upbeat_general

I am actually curious who is using these types of GPUs (H100 currently). Is it exclusively used for LLM training?

I can’t really see another market where it’d make sense.

I've been meaning to try NixOS personally but I spend most of my time on a shared server, so it's not an option. Thanks for the suggestion though, I'll definitely try it sooner rather than later.

Besides the rest of the people I work with not using docker making sharing difficult, docker is just another layer to maintain and deal with.

Especially since I don't work on a fixed project for very long and am constantly installing new dependencies, using different data, etc.

Not to mention I've found adding a layer of virtualization can break some already fragile libraries (e.g., I'd be concerned about multi-GPU training over NVLink).

I was just going through this today. Surely there has to be a better way of managing python libraries with CUDA kernels than what we currently have.

The worst part is, after spending a couple hours building 5 separate libraries from scratch and making sure all the versions match, there’s no easy way to reproduce your environment on another machine.

I may not know enough about the situation but I bought a set of bulbs a few months ago and they pretty much instantly connected to my Zigbee router connected to Home Assistant. Haven’t had any problems with them since (dimming & changing color temp).

I’m not familiar with the specifics latencies for drone racing but hierarchical planning/control is very very common in robotics.

It’s natural to say plan a longer term trajectory at 1Hz, a short-term trajectory at 10Hz, and perform control at the 100Hz, obviously with those rates varying based on the system.

And since (generally speaking) higher levels of planning require more compute resources, it might make sense to partition the compute and take the latency penalty.

I haven't, although I've watched a good number of videos from recent betas.

In terms of traditional automakers and systems that anyone can buy, Tesla is the leader (although I believe hyundai has a very good highway system, not city though). But in terms of self-driving, it's hard to rank precisely but they're certainly nowhere near, say, Waymo.

Plenty of people enjoy FSD and understand that they still have to correct it every so often but to a lot of people "self-driving" means just that.

It’s really unfortunate how much impact these promises have had. A lot of people believe them, even plenty of technical people I know. Lots of people [wrongly] really believe that Tesla’s are self-driving and are the leading company in the field.

+1 for VSCode Remote.

I've been coding for the past 2 years almost exclusively through VSCode remote on Linux servers with few complaints.

Get a seperate SFTP/network file client as the built-in VSCode one is slow and unstable but otherwise everything works great. If you're considering using any Jetbrains IDE over a network connection however, I'd encourage you to think again.

Same

Yep, I'm unable to push.

That’s not the case. It is significantly faster, at least for deep learning workloads. Check out Lambda Lab’s benchmarks.

Lower wattage allows more GPUs per system. For distributed training, that’s allows 8 GPUs per system which would otherwise be very difficult with retail GPUs.

And GPU-GPU interconnects are important for this type of training so putting as many GPUs as possible close together is necessary.

Agreed.

I thought the same thing about Python's indenting originally (and still do). But the ecosystem was enough to overcome it for me eventually.

Yup. I thought it was a userscript bug until I saw this thread.

Tons of popular iOS games are >1GB.

Not allow resetting from a device only with the pin?

Also don’t allow the pin for some operations (or let you disable this). E.g. for viewing passwords or other sensitive operations besides login, it’d be safer for me to not allow pin access and only Face ID.

Slightly off-topic: Does anyone have a recommendation for a pdb alternative?

I’ve found most of the alternatives have a high (for my use case) learning curve. For real debugging I use VSCode but it’s nice to be able to quickly throw breakpoints around.

One possibility is that it doesn’t make sense financially on any other platform for devs.

macOS games are dead and iPad games are better but still adding a complex feature for a small group of iPad users might not make financial sense.

My guess is because there’s a lot of ways of making things digitally and it’d need to be a really complex codec.

On the other hand, deep-learning based compression with autoencoders or something similar is pretty promising, since it can learn the constituent elements in a more general way, independent of what program was used to create it.

If you're referring to the Lex Friedman interview, at the start of his answer, he mentions it was a cost-based decision. And that radar/ultrasonic wasn't worth it for them, due to the additional time it took. Not that it wasn't helpful, just more effort that could be better spent elsewhere.

As someone who used to do webdev and now works in Python, I've been wanting something like this for a while.

The barrier of writing both a front and backend for a simple demo/viz tool is pretty high (relative to the time it takes).

I’ve very often heard cancel culture used to refer to “cancelling” of companies so I’m not sure I agree with that statement.

Also the line can be blurry with celebrities since they will engage in business deals.

Hence why M1 killed them.

Some people do all their work in browsers.

I find it interesting how the first main section for an OS release page discusses the features of an email application.

I wish this were the case! Plenty of more expensive cars (~40k) still don’t have it standard. And almost no ~20-30k cars have it standard either.

Do you know of any good explanations of the techniques you used for those who only touch PyTorch Eager + occasionally torchscript?

I do think there should be a default expiry but I’ve absolutely went back many many months and found old versions of files in cases I didn’t have proper commits.

They’re very useful for deep learning!

iOS apps that support X version can only be released once X version is publicly available so testing is through TestFlight only.