I don't know how many people know it, but FOSDEM's video recording infra is a technical masterpiece:
All the rooms have a live capture device sitting on the speaker's desks. This not only streams every room live on the website, but also deposits the feed on a server .
By the moment your speech is finished, you receive an automated mail with a link to cut your video. It is a limited web editor where you can change the beginning and the end of the stream, and changing audio channel too (talks are recorded with 3 mics).
You can accept the video as it is or do some adjustments and, voilà: a quick ffmpeg and the video is online ready to be viewed in a couple of hours.
I don't know you, but I am not aware of other conferences where the video recording is managed with such a DevOps pipeline.
When there are ~780 talks in ~30 rooms, we understand why automation is required.
It is impressive that they manage to get great quality video for most talks.
The videos from FOSDEM, and most conferences, mix slides and filmed presenter into one stream. I would be great if they'd be combined as two streams in one file, so viewers could switch between fullscreen slides and fullscreen presenter.
Always a delightful super-insightful tour from Chris. There's really a ton of new actionable ways to watch & analyze things that he introduced in this talk!
It makes me a bit sad that Kubernetes kind of feels like it's apart from cgroups so much. It has it's own idea of scheduling, and doesnt seem like it lets us use the incredible knobs & analysis that Chris talks about here because of that.
I count two talks that appeared to discuss machine learning which is better than nothing. But it is a bit sad to see that machine learning, and GPU compute more generally, is yet again practically nonexistent at a major FOSS conference. I'm not just talking about GPU acceleration for drawing but rather GPU compute more generally which of course includes deep learning applications.
I'm not looking for FOSS to be the leaders in the field or to develop any groundbreaking technology, but it would sure be nice if there was at least some acknowledgment and planning related to a rather significant component of computing these days. At this rate even today's prettiest, most polished, Linux distro/desktop will feel increasingly dated over the next few years.
I'm sympathetic to your argument, but isn't this at least partially because the cost of training LLMs and other very large/deep networks is so high (in the tens of millions of $$). The models themselves could be (and sometimes are) open source, but without the trained parameters they're not very useful.
The broader question here is how can open source play in this space? Given the impact these technologies are going to be having should we have some kind of public funding? Otherwise only very large tech companies are going to be controlling this space.
Inference is much more common than training just as using a word processor is much more common than writing one. Even from a training standpoint, there are a lot of interesting things you can do with DNN models that aren't at the large end (GPT-3/ChatGPT/DALL-E/Stable Diffusion/etc) of the spectrum.
Right now there's not really a coherent open source stack to even run pre-trained models on a GPU even though most of the pieces are open source... so I find it a bit baffling that no one is even talking about trying to put the pieces together. The whole GPU compute world still basically needs to be side-loaded on many distros today with the core of it (the GPU compute drivers) still proprietary. There are things like RustICL coming along that give me some hope but I'm more lamenting what appears to be a lack of interest in the larger open source world in changing things even when all the pieces are there. If open source seems 'a bit' (lol) behind on mobile, that's nothing compared to how it's going to feel on the desktop in a few years.
That's fair push back. I absolutely get that re: no centralized control/decision making. However, places like FOSDEM are where I've seen people raising community awareness on larger issues like this in the past. So I just figured I'd throw my $.02 out there in case it gets anyone thinking more about it or starting to work more in that direction as that's what I've got the time to do right now... yeah, I know: lame.
With the exception of drivers (currently), yes. But almost everything that exists is targeting developers and server infrastructure. If open source graphics infrastructure were in such a state, no one would even be running a GUI desktop on Linux, for example. (and things like Steam on Linux wouldn't be possible)
In fairness, I'm coming at this from the standpoint of someone who uses Linux as a daily driver on the desktop so I know I'm an extreme minority. It's also not lost on me that I'm probably also an extreme minority within that extreme minority...
FOSS and open source ML/AI are a perfect combination that can change the world of computing forever. Crowdsource the training of the models, embed them transparently in the apps and tools and never look back.
All the rooms have a live capture device sitting on the speaker's desks. This not only streams every room live on the website, but also deposits the feed on a server . By the moment your speech is finished, you receive an automated mail with a link to cut your video. It is a limited web editor where you can change the beginning and the end of the stream, and changing audio channel too (talks are recorded with 3 mics). You can accept the video as it is or do some adjustments and, voilà: a quick ffmpeg and the video is online ready to be viewed in a couple of hours.
I don't know you, but I am not aware of other conferences where the video recording is managed with such a DevOps pipeline.