It's useless for LLMs and it's actually slower than Hailo 8H for standard vision tasks, so, why ?

How, when foundries are a natural monopoly? How can you realistically create a startup to build a multibillion / multi year fab ?

There are 3 foundries capable: Samsung, Intel, and TSMC. Companies can contract out some to Samsung and Intel. TSMC can (and is) build more fabs.

I don't think that laying bricks and operating a crane is on the same level(or universe) of complexity as making and operating high-end semiconductor machines.

It could be 1 dollar a month, I'm still not paying to use my own ressources

Well, you obviously are using their resources, to kick off and register statuses of the jobs running on your resources, right? That is probably worth $1/month to you?

I wouldn't trust LMArena results much. They measure user preference and users are highly skewed by style, tone etc.

You can litteraly "improve" your model on LMArena by just adding a bunch of emojis.

There’s just no way to stop cheating client side despite what devs love to think. But server side anti cheat is much harder and requires more work; it’s much simpler to just install spyware / rootkits on the client and call it a day.

You can’t prevent wall hacks with only server side anti cheat. The client needs that data locally before the enemy is rendered on screen.

As mentioned in another comment, you can’t do this on the server without expensive checks for every single player that is always checking line of sight, because it’s not just your session running on a single server but multiple sessions.

And let’s say you did this, now you have a latency problem because most modern games to make them feel fluid has client side prediction with server reconciliation. This is what makes your modern games feel more responsive, if you put a constant server check there you have lost this.

No matter what people say online, it isn’t just move all of it to the server, there is data the client needs to know and can’t be spoonfed by the server.

I think it’s an organization accountability issue.

Why would a company pay for anti cheat infrastructure when they can outsource it to some company and blame them if there are cheaters or upset users? Windows is the status quo too, so it’s very easy to point to everyone else when justifying your choice to the execs.

It would be great if steam deck+box start costing studios quantifiable amounts of money that can be used to justify fixing this instead of outsourcing and hand waving.

CTRL-F: "Fokker-Planck"

> 97 matches

Ok I'll read it :)

why am I only getting 26 matches? where's the threshold then? :D

It's all about the en dashes and Fokker-Planck vs Fokker–Planck.

PDF files often break up sentences in ways that the find utility can't follow, so even if they ask have the same dash, it might not find them all. At least those names are uncommon enough you could search for just one.

AI is definitely related to dashes!!

Pandoc isn't a typesetter, it's more a translation engine between various typesetting languages (and it can then call the typesetter).

Pandoc is a ridiculously underrated and powerful tool but it solves a different problem. Someone still has to write typesetters ;)

I think I'm going to subscribe without any intention of using the app just as a "financing" donation.

I love, and hate, LaTeX and the idea of a LaTeX successor / alternative is incredibly appealing.

And the fact that they are aware that microtypography IS important and that they are working on it is a huge huge plus.

First-gen 50-year-old open source suffers from a first-mover problem of not knowing how people will use the thing. Thus, 50 years later, we end up with multiple-gigabyte distributions and messy, inconsistent syntactic approaches to hack together what people want and need.

Typst has 50 years of accumulated TeX experiences to learn from, and fit everything people actually want to use into a 45M binary, and maybe you'll download a few dozen K of package scripts.

I have used it for much more than academic publishing (book, brochure, and even card layout) and it's hands-down the best tool ever made for producing documents of any imaginable kind. Procedurally producing layouts from first-class JSON and CSV support is bliss.

Typst is open source, so you can run it on your computer; it's available as a CLI and has integrations with multiples IDEs (most use tinymist). Using typst is better than subscribing and not using it IMO because you can already start creating content and advocating for it, while telling the team about bugs or pain points

you can use the cli, subscribe (to the pro features of the app) and not use the app (online editor) to provide a bit of financial support WHILE using Typst and create content

Exactly my plan yes :)

I have made direct donations and I now just support financially by paying for a subscription to the web app.

I find myself switching between cli and web app a little more; the web app seems nice to experiment, share experiments especially when you need to demonstrate an issue when getting support, and has good enough git(hub) integration.

I would like Typst to support bounties, because I would throw a bit more towards HTML support.

LaTex is not too bad but it's hard to cook properly. My personal pain point with LaTex is that it depends so much on environment where it runs.

I've had some success setting up nix flakes for this! Using an LLM assist I even got a single font changed. I had previously run a bunch of scripts as root with no idea what I was doing in order to effect the same change. Annoying!

If you're not going to use the web app, you may as well sponsor them through GitHub instead. Also, you get stickers that way.

Have they indicated that they are working on microtypography support?

The example in this link of character-level justification is incredibly nice (enough to get me to try Typst), but it's not clear at least from this link whether they're actively working on microtypography.

> Has the situation changed on AI code legally speaking?

I think the position has shifted to "let's pretend this problem doesn't exist because the AI market is too big to fail"

It is going to be so interesting now that most software is going to be public domain. It's going to be us and the fashion world working just fine without intellectual property rights.

While your example is not AI related (should be handled by the LSP integration with "Rename Symbol") I agree that Zed's Next Edit Prediction model is *extremely* subpar. Imho they should either scrap it and just work on having a good integration story with third party models for the next edit (and maybe propose by default a partner model I don't know) or invest a lot more efforts into it.

But currently I sadly have to say the model's "help" is often a net negative.