Quite a few topics have emerged in the past week, and some exceptionally diverse ones. So there will be about the extinction of projects that reigned in the headlines, the definition of the Open-Source name, and new algorithms! And finally – a model for generating music.
1. “So this is how Metaverse dies…with thunderous applause.”
I say right away that the headline above is a kind of LowQualityBait, but first – I will try to defend my thesis, however, and secondly I was not able to deny myself a reference to Star Wars. While things aren’t all that bad with this so-called “future™️” (as you’ll see in the last section of today’s issue), some of the projects we’ve read about over the past few years are slowly starting to have their oxygen cut off.
This is nothing new, after all, we already reported at the end of the year on the massive reductions in Alexa and the focus shifted from Google Assistant. However, a broader pattern is evident in this. Well, during the layoffs at Google, the experimental departments were hit particularly hard, with the number of reductions being disproportionate to the rest of the company. Area 120, the company’s in-house incubator, will be shut down along with most of its remaining projects. Incidentally, the majority of employees working in the departments were “thanked for their cooperation”. FuchsiaOS – an experimental OS that is supposed to be Google’s future – is, in percentage terms, one of the projects the most affected by the layoffs.
Equally interesting (uninteresting?) is the situation at Microsoft. Reports report that AlterspaceVR – a division of the company working on projects such as HoloLenses and all related VR/AR innovations – has been shut down. Microsoft also got rid of the team working on the Mixed Reality Tool Kit, a set of tools for creating virtual worlds.
- Google’s Fuchsia OS was one of the hardest hit by last week’s layoffs
- Microsoft has laid off entire teams behind Virtual, Mixed Reality, and HoloLens
- Mixed Reality Tool Kit
2. Before you call something open-source, think twice whether it is realistically open
Sometimes it happens that as quickly as an announcement appears, so quickly it disappears from the web. This time it happened to Confluent – the creators of Kafka – and the whole context is so interesting that I wanted to describe it here – as a lesson and warning.
Confluent made an announcement last week that they want to make their proprietary Kafka tool, the Confluent CLI, a standard for the entire ecosystem. In the aforementioned (already 404-ed) post, they said that they are making it open-source software…. and 💩🌊 began.
This is because it turned out that, in the latest fad, Confluent proposed a license, which has little to do with open source – as it prohibited, among other things, the use of the project’s code by competitors. This sparked a discussion online (especially on Reddit), causing Kris Jenkins, the company’s Developer Advocate, to react, apologizing for the unfortunate use of the term open-source and promising that in such a context Confluent would use the less controversial “source-available” in the future.
I suspect that wasn’t the company’s biggest problem of the day, though – the date of the online drama coincided with the 8% staff reduction 😒.
3. Revolutionary pathfinding algorithms are still being developed in 2023
New algos rarely make the headlines, but we are not left without help. When I was still in college, I had this flash of thought: “they teach me about all these algorithms, but if a new, groundbreaking one comes along, how will I learn about it?”. Today I know that my fears were unnecessary – and Quanta Magazine plucks from thousands of academic papers the ones that are really interesting.
This time is no different. Probably each of you has heard of Dijkstra’s algorithm, which allows you to find the fastest route in a directed graph in which the paths have weights. However, there is a simplification in this definition – Dijkstra’s algorithm works, but only if the paths have weights greater than zero. If negative weights can also appear in the graph, the problem gets much more complicated – it may turn out that it is more optimal to choose a “more expensive” path, because at some point its weight will be significantly reduced. That’s why we didn’t have a fast algorithm to deal with these problems…. until now. Thanks to clever clustering of graph fragments, the authors of the publication Negative-Weight Single-Source Shortest Paths in Near-linear Time were able to come up with a solution in near-linear time. Aforementioned Quanta Magazin in its publication Finally, a Fast Algorithm for Shortest Paths on Negative Graphs presented it to the masses, explaining the context of the problem and the complexities involved.
Bookmark this article. I have a feeling that as early as next year you will probably be able to implement it during Advent of Code. After all, it’s the only time, aside from hiring interviews, when programmers have to traverse the graphs en masse.
- Finally, a Fast Algorithm for Shortest Paths on Negative Graphs
- Negative-Weight Single-Source Shortest Paths in Near-linear Time
4. MusicLM – a model that generates music from text
And finally – the obligatory Large ML model. After all, what would a week be without another revolution in this topic?
Indeed, Google has prepared a model that is capable of generating music from a text prompt. It sounds revolutionary, and the whole thing has made its way through all sorts of technological publications, but it hasn’t hit the mass consciousness like ChatGPT. The reason why OpenAI projects and not Google usually fire our imagination is simple – after all, such ChatGPT or even now Dall-E 2 can be played with, so they have virality in the nature of their entries. This one won’t appear in the case of MusicLM – since we can’t create our own samples (at least yet, probably never), we have to believe the creators that their model works as well as they announce, and judge the quality itself on the samples provided by the creators.
So I listened to the attached examples and. are really quite good. Of course, we’re not talking here about some outstanding melodic lines that will get an Oscar for “Best Soundtrack of the Year”, but they definitely don’t hurt. Such as “The main soundtrack of an arcade game,” for example, is very much in line with what can be found in indie productions. Concluding, examples from Google are really enjoyable to listen to.
The whole thing will certainly find its uses, I imagine it would be great if only as personalized music in a store. However, I suspect that the biggest problem in implementing MusicLM in “creative works” will be the lack of control – it seems to me that describing a piece of music well is even more difficult than an image. Everyone seems to have something slightly different in mind when seeing the prompt “melodic techno.” The creators of MusicLM do give some control over the melodic line, but in practice, it requires making music by trial and error. It’s ideal for a noob like me, but those with a bit of hearing will probably create something better without much effort..
Bonus from editors
At the very end, I will share some thoughts on Blizzard and China, in the context of the shutdown of their games in the Middle Kingdom that happened last week.