colinmarc

I find it works a lot better at smaller sizes.

I use my Anthropic sub with Pi. They keep saying they're going to disallow that and then walking back the decision. Or at least I think that's what's happening.

My absolute favorite modality is one I don't use all that much at the moment: Zed's edit completion.

If you're unfamiliar, it's like tab-completion, but it has a context that includes the edits you've made in the last few seconds, and it can predict around the cursor.

The model isn't advanced enough to understand complex tasks, but it has more the feel of the "crafting gun" in Subnautica or other survival crafting games, if that analogy makes any sense.

Personally I hate working with a chatbot - it's low-bandwidth and rage-inducing. If I could imagine a perfect workflow, it would be something like me whispering my train of thought as I program, and then pointing a very fancy "autocomplete gun" at the code.

I really like Orion, but I had to switch (currently trying Zen) because the 1password extension was so so broken. I know it's probably not their fault, but it really got in the way every day.

Maybe the game has changed with LLMs, but its been a running joke that engineers will build a startup/product/library/thing only to then realize they can’t get any users and that marketing and sales are hard.

I agree with your premise, but you're still viewing this technology through the lens of "a successful product" versus "a successful piece of technology".

Plenty of open source projects stay open source and are popular without ever making any sales whatsoever. I'm not trying to project my own motivations on the Iroh team; they may want to build a product out of it. For me, though, the project has a lot of appeal already, because it exactly and excellently fulfills a technological need, not because they brought me in with a "it's x but for y" narrative.

It's primarily an open source library, not a product.

I don't understand why HN seems so concerned about nailing down its "value proposition".

We use Iroh in production at work, and I'm absolutely in love with it. I'd describe it primarily as "Tailscale-style hole punching as a rust crate", but of course you can sprinkle a lot of cool p2p stuff on top of the basic QUIC connections.

Not affiliated, but I am a very happy user of Tailscale and a very happy user of Iroh; we use the latter in production at work.

Tailscale is a great service that happens to be open source, but Iroh is clearly structured as a library that you can build into whatever you want.

I would very much like to read the German, if anyone has it.

This is trivially circumventable by changing the system prompt (they string match against a blacklist).

I can't tell what's worse: intentionally obscuring the fact that the vast majority of people would pay ~no wealth tax or unintentionally forgetting that the vast majority of people would pay ~no wealth tax.

Wow - I love programming in YAML! You know what would make this really fun? Sprinkle in some Jinja. Then we'll be cooking with gas.

Previously: https://news.ycombinator.com/item?id=47691021

I know of one: https://lubeno.dev

I use mise to install it in a manageadble way (and basically everything else on my computer).

I don't think Earendil and OpenAI seem very similar. The former is a tiny startup of friends working on exclusively open source stuff.

Pi is a coding agent harness, like Claude Code, but significantly better and more elegant. Here is a good post describing it: https://lucumr.pocoo.org/2026/1/31/pi/

I switched a few months ago and have not looked back. Unfortunately Anthropic blocked access from Claude Subscription users today, but that's a different story.

I would absolutely be in favor of those measures.

There are obviously certain harms we are comfortable (trying to) prevent 13yos from having access to. So it's a matter of degree.

The internet you describe has been gone for a long time. The internet that replaced it is several degrees more harmful, to adults and children alike.

I'm surprised by the complete lack of dissent or even nuance in the discussion here. I'm much more ambivalent on this: the historical record for prohibition is not good, but instagram and the like are uniquely and disastrously harmful and the companies pushing them on children are powerful in a way that has no real historical precedent. In the balance, anything the reduces the power those companies have over our lives (and our politics) has to be at least considered. In other words, I don't think this is necessarily the right measure - but I'm desperate.

Didn't regulating cigarettes kind of work?

I opened this link intending to post exactly this. 1Password is just extremely broken on Orion. It's a testament to how much I like the browser that I'm still using it despite that (and despite the fact that Github was completely broken on Webkit this summer, but that's not Orion's fault).

Would the Swift UI also work on an iPad?

Yes, but probably not for the first version.

Do you have any comparisons with other tools (eg steam streaming, moonlight)

Steam streaming just doesn't really work on linux. Moonlight is somewhat similar in terms of direction, and has an established client base. I know of at least two projects to build servers for the Moonlight protocol[1][2].

The Moonlight protocol is a bit weird, because it's an open-source reverse engineering of a NVIDIA project, GeForce now. There are fundamental limitations to the protocol, for example that the cursor must be rendered in-stream or simulated. Using my tool, the cursor is rendered locally, and custom cursor images can actually be pushed to the client, for a seamless experience. This sounds like a minor detail but it matters a lot for subjective latency. I'm also working on employing tricks like hierarchical coding using FEC in the protocol, because I hate VBR encoding for games (it makes text blurry and breaks immersion). Those tricks aren't really possible in Moonlight.

All of the Linux solutions I know about have significantly higher latency compared to Magic Mirror, although I don't have numbers for exactly how much higher. (I have a benchmark to test the latency of my tool, but the others don't.) I'd encourage you to try them out and get a feel for the difference.

Finally, I think Magic Mirror is the easiest to install and get going on the server. It has almost zero runtime library or service dependencies (there's a pesky dynamic link against libxkbcommon which I haven't managed to remove), so you don't need to mess with pipewire or docker or anything - it's completely self-contained.

All that said, the existing tools have the advantage of a larger user and contributor base, whereas Magic Mirror is just me on a mission so far :) So they're likely to be much more stable and usable.

[1] https://github.com/LizardByte/Sunshine [2]: https://github.com/games-on-whales/wolf

Edit: an earlier version of this article mischaracterized the status of GeForce now :)

Beat me to the joke!

A large number of people at Stripe (I want to say more than half, but that might be wrong) dropped out of college or didn't go at all. Given that a bigger portion of our peers outside Stripe do have college degrees, I think that suggests a different correlation entirely.

On my part, I dropped out of school almost immediately (I was an Econ student at Suffolk in Boston). I joined Stripe right after graduating from Hacker School batch[2].

heh, I didn't think anyone would actually use that - I originally wrote it meaning to use it as a tutorial for the blog post, then scrapped that idea.

Thanks for the kind words!

Alum from the third batch here. In the hopes that I can convince someone else to apply, my story:

I applied to hacker school, on a whim, while I was working on a php app at an advertising company in Boston. I didn't think I would get in, and I hadn't made any plans to support spending three months in NY. When I got in, I decided to drop everything, and left a month later on the dot. It may have been the best decision of my life to date.

Going in, I was a programmer, but I wasn't very good. Hacker school was so transformative, so incredibly valuable that three months later I was accepting an offer at Stripe, where I am now.

The strange thing is - the value doesn't come from lectures, or workshops, or anything like that. Hacker school is valuable because it gives you the space to grow into your full potential on your own. Just having three months where I could dabble and play in the company of similarly-minded people was the best thing for me, and it's hard to overstate how much I learned.

Do it! Apply! I don't know of a better way to spend three months.

I don't see how this attack would work on SPDY, unless you could add arbitrary headers to the request. CRIME works because the entire request is compressed together, whereas in SPDY the header block is compressed separately.

I think it could be a bit more declarative.

That last language is REALLY concise.

Some really clever tricks here. I particularly like the __getitem__ one. I'm simultaneously excited and filled with horror. =)