Huh, which model? I didn't pick up any of the tells. It seemed normal enough to me. (I vouched it, oops.)

Telling which model wrote it is much harder than telling that some model wrote it.

Interesting. I wonder where the "sweet spot" on the delay is. I've given up online news and twitter and all that entirely, and just read my print Economist, which comes weekly. That already has the news somewhat delayed and presumably with less of a rush to report has a little more fact checking.

This advertises up to a 3 month delay. I do like that with that much lead time you can see not just news but sometimes a bit of the fallout and consequences.

Not a fan of the name, though. It reads like those Claude comments that I absolutely despise, where it feels the need to comment in a reactionary way to the feedback it's getting and changes it's making, rather than a forward-looking way that's cohesive with the new state of the codebase.

Maybe I missed it, but is there a back issue available I could page through, or even just a full story somewhere, that I could read so I can evaluate what I think of it?

Luckily the magazine came long before Claude, but I suppose your comparison describes their mission pretty well. It's a haven of non-AI writing.

Similarly, what is considered health insurance is not really insurance but a convoluted way to pay for healthcare. I think there's a place for what you're calling disaster insurance but to me sounds truly like insurance for health-related issues (i.e. risk-pooling high cost low probability events). The conflation with what we've got now with that is one of the key reasons our health care system is so out of control, in my opinion.

I'm not sure Anthropic even cares about "releasing" the next version. The rust one has been in use in Claude Code for more than a month now, used by millions of people, and that's as far as they probably really worry about it. They bought Bun for Claude Code and I doubt the open source project matters to them otherwise.

If that is really what they are concerned with, they probably could have just had Claude Code rewrite Claude Code in rust? That likely would be far more efficient than changing the language of your typescript runtime that ships to run your react-based CLI app.

I imagine the $800k or whatever this has cost is coming out of Anthropic's marketing budget, so they can make a big splash about it.

> If that is really what they are concerned with, they probably could have just had Claude Code rewrite Claude Code in rust

Rewriting a runtime with well-defined interfaces and behavior is quite different from rewriting a user-facing application under active development by a decent chunk of their organization.

I agree, but this difference is very frustratingly omitted from the hype-sphere surrounding the rewrite, even though most software developers don't work on a runtime with well-defined interfaces and behavior. I know a number of people who've developed severe FOMO that their projects take longer than 2 weeks and maybe it's because they're not AI native enough.

(I don't mean to be critical of the Bun maintainers, who have always been open about the fact that they leveraged the specific context of their project to do this effectively.)

If you look at the reasons for rewriting Bun [0], they weren’t about efficiency and they don’t apply to Claude Code. It’s already written in a memory-safe language, and doesn’t have to interface to a library with a tricky set of invariants.

[0] https://bun.com/blog/bun-in-rust

I very much doubt they would want to abandon the wider community. They do benefit from others using it.

To the contrary, I think any problem with this release would be jumped on harshly so they're being more careful than usual. There's not a rush for the community to move to 1.4 and any issues could poison the community trust.

Like this rewrite did not terminally poison trust?

Has it impacted your use case negatively?

Yes, but it’s slower on non-bun runtimes. Nowadays, I believe they bundle bun with CC

They chose to use a react rendering to native TUI renderer, which was a source of a lot of performance issues.

AFAIK, they have written a new renderer.

> They chose to use a react rendering to native TUI renderer, which was a source of a lot of performance issues.

The source of the performance issue here is their own incompetence, actually. I have never in my entire career heard of a rendering pipeline as idiotic as the custom one they built.

>AFAIK, they have written a new renderer.

Which is equally terrible and has earned them every single rendering engineer in the world taking the piss out of them when they proudly announced it was as complicated as rendering a video game.

> They chose to use a react rendering to native TUI renderer, which was a source of a lot of performance issues.

That's such a stupid engineering choice that it really makes one question if the people working at Anthropic have any software engineering ability at all. There's no excuse for running React in a freaking TUI.

because their js code is so bad and slow that they need the bun performance hacks and optimisations.

Do these hacks make bun less secure than other runtimes? I find it a little hard to believe you can get much more robust and performant than V8 or SpiderMonkey without cutting some corners that notably went uncut by either for all these years.

Bun uses JavaScriptCore, which is Safari’s JS engine, for the actual interpreter portion of its runtime.

Which makes the parent comment’s “Bun performance hacks and optimizations” sound a little far-fetched, at least in this scenario I don’t think that’s gonna be the deciding factor. Their vendored JSC is also not touched by this port-rewrite at all whatsoever (still Cpp), so I’m not sure if it’s even possible to hack-improve all that much there.

yes, especially compared to deno bun does not have proper isolation or permissions at runtime layer which is why anthropic tries to fix the lack by a mixture of app layer (build into claude code) and os layer (srt) however using runtime permissions would allow much better security if done right.

Why wouldn't they?

Bun is a major part of the web ecosystem, having it developed by them via AI, sounds like a gigantic pr win.

> Bun is a major part of the web ecosystem

It was never that large but it did have a shot at getting more popular - until this drama.

No one worth their salt is building on Bun anymore, the creators and maintainers of bun have demonstrated their complete lack of care around engineering and support. It’s just too risky.

I regularly use bun, it's a decisive speed improvement over node. Large parts of the ecosystem build on bun.

If you install bun you're still getting the zig version.

I don't see major reasons to doubt the rust rework, especially as so much capital and token investment is being thrown at it.

