jowday

Every company knows this, the problem is that you can’t fit the compute you need for AR/VR + AI into a form factor like this, at least not right now.

Also, completely anecdotal, but the only people I’ve ever seen wearing them are Facebook employees.

If anything, what I’ve heard from friends at or adjacent to Meta is that they’re paring back metaverse ambitions and capabilities on their future devices because of the success of the Ray-Bans glasses and the relatively middling sales of the Quest 3.

The Ray-Bans are also a weird anomaly since they’re leaning on the Ray-Bans designer pricing to justify a lot of the cost. If you’re already buying a pair of luxury sunglasses that cost close to $200, what’s another $100 to get the smart version?

Dude your kid is fine. He’s going to a top tier college for his program and making a B average after his first year. Maybe he’s playing too much Roblox? Is he happy, healthy? Does he have a social life? If so, not a big deal in the grand scheme of things. I don’t even check the GPA of new grads I hire.

Edit: If it’s any consolation, I did know people who were truly addicted to gaming in college. They couldn’t manage their time and were on academic probation by the end of freshman year. Slept through multiple finals because they were up late playing league of legends. Never got any internships. Your kid doesn’t sound like the people I knew who really couldn’t manage their time. Also, Freshman year is generally academically the rockiest. A 3.0 is fine. The kids I knew who really couldn’t cut it were failing by the end of their first semester. And this is at a far less competitive program than where your kid is at.

What do you mean by “not great” grades? Failing, or not 4.0?

It’s honestly a little weird that you’re this into your child’s business after they’ve gone off to college, especially since this doesn’t even seem like a major crises (drug or alcohol problem, major psychiatric crisis) that I saw a ton of people go through in college.

Assuming they’re not failing, that they even got an internship after their first year shows a ton of initiative. I wouldn’t be too worried.

It does sound like you might be a helicopter parent though. In that case, did they get this internship largely of their own initiative or is it something you pushed them into?

I think it’s pretty clear that the OP meant that a downside of this is that you need a GPU to run it quickly locally, not that it being local first is a bad thing.

Out of curiosity: did the name change because of issues with Blizzard?

There's a lot of issues with it, but perhaps the biggest is that there aren't just troves of easily scrapable and digestible 3D models lying around on the internet to train on top of like we have with text, images, and video.

Almost all of the generative 3D models you see are actually generative image models that essentially (very crude simplification) perform something like photogrammetry to generate a 3D model - 'does this 3D object, rendered from 25 different views, match the text prompt as evaluated by this model trained on text-image pairs'?

This is a shitty way to generate 3D models, and it's why they almost all look kind of malformed.

I gotta say, your responses in this thread more or less typify the interactions with Vercel devrel that I complained about in my parent comment. Every time I, or someone I know, complains about Next or try to raise an issue on social media, you or someone else at Vercel shows up asking for more info. If we give you any info we get a polite "Ah thanks for letting us know, yeah that sucks!", and nothing is done, as far as we can see, to address the issues we brought up or course correct Next's development.

Hard to say specifically, especially because I spent about a month trying to deal with this when we first migrated from Next 12 to 13 - but it essentially seemed that if you were using any web tech that wasn't built with SSR in mind, there were issues that were unaddressed in the documentation at the time.

Someone else in these comments mentioned:

- Fuck spending 3 hours working out why you’re not able to use relative image paths in MDX files and have to shove everything in /public.

- Fuck fighting five layers of configuration and bundlers and libraries and GitHub issues to try and load a WASM file without having the whole thing break.

These were both issues we had to deal with, and similarly lost many hours to trying to resolve.

At one point you guys seemed to have shipped a release that, for some large fraction of users, caused the devserver to start OOMing rapidly. When this started happening to me, I spent about a day trying to debug it, since I assumed that such an a huge and imminent issue would arise from my mistake, not the web framework's. I eventually texted a friend of mine about the problem, only for him to tell me that they had been having the same issue, and linked to this thread.

https://github.com/vercel/next.js/issues/54708

Working with a web framework that's so unstable that fundamental features like the devserver break periodically isn't fun. I stop trusting the web framework, and when a bug arises I find myself having to check both my own code and Next's. Now this applies to some degree with any open source framework, of course - there's bugs in everything. But I have so little trust for Next at this point. We've collectively lost weeks of eng hours to just dealing with quirks of Next.

