HNHacker News
TopNewBestAskShowJobs

jkelleyrtp

2,460 karma · joined September 24, 2018

https://jonathan-kelley.com
submissionscomments
jkelleyrtp··on Coconut by Meta AI – Better LLM Reasoning with Chain of Continuous Thought?
I think this might be the “it” moment for AI/LLMs. I was hiking with a friend recently and we talked about this at length.

The arc-AGI results from O3 are apparently a result of chain of thought given enough time to explore a solution space. Reasoning might be simply a higher dimensional form of rubix cube solving. BFS, search, back-tracking, etc. It seems unlikely that humans think in “tokens” so why do LLMs?

By staying in latent space, the models are free to describe an “idea” in higher resolution than what language allows. English is coarse, granular. Latent space is a much finer representation of ideas and their interplay.

Latent space is also much cheaper to execute in. The model can think without the language encoding/decoding step. This lets it branch out hundreds of ideas and explore only the most useful ones in a fraction of time that reasoning “out-loud” would take.

The states also don’t need to be tied to language. Feed in a robot’s state, time series data, or any abstract data. Reason in category theory or linear algebra or complex analysis. Humans are hard wired for one set of math - an abstract latent space can represent anything.

I’m a bit disappointed OpenAI didn’t stumble on this first. I’ve been skeptical of LLMs since their big debut last year. LLMs seem like a great way of solving language, but reasoning is much more complex. Once you grok the math behind the current models, you immediately question why the encoding/decoding step is there. Diffusion models are incredible but it felt that LLMs lacked the same creativity. Encoding/decoding forces a token-based discretization and therefore a loss of complexity.

With the byte-latent paper it was quite clear we’d see this paper. This truly might be the “it” moment.

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
This version fixes a lot of inconsistencies in our hot-reload. We can't reload all Rust code, but we can reload some simple Rust expressions and way more RSX.
jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
> So you have a DOM, but if the target renderer is not webview or browser, you create a DOM in some other way?

Yes - we have a project called Blitz which is trying to emulate just the DOM without the browser. It's built on Servo/Firefox tech.

https://github.com/dioxusLabs/blitz

> The browser is already there, and called webview on all major platforms. I'd say Electron is popular for maturity reasons, that architecture makes limited sense even if you are using web. Tauri would be a better comparison.

The default dioxus desktop/mobile renderer uses the system's webview for exactly this reason. The blitz stuff is still young but is planned to eventually replace the webview renderer. Fortunately, webviews are on many platforms already.

> .. That also means users have to come up with the cross-platform API surfaces themselves, right?

Yes, but this is a challenge in every cross-platform framework. Flutter has pub.dev, react-native has expo modules, etc. In 2025 this will be a huge focus for us, and in all likelihood we'll just find a way to bind to existing widgets in the respective ecosystems. Easier than to try and bootstrap a pure-rust ecosystem (though it would be nice). Projects like our SDK do this, but the surface area is quite small when compared to what we eventually want to support.

http://github.com/dioxusLabs/sdk

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
It's very much inspired by Jetpack Compose and SwiftUI.

At the end of the day we just use the Rust Struct parser!

It's very similar to JSON or RON if you're familiar.

https://github.com/ron-rs/ron

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
I never know how to position Dioxus, so would be happy to have some 3rd party insight.

When I say we use a webview, people immediately jump to Tauri. But we're very architecturally different than Tauri. No WASM on native, just a regular old executable.

When I compare us to Flutter, people immediately think of Flutter's negatives (poor web support). We have excellent web support (DOM is 1st class).

When I compare us to React Native, people immediately think the web is 2nd-class and then wonder why we don't implement NSScrollView and all that.

We are shooting for something like Flutter (web, desktop, mobile, embedded) but where the web is the defining feature.

As someone outside Dioxus, how do you think I should describe it?

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
Not sure if you grok 100% what we're building.

Dioxus-web is basically React/SolidJS/Svelte. It writes to the dom and handles dom events just like any other web framework.

