stack and mapping. Obviously, any self-driving car needs a way to perceive the world around it, and Tesla’s approach is very different from Waymo.
The software that actually does the driving, ie turns the wheel and works the pedals, is very similar between the two companies. Tesla took a big step in Waymo’s direction earlier this year when they replaced their optimization based path planner with a neural network.
I would recommend watching this video, where the same YouTuber is using FSD which takes an exit and needs intervention to avoid running directly into a divider.
You seem to be stating this as some kind of gotcha.
1. Since that video, FSD 12.5.6.1 has been released, and v13 (which is what is rumored to have been used on Cybercab) is just around the corner. It is completely disingenuius to point to a video 7 months old (on FSD 12.3!) and insinuate that it is representative of the current experience.
2. In FSD, interventions are in your hands. With Waymo in the same situation, your fate lies solely in the remote operator watching your vehicle and how quickly they are able to react. FSD is obviously not perfect but the rate of interventions plummets with every new major release.
Your understanding of remote support is entirely wrong. At no point can a remote operator "drive" a Waymo. They confirm or change plans when the car gets stuck -- that's it.
If you look at the rate of progress over time you see a monotonically improving system that has no apparent halting of improvement. Likewise you have examples of competitors making progress as well in different areas each converging to a pretty, what appears to be, inevitable conclusion of full autonomy.
No one knows when that happens. But it feels pretty certain it’s happening.
I’ve been using FSD for 5 years now. It’s gone from glorified cruise control to something I generally don’t need to intervene with on city streets in that time. Will it improve that fast over the next five years? I doubt it. But it doesn’t have to because the residual problems are much fewer if harder. At this point, especially given the rate of AI improvement overall, I am confident in that five years those problems will largely if not entirely disappear.
Do I take a nap in the back seat? Of course not. Should it be marketed in its condition? I don’t know. But I do know the joint probability of me making a mistake as the attentive operator and it making a mistake while in control is significantly lower than either alone. The fact it makes mistakes at times is obviously concerning to me as a driver, and the fact I also make mistakes actually doesn’t concern me nearly as much as it should. However - I catch its mistakes, and it doesn’t make mine. Why is it rational to be more upset about the machine making a mistake than a human? It’s not - but humans are taught logic and are never rational.
But it's not a robotaxi. Even its level of sensor and compute redundancy is not ready to be a robotaxi. Nothing shown at the Cybercab event changed that. They go for form over reliability every time.
With HW3, they ate their redundant compute node because they underestimated the compute required for the task.
Now even with the redundant node utilized for non-redundant purposes, that doesn't seem to be enough as they are finally admitting HW3 will never not be "supervised".
And then the many years of lying about its upcoming readiness. There are websites out there where you can find all of Musk's quotes about it being just around the corner, or their current generation of vehicles all becoming money-making robotaxis with a little software update worldwide.
There's no indication at all they'll break out of the 100-120 miles per safety disengagement they currently sit at (community tracker, Tesla themselves doesn't publish reviewable safety data).
You being happy that your car can finally make zero intervention trips is NOT the standard necessary for taking the driver out of the seat.
These comments on that video perfectly capture my reaction:
> This is FUD. Who or what speeds up to 55 mph to enter a rotary? How is it that the posted limit and presumably map data indicates 55mp seconds before the rotary? What do you expect FSD which is training on humans using vision and the maps to do? I saw a dinky little rotary sign AT the rotary. I'd slam on the brakes or have an accident too.
> Why would the car come to a stop? I don't see a stop sign, and most roundabouts are yield and I don't see another car blocking your way. Why enter a roundabout at 55? You are wrong, not the car or FSD. You don't know the correct way to drive a roundabout.
FSD mimics human behavior. If you are speeding into a roundabout at 55 mph, you are the one in the wrong, not FSD. It's honestly kind of incredible the ridiculous lengths people go to try to discredit FSD.
That YouTube commenter you quote pretty clearly did not pay attention to the video.
> FSD mimics human behavior. If you are speeding into a roundabout at 55 mph, you are the one in the wrong, not FSD. It's honestly kind of incredible the ridiculous lengths people go to try to discredit FSD.
That's just rephrasing the YouTube comment. Try watching the video yourself. Particularly watch test #7.
Here's a summary:
• The car is on a highway, traveling at normal highway speed of 55 mph. There is no visual indication that there is a roundabout somewhere up ahead.
• After traveling ~3800 feet there is a sign that indicates a roundabout and says the roundabout speed is 15 mph. The roundabout is not yet visible.
• The car continues at highway speed past another sign ~600 feet past the first that also shows that there will be an upcoming roundabout. The road starts curving after that sign, and the roundabout starts coming into view ~600 feet further down the road.