Right now we're really affected by the fact that the recommended way to cache database connection objects is broken in our repo (and some other users' repos, evidently).

https://github.com/vercel/next.js/issues/47099

This is another example of something where I naively assumed that the error was on my end, and spent days trying to debug our database configuration, hosting provider, etc, before realizing that at some point something broke and the recommended way of persisting these connections on the global object doesn't work anymore. What's the point of using a framework if I have to question whether it's working correctly every time an error arises?

We've had an awful experience with Next. There are undocumented, broken things all over the place. Migrating to the App Router was a huge mistake. 'use client' didn't work as expected, and we had to install workarounds and try out different hacks to get access to the most pedestrian features we could expect from any other web framework.

This is the broad consensus among everyone else I know using Next - they felt like the Next 13 release and the app router was a rug pull from a previously pleasant experience.

And whenever you complain about this or ask for support, some devrel guy from Vercel shows up in your replies, saying something to the effect of "Wow! That's crazy, we worked really hard on Next 13! This is my first time hearing about this!"

Edit: Since I know the devrel guys are in this thread: If you want people to keep using Next, something fundamental about how you guys write the framework needs to change. The instability and lack of documentation makes developing with your framework a massive pain in the ass - anyone I know who can is migrating off it as quickly as possible.

They have this, by their own admission and from my personal experiences in the car. A detail you might be mistaken on - they’re not really fully driving the car remotely. They’re basically sitting in a call center answering questions for the car or plotting a course, or giving it the all clear to proceed if things are uncertain. And generally this only happens if the situation is one where the car can ‘safely’ stop and wait for a response. In situations like the one discussed here, the car doesn’t have time to wait for a response from a human operator, so it just does what it thinks is best locally.

Any actually-deployed driverless car has the same setup with remote operators - Cruise, Waymo, Zoox (if they've managed to deploy yet).

They have that, and these cars completely depend on it for fundamental parts of the stack. If they detect that their understanding of the scene doesn't match their map, they (generally) have remote operators clarify the best course of action for the car. This isn't a remote takeover - they just plot a path for it or select from a number of different options the car presents to them, or give it the all clear to proceed.

I worked in the AV space for years - not at Cruise - but during one of my first driverless rides in a Cruise car, we ran into a few blocks of SF where the power was out. About 3 street lights in a row weren't functioning. At each one the Cruise stopped completely for about 45 seconds before crossing - even if there was no traffic coming from it's left or right. That delay was so that a remote operator could clear it to proceed after it detected that there wasn't a functioning street light where it expected there to be one. After the first street light we were getting honked at and swerved around by cars behind us while it paused and waited for clarification.

Comparisons between the safety of robotaxis and human drivers are strange to make when 'robotaxis' still depend on remote operators for fundamental parts of the driving task

Reminds me of Mixamo, which is entirely procedural.

Or they pay for professional headshots for their LinkedIn page?

The event is held several miles “into” the dry lake bed - once you get off the paved road, you drive about 1.5 miles on the dry lake bed before you get to the “gate” of the event boundaries. You can see this pretty clearly if you look up its location on Google maps. Right now, the event is completely surrounded, for miles, by muddy lake bed. If you tried to drive your wheels would get jammed up.

Under the “principles of burning man” (which are largely not known or ignored by a sizable portion of of attendees in 2023), you should be prepared for this sort of event with enough extra food and water to wait it out. And I think a large fraction of the attendees are. If this was the camp I used to run with, the collective response to this would be to view it as an annoyance, but not life threatening - everyone was on top of their shit and had enough food, water, and shelter to last out something like this. And in fact, people would give you shit if you came around asking for help saying that you didn’t expect rain like this. You’re very explicitly told when buying a ticket that shit like this can happen and you should be prepared to deal with it yourself.