Where Dioxus differs is on native platforms, where we basically try to provide a Web-like API to native widgets. Think electron but you're not shipping a browser, just the rendering engine (that fits into 3.5mb!).

Our users on native platforms like being able to access system APIs with no intermediary. You can spawn threads, talk unix sockets, call FFI, etc. Stuff like electron is heavy, slow, and puts an IPC boundary between your UI and the system. We're trying to dissolve that.

Long term we want to expose JS and Python bindings for our native engine - Rust is not necessarily the "killer feature" there.

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
Both would help. Sledgehammer is fast because it batches and interns strings on the Rust<>JS boundary.

Rust strings are utf-8 and JS strings are utf-16 (don't ask me why...).

The Dioxus VirtualDom is ridiculously fast - it's closer to the micro-dom SolidJS has (yes there's a micro-dom) than an actual virtualdom that React has.

Sledgehammer queues up all the Dioxus mutations into one fat binary batch and then runs a hand-rolled utf-8 -> utf-16 converter since the browser's converter is so slow. All the dom mutations are done from javascript. If direct DOM access landed, then we'd be able to do that from Rust directly.

And ultimately, all this work is like 10% of the actual "time" taken that you see on benchmarks. The vast majority of time is the browser running layout and re-painting. Reconciling a dom is not that hard or expensive of a task, frankly.

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
This is more of a culture thing with the Rust ecosystem more than anything else.

If the culture of the Rust ecosystem was to publish a 1.0, we would do it. And frankly, I'd really like to have an extra "version" to play with - 0.x.y is quite limiting.

There's a few core things I want to clean up and split out in 2025. If all goes well, we might see a 1.0 next year.

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
If you don't want to use web technologies but still benefit from the non-web parts of the ecosystem (SDK, hooks, build tools) there's Freya. The creator is a wonderful guy and has put a lot of great work into making a nice Skia-based renderer for Dioxus.

https://freyaui.dev

WRT the other comment - `rsx!{}` is carefully designed to not leak details about the web. RSX is entirely abstract over the element+listener set so you can build whatever you want with it.

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
Just chiming in because I think the discussion is interesting (creator/founder here).

We don't need the OSS crowdsourced funding. Early 2025 I have plans to distribute whatever OSS capital we still have back into the Rust community, so at least donors won't feel like their money went directly into a company.

Fortunately, we did raise a good round and have plans to get "more serious" about it. Hiring is hard and this year it was challenging to balance hiring with solving users' issues in a timely manner. It's my first time doing it. Lots of the project's details are tucked deep in the depths of my brain.

We sadly have the curse of maintaining a project where the devtools need to work on mac/win/linux/wsl and the apps need to work on mac/win/linux/wasm/wsl/android/ios. I don't think people grok the scope of this work and the permutation explosion it brings. We test against mac/windows/linux of course, but there's always some weird issue with WSL or Windows Paths or some dependency's missing system library.

For example, we had to drop 3rd party tooling for mobile this release since that was constantly breaking our setups. There's still more dependencies we don't own completely that break frequently. I'm hopeful that hiring a larger team and prioritizing quality for the next few months will "solve" the majority of issues people run into.

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
There is a weird issue with `cargo`'s output that not all errors get emitted as an "error" - I just kicked off a new `binstall` build `cargo binstall dioxus-cli@v0.6.0 --force` that fixes the filtering issue.

You can also press `t` for `tracing` logs which will emit the relevant logs. I'm sure it's something pretty simple / easy to figure out like a wrong version.

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
In, my opinion it's good for ambitious projects to have a stable financial foundation. What negative things one project does because of venture capital does not mean another will.

There's no CLA. Dioxus is MIT/Apache-2 licensed.

> How many of the contributors are directly payed for their contributions?

I unfortunately feel like you're asking this in bad faith. Our team is very small, we're very lean, and we have funding sources that aren't just venture. We also have a very active community and people are building new libraries and becoming dioxus contributors every day.

https://github.com/DioxusLabs/dioxus/releases/tag/v0.6.0 (scroll to contributors)

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
The DSL is basically just Rust structs with optional fields. We essentially use the same struct parser rust uses but with some small tweaks.