• The car continues approaching the roundabout at highway speed until the human intervenes. He tried to give the car as much time as possible for FSD to decide to slow down. In some of the tests he waited long enough that when he did hit the brakes he had to brake very very aggressively to slow down to 15 mph before entering the roundabout.
Even if it does not have that roundabout in its map and did not read the signs so it is not expecting a roundabout there shouldn't is see it as a sharp bend in the road that should not be taken at highway speed and slow down?
> > This is FUD. Who or what speeds up to 55 mph to enter a rotary? How is it that the posted limit and presumably map data indicates 55mp seconds before the rotary? What do you expect FSD which is training on humans using vision and the maps to do? I saw a dinky little rotary sign AT the rotary. I'd slam on the brakes or have an accident too.
This is a laughable hot take. "seconds before".
Watching the video, it starts with him at 43mph, and he drives at 55mph for FORTY SECONDS before encountering the roundabout.
All these clowns saying "Oh, in the real world he'd have slowed down for that roundabout".
No. He wouldn't have started slowing down two-thirds of a mile away (40 seconds at 55mph). This is a garbage argument.
You're not going to be taking that roundabout at 50 mph. That why, significantly before the roundabout, there's a 15 mph roundabout warning that FSD completely ignores.
Almost crashing the car IS a gotcha for any vehicle purporting to be autonomous. Especially when Musk seems to be betting Tesla on it working vastly better than it currently is.
One is doing it completely driverless and has to get it right every single time. The other has a driver ready to intervene and just needs a single intervention-free drive.
The biggest difference is the presence of lidar in Waymo vehicles which means they have an accurate physical representation of the world vs the vision only based approach that Tesla is taking.
So you’re going to ignore all the data Tesla is collecting, and it’s purely AI approach?
Very biased comments here, I get it, everyone hates Elon, but let’s not lie to ourselves.
For all their faults, one thing you can't say about Google is they don't know how to scale. Prior to Google Maps, Street View would have seemed impossibly expensive, but now we take it for granted. If they need to do a LIDAR run with a car over every public road in the US for them to deploy it, it seems like they could just do that. All they'd have to do is add LIDAR sensors to the existing Street View imaging fleet and continue running them.
If on the chance Lidar is needed, which I know it's not because my car drives me in every type of situation I've thrown at it, then camera only data is still valuable because it gathers so many crazy edge cases.
Waymo's latest hardware uses 13 cameras, 4 LIDARs, 6 radars and at least one mic. Def not early-optimized. I have not found a BOM for the whole package. That's a significantly different idea of how much, and how varied, sensor input it takes.
to my knowledge, Tesla has gone the computer vision route where they are solely relying on cameras and algorithms, while Waymo went the way of more traditional LIDAR and other scanners to close achieve the safe full self drive.
The disadvantage of using the LIDAR and full sensor stack is largely price.
So does Tesla’s. I use it daily. From home going through a busy city, onto a major highway with rush hour traffic, into a downtown area to work. It can do this without me touching the wheel or pedal for the entire length of the drive. I have a hw4 S plaid and it’s made dramatic improvements over this last year. I’m blown away at how good it is (also blown away by waymo).
Well, we know exactly how Waymo's remote operators help out: https://waymo.com/blog/2024/05/fleet-response/. They can't prevent accidents in real time like the Tesla drivers do and can't "control" or "drive" the vehicles.
Tesla FSD is impressive for a driver assist system. But that's all it is — a driver assist. They need orders of magnitude improvement to match Waymo's performance and go driverless.
It still needs to be supervised for the edge cases, but the standard city roads and highways are a solved problem. I think some of the complex roads where you have to quickly cross two way traffic that doesn’t stop can be difficult, I don’t use fsd in that situation, it’s even hard for a human. Sometimes I’ll give it a nudge when it’s being too safe. There’s a construction area that I hit which would have caused the car to take a non optimal path, so I take over there on a regular basis, those issues do get fixed though. That’s about the only issues I have. It can now do things like drive down my long private unmapped driveway without issues.
My work is about 10 miles way in the Seattle area. I can go to and from with zero interventions until I get to my works parking garage
You kind of have it correct, but Tesla is using vision, AI, and huge amounts of data. It’s like the chat-gpt of autonomous driving.
The data is the most important part, to solve real world driving everywhere, you need huge amounts of data for all the edge cases. Tesla has millions of cars on the road gathering this data, vs a couple of thousand for Waymo
Data quantity is useless if the data is of low quality. You need to be able to judge the car's performance in simulations to guide training. Elon admitted in the latest quarterly this is a huge problem for Tesla -- they have to do many millions of miles of simulations to compare two models. Higher fidelity data would cut this number by many orders of magnitude.