Whether this turns into something requiring external help or evacuation depends on what portion of the attendees really are on top of their shit in 2023.

Rain like this is rare buts it’s shut down exit like this before, this isn’t unprecedented. Even during a normal year they shut down entry/exit when sandstorms reduce visibility - often for a couple of hours every few days.

You’re told when you buy a ticket to prepare for this sort of event - it’s happened before. So anyone who’s actually on their shit will have enough food, water, etc to wait for it to dry out and leave normally. They even explicitly advise bringing a bucket to turn into a makeshift toilet if something like this happens and shuts down the porta potties.

That said, I’d bet a sizable fraction of the attendees are completely unprepared or dependent on others to get them ready. Probably people will share extra supplies with them. Maybe this entire event will expose just how commercialized it’s gotten.

It was a deeply divisive event - for a very large fraction of the playerbase, it was the most fun they've had playing WoW. The entire way the game functioned changed for a week - some people set out to grief, others worked to try to cure people infected. And it was done just before the release of a new expansion - any gear or progress that occurred in that week would be invalidated two weeks later when the new expansion raised the level cap by 10 - all of the best equipment you could earn prior to the new expansion would be replaced by normal leveling gear. The game was, so to speak, in a lame duck state at the time, so why not have some fun with it?

I worked in the AV industry - not at Cruise but based on incidents like this and other things I’ve seen from riding in Cruise cars - they absolutely need a cell phone connection and will strand or pull over (or stop in the middle of the street) if there’s anything more than a very transient break in remote connectivity. Fundamental parts of the stack assume that if there’s any uncertainty in a situation, a remote operator will be available to clarify how the car should proceed. The car can’t make certain safety or operational guarantees without this.

Can you add some more context on what you mean by “inappropriate”?

For people in this thread convinced that Google is scrambling to come up with a ChatGPT competitor - that’s only true on the product side. Internally, LamDA has been available with a chatGPT style interface for years (maybe more like 18 months).

Publicly available satellite imagery can’t track my car in any meaningful way. At best you’d be able to figure out the color and (rough) size of it if you knew my address.

Whatever API they release is going to be way more restrictive than what people can do with Stable Diffusion. I doubt we’ll see anywhere near the same amount of integrations unless OpenAI just lets you download the weights and run Dalle locally.

Imagen can produce images of humans - they’re just filtered out from the results by supervised models (for now). OpenAI did something similar with Dalle for a while IIRC.

For a generative model to hit the quality threshold you see on this site, it’s going to need so much training data that you can presume there’s already a ton of stock photos that match that description that it scraped from the internet.

If I had to guess, I'd bet they have a supervised classifier trained to recognize bad content (violence, porn, etc) that they use to filter the generated images before passing them to the user, on top of the bad word lists.

Sad to say I've been dissapointed in DALLE's performance since I got access to it a couple of weeks ago - I think mainly because it was hyped up as the holy grail of text2image ever since it was first announced.

For a long while whenever Midjourney or DALLE-mini or the other models underperformed or failed to match a prompt the common refrain seemed to be "ah, but these are just the smaller version of the real impressive text2image models - surely they'd perform better on this prompt". Honestly, I don't think it performs dramatically better than DALLE-mini or Midjourney - in some cases I even think DALLE-mini outperforms it for whatever reason. Maybe because of filtering applied by OpenAI?

What difference there is seems to be a difference in quality on queries that work well, not a capability to tackle more complex queries. If you try a sentence involving lots of relationships between objects in the scene, DALLE will still generate a mishmash of those objects - it'll just look like a slightly higher quality mishmash than from DALLE-mini. And on queries that it does seem to handle well, there's almost always something off with the scene if you spend more than a moment inspecting it. I think this is why there's such a plethora of stylized and abstract imagery in the examples of DALLE's capabilities - humans are much more forgiving of flaws in those images.

I don't think artists should be afraid of being replaced by text2image models anytime soon. That said, I have gotten access to other large text2image models that claim to outperform DALLE on several metrics, and my experience matched with that claim - images were more detailed and handled relationships in the scene better than DALLE does. So there's clearly a lot of room for improvement left in the space.