You want to use a macro for this so you can do hot-reloading - doing it with native Rust is currently impractical for rapid dev.

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
Dioxus was awesome for a long time without money :)

As my little side project, we pioneered html-templates, wasm-bindgen string interning, macro hot-reloading, macro auto-formatting, multi-platform builds, native HTML/css rendering, and a number of other "awesome" things before raising a dollar.

Users want a guarantee the project isn't going to disappear tomorrow and sadly 100 bucks a month on open collective isn't a great guarantee.

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
In the same way React Native still uses JSX, Dioxus uses RSX even if those widgets are rendered natively (say, on the GPU).
jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
Our hot-reloading is "magic" for the Rust ecosystem (this post's original intended audience). Hot-reloading formatted strings and simple Rust code is basically magic in Rust land.

I use blazing-fast tongue-in-cheek but Dioxus is really really fast. We did a ton of R&D into making Rust<->DOM very fast - our sledgehammer binding layer is faster than SolidJS on the JSFrameworkbenchmark [0].

As for rendering - we have two options: webview and GPU. The GPU stuff is nascent but capable. The final vision is something like React Native Skia where the majority of the interface is drawn on the GPU but interactions are done via Native System APIs. That way you get apps that look the same across platforms but "feel" native. To render, we have to step through the platform's native containers anyway, so you can always composite in a native system widget as a component.

https://krausest.github.io/js-framework-benchmark/2023/table...

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
Oh, my bad. That language has never been accurate and I'm not 100% sure why it's phrased that way. Just pushed an update.
jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
Oh, my bad. That language has never been accurate and I'm not 100% sure why it's phrased that way. Just pushed an update.
jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
We don’t necessarily use Tauri… Tauri has a library called wry that’s basically a more maintained fork of the Rust web-view library.

https://github.com/Boscop/web-view

We are great friends with Tauri but I think people misunderstand our architecture and philosophy just because we use one of their maintained libraries.

We’ve been working on a native HTML/CSS library for a while called blitz which is reaching stability soon.

https://github.com/DioxusLabs/blitz

jkelleyrtp··on Dioxus 0.6 – Crossplatform apps with Rust
Dioxus is similar to modern React Native in architecture. Your rust code is running natively and you can call into JNI and Objective-C freely without a VM in your way.

It’s retained mode and stays close to the architecture of a React/Svelte app.

On the web, it talks DOM directly. We have SSR, hydration, streaming HTML, and a fullstack nextjs-esque framework to go with it. All done via webassembly.

The async story is great. Our state management is multithreaded and many users are using web workers and PWAs.

We have focused less on building a good multi-tenant story but you can definitely integrate React into Dioxus and Dioxus into React. We’re close to runtime compatible but React core is a shifting foundation to build on.

The API feels closer to solid now but the scope of the project is big and APIs are bound to change. This update, we changed a few things but upgrading is quick. It will be a while before we want to rework interfaces again, and we’ll likely provide codemods at that point.

I think the community is quite large (gaining critical mass for library support) picking up steam, but JS land certainly is going to be much bigger.

jkelleyrtp··on Show HN: Whirlwind – Async concurrent hashmap for Rust
Tokio has a task budget that will cap at 32 or 256 polls of the same task before switching to another task. So, yes a spinlock, but not likely to deadlock.
jkelleyrtp··on My negative views on Rust (2023)
The title is inflammatory and yet there are few nuanced takes in the article. It's weird to see it shoot to the top of HN.

I think the loglog article is a much better, nuanced, full critique of Rust.

https://loglog.games/blog/leaving-rust-gamedev/

The internet is just so full of negativity these days. People upvote titles but don't read articles. Reading about people's views on subjects is useful, but I don't think this one is.

jkelleyrtp··on My negative views on Rust (2023)
I mean, maybe?

If you come into Rust thinking you're going to write doubly-linked lists all day and want to structure everything like that, you're going to have a bad time.