You either didn't understand what Elon said, or are deliberately misinterpreting what he said - I listened to the earnings call myself. He said it's taking longer to train the models because the miles between interventions is getting so large that it takes a while to see which model is better when they're comparing different models. It's not a "huge problem", it's a good problem.
> Tesla has millions of cars on the road gathering this data, vs a couple of thousand for Waymo
Except that Alphabet has been mapping and scanning for years, since before Waymo. And, Waymo vehicles are on the road while waiting for a fare, so they can use that time for mapping, while Teslas are reliant on where their owners go.
Tesla thinks they can do everything with nothing but cameras. Everyone else is using multiple sensors to give the computer additional info to verify what the vision is sensing. At this point it seems extremely unlikely Tesla will ever produce fsd with vision only. Probably why they added radar back into the model s refresh. It is currently not enabled afaik.
I find it fascinating the crew of Tesla fanboys on this site that run around downvoting anyone who isn’t drinking the koolade, but never have anything factual or constructive to add to the conversation.
There is literally no indication at all that Tesla is right and every other expert in the industry is wrong about vision only never working. We’re a decade in and they still can’t prevent phantom braking that a 2010 ford Taurus didn’t suffer from.
There isn't a big story to this. Does Waymo share details about how and when its vehicles are driven remotely? Since the answer is no, you can only assume that it is pretty often.
Really kudos to Zoox for turning a profile about them into this unveiling of Waymo. I fell for Waymo's sleight of hand. Aspirationally, we really do want self driving cars.
Occasionally there are problems of being stuck (at 0 mph) due to a lack of aggressiveness or the refusal to violate a rule, or being in a loop the car can't figure a way out of. Often times this is just a safety precaution, at this early stage, of wanting a human in the loop to confirm a plan, but sometimes it's because the system isn't robust enough to properly make the plan and I'm sure that's where a lot of the development effort is currently pointed -- reducing these events to as close to zero as possible. That's the long tail. It never really quite goes away.
They aren't. The car asks the remote operator about its options, the operator suggests an option, the car proceeds. Nobody is driving in realtime over LTE, that would be insane.
> Nobody is driving in realtime over LTE, that would be insane.
To be pedantic that's not really a good argument against the possibility of remote driving. All of Waymo's service areas are inside the coverage areas of AT&T's 5G , Verizon's 5G Ultra Wide, and T-Mobile's 5G Ultra Capacity networks.
Just because they figured out an efficient way for the cars to be driven remotely doesn't mean that the cars are not being driven remotely. This ventures a bit into a subjective sense of what is self driving, and clearly, Waymos overstate their self driving ability through a variety of very clever, very powerful sleights of hand.
Waymo shared that their disengagement rate is 17k miles per disengagement meaning they are driven by human test drivers once per 17k miles, across all causes. It doesn't include assists by humans because that's not driving: it's merely remote humans answering a question about where to drive in case of a novel obstacle on the road. You'd be nuts to think remote humans actually drive: the internet infrastructure is not designed for guaranteed low latency.
I also wouldn't trust NYT reporting on anything related to Big Tech: they have a well-documented pattern of anti-big-tech biased reporting.
In this comment, you are saying the verbatim words "where to drive" to mean the navigation that Google Maps does, but then you used the verbatim words "where to drive" to mean resolving questions about obstacles. So I guess, yes, resolving questions about obstacles is indeed remote driving, and navigation questions from Google Maps is not remote driving. But everyone already knows that.
The sort of obvious definition of self driving means no human intervention whatsoever, which Waymos also fail. So I don't know. Why doesn't Waymo use Google Maps instead of people to tell the car, "where to drive"?
Very simple. Because there are real-time conditions on the road not known by Google Maps. Google Maps can tell me to turn right at an intersection, but I can see that police cars have blocked that road so I decide to proceed straight. Same thing with Waymo; its routing indicates turning right but it finds obstacles on the right. So it asks a remote human what to do. Both of these are "where to drive" questions. It's just that their answers come from different data sources, and one is easier to answer than the other.
The disengagement rate is calculated when a human test driver is in the vehicle and takes over driving. It's a specific term from autonomous vehicle testing. If a self-driving vehicle has no human drivers in it and it does not know how to drive, it simply pulls over if possible and stops. That's not among the definition of disengagement.
