pavpanchekha
Joined 16 years ago
Karma
2,869
Submissions
459
Member since
May 16
Recent submissions
Comment · 1 month ago
Last author here. We've talked to the Chrome folks and they are interested, but it's difficult work. Chrome is big enough that emitting different sequences is hard and would requiring changing a lot of internal abstractions. Skia would love to do it optimization like this but it's a small team with a lot of other priorities. Integrating outside code is hard.
Comment · 1 month ago
I am a big fan of DB-style thinking, very much on the same wavelength as you :)
Comment · 1 month ago
Let me also add that nanobench was a huge help, not just because it was a good benchmarking tool but also because it gave us some confidence that we're measuring the right thing. It's easy to make _something_ faster but hard to know if it's the right thing. Having that come pre-packaged from the project answers a lot of tricky questions that would otherwise be easy to get wrong.
Comment · 1 month ago
Thank you! The SkRecord system was _perfect_ for doing these optimizations. I don't think it would have been possible to do this project without it.
Comment · 1 month ago
Last author here, this is very much what I've worked on for most of my career. In this project, I had the idea of optimizing rendering instructions years ago, while I was writing https://browser.engineering/, but the hard part of this project was being very careful with the semantics of Skia itself. It's _super_ easy to write down rewrite rules that _seem_ correct, but are actually only correct when, say, something is opaque, or has the right blend mode, or two things don't overlap, or something like that. Which is why this paper focuses os much on carefully defining that semantics. We actually did the semantics in Lean because otherwise we couldn't consistently write correct rewrite rules.
Comment · 1 month ago
Hi folks! Last author here, happy to answer questions, very surprised to see this on HN. We had a blast working on this. Let me add that the Skia team at Google was super supportive, met with us many times to explain a lot of stuff.
I had the idea for this project years ago while writing Web Browser Engineering with Chris Harrelson (see https://browser.engineering/). Then a few years ago I made a first attempt at this project with Yuvaraj (https://droidkid.github.io/), but for various reasons we never got very far. I restarted the project with Bhargav (https://bhargavkk.com/) about a year ago, and focused much more seriously on the semantics of Skia itself, which made progress much more rapid. Still, I was, frankly, shocked by how good the results are.
Comment · 2 months ago
I do a substantial amount of coding in Racket, including maintaining the Herbie numerical compiler (https://herbie.uwplse.org/) over the last decade.
Racket is great! The runtime is reasonably fast, and the standard library is exceptionally featureful, including, for example, a decent plotting library, an HTTP server, decent HTML and JSON support, several forms of multi-threading, and a quite good FFI, all of which Herbie uses extensively. I suppose the parentheses are a question of taste (I like them!) but a lot of the specific syntactic decisions, like the `for` and `match` macros, are quite nice.
Comment · 2 months ago
It's about chips with a large enough scale up domain. Larger domain allows for bigger model, which is what's driving this jump. You've got to get the chips, test them, tune kernels, then start a big pre train, mid & post-train, and only then do you actually get the model. So it takes time. Anthropic got there first partly because they use different hardware (TPU I think, maybe Trainium) which had larger scale ups earlier.
Comment · 3 months ago
A lot of algorithmic improvement in AI is ultimately bottlenecked by compute. It is very easy to come up with ideas that could improve models! But to prove that they do, especially at scale, is expensive and takes a long time.
Comment · 3 months ago
Making Luna, which was already very cheap and extremely capable, 5x cheaper is crazy. I use Sol at work but Luna at home, and while there's definitely a difference, it doesn't feel like night-and-day. After a year of ever-increasing prices it suddenly feels (between this, Kimi K3, GLM 5.2) that prices are falling again.
Comment · 3 months ago
For compiler work I found that Sol is noticably better than 5.5 (and I generally use OAI models because I like the Codex app), but Fable was still obviously better.
Comment · 5 months ago
OpenAI used to make Codex-specific models, but they stopped. What I've gathered from interviews and similar is that training two models isn't worth the (small) lift from having a coding-specific model. You're pre-training on everything anyway, and coding RL is reasonably useful for general-purpose models too.
Comment · 7 months ago
University of Washington Programming Languages and Software Engineering (research group).
I'm not at UW any more, I'm now at Utah, but some of the Herbie team is at UW and they provide the infrastructure
Comment · 7 months ago
Documented here but yes it's an average, of something similar to but not exactly the same as relative error: https://herbie.uwplse.org/doc/latest/error.html
It's true that averages can be misleading but we encourage users to think about it instead as a percentage of inputs. In practice the error distribution is very bimodal, the two modes being "basically fine" (a few ulps of error) and "garbage" (usually 0 instead of some actual value)
CS Professor at the University of Utah. I study web browsers, floating point, programming languages, and automated reasoning.
Website: pavpanchekha.com Email: [email protected]