But then in python you run into stuff like:

```

def func(list = []):

   list.append(1)
```

and list is actually a singleton. You want to pull your hair out since this is practically impossible to hunt down in a big codebase.

Rust is just different, and instead of writing double-pointer code, you just use flat structures, `Copy` keys, and `loop {}` and move on with your life.

jkelleyrtp··on My negative views on Rust (2023)
"There are only two kinds of languages: the ones people complain about and the ones nobody uses".

---

Glad to see fluffy negative articles about Rust shooting up the first slot of HN in 20 minutes. It means Rust has made finally made it mainstream :)

---

The points, addressed, I guess?

- Rust has panics, and this is bad: ...okay? Nobody is writing panic handling code, it's not a form of error handling

- Rust inserts Copy, Drop, Deref for you: it would be really annoying to write Rust if you had to call `.copy()` on every bool/int/char. A language like this exists, I'm sure, but this hasn't stopped Rust from taking off

- Fetishization of Efficient Memory Representation: ... I don't understand what the point is here. Some people care about avoiding heap allocations? They're a tool just like anything else

- Rewrite anything and it gets faster: okay sure, but there are limits to how fast I can make a Py/JS algorithm vs a compiled language, and Rust makes writing compiled code a bit easier. People probably aren't rewriting slow Python projects in C these days

- Rust is as complex as C++: ...no, it's not. Rust really hasn't changed much in the past 6 years. A few limitations being lifted, but nothing majorly new.

- Rust isn't as nice of community as people think: subjective maybe? People are nice to me at conferences and in discussion rooms. There's occasional drama here and there but overall it's been pretty quiet for the past year.

- Async is problematic: Async Rust really is fine. There's a huge meme about how bad it is, but really, it's fine. As a framework author, it's great, actually. I can wrap futures in a custom Poll. I can drive executors from a window event loop. Tokio's default choice of making `spawn` take Send/Sync futures is an odd one - occasionally cryptic compile errors - but you don't need to use that default.

I'm unsure why this article is so upvoted given how vapid the content is, but it does have a snappy title, I guess.

jkelleyrtp··on Launch HN: Firezone (YC W22) – Zero-trust access platform built on WireGuard
Hey! I worked on WARP at Cloudflare. I believe Cisco has anyconnect and then there's zscaler.

I'm curious how you guys are competing with the other folks in the space. WARP was/is a really tough product to maintain (crossplatform networking is very difficult). CF was doing well with WARP mostly due to the distribution advantage. I imagine it's harder for startups to break into the space.

jkelleyrtp··on Sohu: The First Transformer ASIC
?

If you can design and ship ASICs faster than anyone else for new problems then you can become the dominant player. All the new LLM startups just use the OpenAI API interface since they were first to market dominance.

jkelleyrtp··on Silicon Valley's best kept secret: Founder liquidity
Secondaries have become more popular over the past few (5?) years but you need to know to ask.
jkelleyrtp··on Swift Static Linux SDK
I have a slight inkling that Apple internally is using Swift for backend stuff somewhere, especially with all the ML stuff they just shipped. Word on the street is their datacenters are full of M-series chips. I could see those servers running Swift + distributed actors + mlx.

The move to make swift run on Linux could be part of that work - their sysadmins, devops, and backend folks probably use Vim/VSCode and deploy on Linux.

We might see the release of a very good backend framework by apple for Swift that works on Mac/Linux.

The new macros they added kinda bolt on serde into the language too - and the distributed actor stuff is basically RPC in the language.

jkelleyrtp··on Swift Static Linux SDK
The default testing framework made my jaw drop a bit - I have a very hard time seeing Rust have the enthusiasm or momentum to implement something like that right now.
jkelleyrtp··on Swift Static Linux SDK
Some people want generics and ADTs in Go. If you're using go because it's

1) fast

2) expressive

3) concurrent

4) compiles natively

then I think Swift is the better of the two.

If you just want a very simple language, then sure, Go is there. But I don't think that's necessarily what makes Go, Go.

← PreviousPage 2 of 9Next →