This has happened multiple times in San Francisco. You can find videos of driverless Waymo cars getting stuck. You can also find a hilarious instance where a stuck Waymo was driven by the emergency personnel after it became stuck.
i'm at my point in my career (20 years xp) where i think every project should be built around pure functions and structs, sorted in modules. And only once you're sure there absolutely no other choices, add a bit of interfaces, class and inheritance. Which, imho should happen extremely rarely.
I've come to realize that the amount of useless abstractions we add just because the language lets us, instead of thinking more deeply about what exactly is the problem we're facing, is just insane.
Sidenote : i recently tried cursor, in "compose" mode, starting a fullstack project from scratch, and i'm stupefied by the result.
Do people in the software community realize how much the industry is going to totally transform in the next 5 years ? I can't imagine people actually typing code by hand anymore by that time.
Yes, people realize this. We've already had several waves of reaction - mostly settling on "the process of software engineering has always been about design, communication, and collaboration - the actual act of poking keys to enter code into a machine is just an unfortunate necessity for the Real Work"
I think all of those of us who are paying attention expect it to change drastically. Its just how I don't know (I accept "there will be nothing like software development" among the outcome space), so I am trying to position myself to take advantage of the fallout, where ever it may land.
But I also note that all the examples I have seen are with relatively simple projects started from scratch (on the one hand it is out of this world wild that it works at all), whereas most software development is adding features/fix bugs in already existing code. Code that often blows out the context window of most LLMs.
> I can't imagine people actually typing code by hand anymore by that time.
I can 100% imagine this. What I suspect developers will do in the future is become more proficient at deciding when to type code and when to type a prompt.
Yes, I tried it, too, and while
impressive, it still sucks for everything.
For the industry to totally transform it has to have the same exponential improvements as it has had in the past two years, and there are no signs that this will happen
At the moment the model companies aren't really focussing on coding though. There's a lot of low hanging fruit in that space for making coding AI a lot better.
i've had a first attempt, which was very mediocre ( lots of bugs or things not working at all), then i gave it a second try using a different technique, working with it more like i would work with a junior dev, and slowly iterating on the features... And boy the results were just insane.
I'm not sure yet if it can work as well with a large number of files, i should see that in a week. But for sure, this seems to be only a matter of scale now.
You're assuming we'll see the same exponential improvements as it has had in the past two years, and there are no signs that this will happen
> The thing is, once you're used to that kind of productivity, you can't come back.
Somehow everyone who sees "amazing unbelievable productivity gains" assumes that their experience is the only true experience, and whoever says otherwise lies or doesn't have the skills or whatever.
I've tried it with Swift and Elixir. I didn't see any type of "this kind of productivity" for several reasons:
- one you actually mentioned: "working with it more like i would work with a junior dev, and slowly iterating on the features"
It's an eager junior with no understanding of anything. "Slowly iterating on features" does not scream "this kind of productivity"
- it's a token prediction machine limited by it's undocumented and unknowable training set.
So if most of its data comes from 2022, it will keep predicting tokens from that time even if it's no longer valid, or deprecated, or superseded by better approaches. I gave up trying to fix its invalid and or deprecated output for a particular part of code after 4 attempts, and just rewrote it myself.
These systems are barely capable of outputting well-known boilerplate code. Much less "this kind of productivity" for whatever it means
What you describe was my experience (with swift code too, on mobile). Until i tried it with web dev. Then maybe it’s due to the popularity of web tech compared to swift, i don’t know ( I should try it with react native to see), but there is absolute no doubt in my mind the time it took to build my website is 10 or 100 times faster ( 2 hours for something that could have taken me a week).
It’s easy coming up with the first version of a web app, especially if you have a mockup. There’s a lot of css and JS frameworks because of how common the use cases are and how easy it is to start solving them. It’s the iteration that sucks. Browser mismatch, difference between mobile and desktops, tools and libraries deprecation,… that’s why you take lot of care in the beginning so you don’t end up in a tar pit.
“starting a full stack project from scratch” - that’s just it, i’ve found AI tools to be great at starting new projects. Using it for a large existing project or a project that has many internal company dependencies is…disappointing.
The world isn’t just startups with brand new code. I agree it’s going to have a big impact though.
I do relatively niche stuff (mostly game development with unity) and I've found it very capable, even for relatively complex tasks that I under-explain with short prompts.
I'm using Claude sonnet 3.5 with cursor. This week I got it to:
- Modify a messy and very big file which managed a tree structure of in-game platforms. I got it to convert the tree to a linked list. In one attempt it found all the places in the code that needed editing and made the necessary changes.