As long as it fits better than alternatives, I see no major reasons to change.

> If you install bun you're still getting the zig version.

> I don't see major reasons to doubt the rust rework, especially as so much capital and token investment is being thrown at it.

You shouldn't trust the rust rework. Despite the mountains of capital and token investment which have been thrown at it, they still only ship the zig version to users.

It’s not. Node is part of the web ecosystem. Bun was a way to upsell paas

Sprites are clearly the future. Are there plans to give it a web UI? I haven't minded ssh'ing in, but I've wanted to share this with my wife who would use a sprite on her iPad. Basically, ChatGPT but with a full sandboxed Linux environment that can host stuff is a pretty compelling offering.

Check out exe.dev for basically what you are describing. Also remember these aren't really sandboxes per se, they're just computers/microvms. Exe also has a neat auth/https proxy that also does TLS/certs they have lots of cool tools and integrations.

I’ve been working on a project for this - https://mymachines.dev

It provisions a sprite and spins up Claude code/Codex on it for each session you create. Finishing up an iOS client for it too

I've played around with a few sprites and it's a very interesting concept. You ssh in and then it feels very much like a local development environment, except the dev server you start up and run locally is actually out there and accessible. It's a great way to vibe code a low traffic, personal site. I wasn't super clear on just how much scale it actually can reach, though. Maybe I missed it but I didn't really know how beefy of a server it was on or if I could scale up. But maybe a big scale is not really the intention of this product?

Wow, I can't believe Neal Stephenson of all people writes exclusively by hand. His books are tomes.

I totally buy that this is the case for him. I wonder how much is just what he's used to vs what's actually better. I feel like I write so slow compared to how fast I type that I struggle to put the words together on the paper. I can spurt out something with a keyboard 10x as fast, and then re-review and re-organize and all that.

Knowing that makes a lot of sense. Sometimes his prose can read like a sentence that was started but he’s too much effort in to give up on. It probably helps his writing to do it by hand - you can’t edit quickly on the fly with paper (unless you’re cutting stuff out). I could see this as being part of his process - just write and let the editing happen later.

Whatever the process it is, long as his books may be, it works.

This actually explains why I've never been able to get into his writing despite rave reviews and countless recommendations by people I know. Something about his writing always just felt tedious to me. Im an avid reader btw.

I find his earlier works a bit too bizarre and quirky. Although Snow Crash is the essence of cool.

Seveneves is probably my favourite, I guess I prefer harder scifi. Fall; or, Dodge in Hell and The Rise and Fall of D.O.D.O. were pretty good too, although less so. I want to like the Baroque Cycle but I haven't managed to get into it so far.

If writing by hand is good for your brain (and I suspect it is), the slowness may be part of the reason for that.

I wonder if fiction writing at that level is like coding. You can learn to type and navigate really fast, but unless you're working on blub software, if you're doing something interesting and novel, you won't actually get to the goal faster. The bottleneck is the ideas, the structure of the story and the development of the characters, amongst many other concerns.

If I was a writer, I wouldn't use paper not because I want to write faster but because my brain works best in iterations. I want to try out an idea and see how it feels and course correct. Doing so after writing a hundred pages would be tedious. But then what do I know, I haven't actually written any books.

David Foster Wallace was a longhand manuscript writer too.

I gave up on the Baroque Cycle, personally, because it felt like it was taking as long to read as it was taking to hand-write.

The point is to write better, not faster.

I wonder if Brandon Sanderson writes by hand :-D

No: https://faq.brandonsanderson.com/knowledge-base/what-program...

This is an interesting article. I don't really work with low level enough languages for this to matter (unless - does this ever show up in Javascript somehow?).

I guess I don't understand the "reduce" step. It seems like you have to be careful not to "undo" all the benefit from SIMD. Sure, it can compare 8 values in parallel, but then if you have to look at each of the 8 answers in turn you're back to where you began. Is the `@reduce()` function in the example a special Vector one that tells you if all the values are true or not in one "step"?

> Is the `@reduce()` function in the example a special...

Yes, it is. In SIMD you can do a reduce for a commutative function in O(log N) steps. Sum, product, min, max, all, any, etc.

Although in this case it might be better if you just take the mask (vector of booleans), convert it to a bitmask and check if it is (non-) zero.

If the language provides it, you might also use .all() or .any() for a mask.

Yes. "`@reduce(.And, ...)` combines every boolean using `and` and returns a single boolean."

If it's true (the common case here) then you proceed to look at the next 8 bytes.

If it's false, you apply a @bitcast (turn the booleans into bits) and @ctz (find the first 0) to get the index of where it was false.

Maybe I'm wrong, but should it be @clz instead?

No, the diagram looks like a binary literal but it's actually backwards from that. Lane 0 becomes the LSB, but that's on the left of the diagram.

Aha, you are right. Smaller indexes are for least-significant bits.

It can show up in JavaScript if you write your code in a way that lets the browser engine lower it to SIMD under the hood

how do I do that? :-)

It depends on what you're doing and which browser engine you're using, unfortunately.

Sounds like you could use some time in a group that bans discussion of politics.

Surely this is a bug in the harness and not in the model (where it's called "alignment"), right?

I mean, an LLM is just a pile of weights. All this happened because OpenAI had a little program running which called the model in a loop, and had tools that let it do all kinds of stuff. If your agentic harness isn't monitoring network calls and so on, and you just let the thing run without oversight, you're bound to run into issues eventually.