- I had a player character which used a thruster based movement system (hold a key down to go up continuously). I asked the ai to convert it to a jump based system (press the key for a much shorter amount of time to quickly integrate a powerful upward physics force). The existing code was total spaghetti, but it was able to interpret the nuances of my prompt and implement it correctly in one attempt
- Generate multiple semi-complex shader lab shaders. It was able to correctly interpret and implement instructions like "tile this sprite in a cascading grid pattern across the screen and apply a rainbow color to it based on the screen x position and time".
- generating debug menus and systems from scratch. I can say things like "add a button to this menu which gives the player all perks and makes them invincible". More often then not it immediately knows which global systems it has to call and how to set things up to make it work first go. If it doesn't work first attempt, the generated code is generally not far off
- generating perks themselves - I can say things like "give me a list of possible abilities for this game and attempt implementing them". 80% of its perk ideas were stupid, but some were plausible and fit within the existing game design. It was able to do about 50%-70% of the work required to implement the perk on its own.
- in general, the auto complete functionality when writing code is very good. 90% of the time I just have to press tab and cursor will vomit up the exact chunk of code I was about to type.
i generated the project, then added features, which meant adding new tables , forms, api endoints, navigation. Then asked for subtle changes in the way the fields were edited.
At one point i asked it to "make the homepage look a bit more professional", and it did.
I can do all this in my sleep. Except "a bit more professional" as I suck at design.
I could do all this in my sleep when I was in my second year of career, and now I'm in my 24th year (god, I'm old).
What you described isn't just easy, it's trivial, and extremely boilerplate-y. That's why these automated token prediction machines are reasonably good at it.
i think we’re not talking about the same thing. I’m not saying it’s hard for a experienced software dev. I’m saying it requires a level of skill that is on par with a professional software developer. Meaning this system can already replace a huge chunk of the jobs in the industry.
What? Somehow people in the west getting arrested for protesting against genocide, is acceptable/redeemable just because another country is presumably worse??? Do you actually care about freedom/human rights at all or are you just abusing those concepts to feel superior relative to other countries and to play petty tribal us-vs-them politics?
Usually not arrested for protesting. It usually for trespassing or vandalisation or something else. Or sometimes something stupid like blocking traffic
When $NON_WESTERN_COUNTRY arrests people for, let's say, "picking quarrels", people here usually see that as just an excuse for cracking down on activities that go against government interests. But when western countries do the same, people take those excuses at face value???
didn't say it was ok to arrest people protesting. Just that it's nothing comparable with non-democratic countries like china or russia. You can protest in front of the white house. Just don't try that in moscow or beijing.
You can do stuff in front of the White House as long as you're a nobody and it doesn't actually threathen powerful interests. How many of these protests where nobody was arrested, actually resulted in change? What's the point of protesting if nothing ever changes? Thousands of children have already been killed in Gaza, but we can protest in front of the White House while the killing continues, and that makes us better than $OTHERCOUNTRY. Why are you content with a circus like that? What's even the point of comparing with $OTHERCOUNTRY if you don't have effective power to change for the better at home? Shouldn't we set higher standards for ourselves?
I think it's really, really weird that some people care more about how bad $OTHER is than problems at home.
genuine question (the maths lost me): this seems like a philosophical problem. Did the writing of that dilemma into mathematical language bring any interesting result, compared to using regular english ?
Does jamming over the internet requires special techniques on the musician side ? Whenever i played with my DAW, any kind of hardware latency made it impossible to play correctly. I can’t imagine how that would work with two or more people over the net.
Jamming online in real-time is possible only between participants in the same city or district. It all boils down to internet-induced latency. More than 20ms and no real-time is possible. For long distance (high latency) online jamming, there is Ninjam [1], developed by Cockos, creators of Reaper DAW, where latency is actually expanded to match a musical division (bar), so everyone can jam in non-realtime but still musically correct.
For long distance collaboration and recording (not jamming), I'm currently working on developing a Reaper Sonobus (which uses AOO) solution which I named ReaConnect [0].
With good hardware and a stable system you should be able to get the harware latency down to 2-5 ms. Then everybody needs a very stable and fast internet connection with low jitter, so you get average ping times of 20 ms or less. On top of that you have a jitter buffer latency of maybe 20 ms or less. (If you're very far apart geographically, you also need to consider the speed of light for the network transmission.) Under ideal circumstances you may end up with a total end-to-end latency of 30 ms. This is the same latency you get when standing 10 meters apart on a large stage. You probably won't be able to play bebop or technical death metal, but pop ballads or stoner rock should be feasable :)
It's coming along.
I don't think anyone is using it for anything serious yet, but it is starting to feel like a real language.
My guess is that it will start being used as a library language (i.e. have libraries written in Mojo being called from Python) before it really gets going as its own thing.
Then, because they probably have private atlantic cables, you can replicate at good reliable speed.
reply