jes5199

I’m building a heavily parallel dataflow system. I thought Rust might be good for concurrency.

what’s the difference

this is the same thing we said about offices in the fluorescent era

I tried that, but the Rust build process was too painful, and agents seemed to burn a lot of tokens guessing how to get the code to compile. I rewrote my project in Elixir and it’s been going much more smoothly

a tmux session where every window is a claude code instance in a different checkout of the repo

and then an MCP+Channels system that let’s the claudes DM each other

plus the Telegram channel so one of the claudes can talk to me over text message

I think this article underestimates how weird things could get if we had low employment but a vast surplus of material goods. I don’t think a society has ever really been in that position

no, but they weren’t exactly churning out new features in the 2010s either

I could not agree less. People used to call Python “executable pseudocode” - in that spirit, Haskell is executable pseudo-math. If you’ve done enough higher math that a professor’s whiteboard notation feels natural to you, then Haskell might feel like a reasonable approximation of that style. Otherwise: it’s line noise.

(I write Haskell professionally)

it was kind of a miracle that it held together as well as it did pre-Microsoft. I think to some degree, they got lucky, and were able to coast on being in the right place at the right time. And then because they were so central they attracted some amazing talent who managed to keep the thing scaling up _despite_ the culture.

somewhat better leadership, I think

This is just CRDT merges and better diffs?? I think the future of version control is much, much weirder than this. Like if you have CRDTs why not have ephemeral branches with real-time collaborative editing and live CI as you type

one of the other head-exploding experiences from that startup was when a major cell-phone company sat down with us and said, we have an idea: the ad-free cell phone. What if, every time a website would normally show an ad, we just paid them not to, at about the same rate the ads are paying. How much would that cost?

and the answer is: not much money at all. we ran the numbers and a typical user’s browsing was worth something like $20/month total across every site and every app combined

but no one can figure out the logistics, so we’re stuck with ads

it was a single monthly subscription to a bundle, and the clever part is we would measure time spent on each site and divide up the money proportionally, so the site you spent the most time reading would get paid the most. Our founder had the idea that this would incentivize higher-quality content. We never got enough paying subscribers to really pull it off

yes perhaps only the big players have this problem

I used to work at a startup that was trying to replace ads as the funding source for news (we failed, obviously)

but the crazy thing we discovered is that the people who run news websites mostly don’t know where their ads are coming from, have forgotten how the ad system was installed in the first place, and cannot turn them off if they try

we actually shipped a server-side ad blocker, for a parter who had so completely lost control of their own platform that it was the only way to make the ads stop

anything can be reliable if you have good tests

instead of committing code, we should just save videos of all of the zoom meetings about the code

there’s a million small scale AI apps that just aren’t worth building because there’s no way to do the billing that makes sense. If anthropic wanted to own that market, they could introduce a bring-your-own-Claude metaphor, where you login with Claude and token costs get billed to your personal account (after some reasonable monthly freebies from your subscription).

But the big guys don’t seem interested in this, maybe some lesser known model will carve out this space

I was wondering when I’d see someone try this! I started work on a very similar idea last year but kept getting distracted by weirder and weirder ideas along the way, and never shipped anything. So, bravo!

I asked Claude a few days ago and it said it didn’t have access to the /clear command? maybe it was wrong or maybe that has changed

I guess I feel like if we’re rewriting device drivers then we’re in a turing tarpit. I think there’s room for innovation at what is traditionally considered the application level - we run git, postgres, document stores etc as applications. I think the way to solve the next generation of coordinating is by doing more interesting stuff on this layer.

I’m actually building an “OS” that’s up a level. it’s more like git, it has a concept of files but they’re documents in a distributed store. I can experiment with interaction patterns without caring about device drivers

I forked the anthropic Ralph Wiggum plugin: https://github.com/jes5199/chief-wiggum

there’s some debate about whether this is in the spirit of the _original_ Ralph, because it keeps too much context history around. But in practice Claude Code compactions are so low-quality that it’s basically the same as clearing the history every few turns

I’ve had good luck giving it goals like “keep working until the integration test passes on GitHub CI” - that was my longest run, actually, it ran unattended for 24 hours before solving the bug

I’m not sure it still makes sense to do OS research so close to the metal. Most computing is done up on the application level, and our abstractions there suck, and I haven’t seen any evidence that “everything is a file” helps much in a world of web APIs and SQL databases

I've been using Telegram bot to talk to a Claude SDK agent who talks to my Claudes via tmux commands (all running on a DigitalOcean VPS)

I'm running Claude Code in a tmux on a VPS, and I'm working on setting up a meta-agent who can talk to me over text messages

Cursor makes it easier to watch what the model is doing and to also make edits at the same time. I find it useful at work where I need to be able to justify every change in a code review. It’s also great for getting a feel for what the models are capable of - like, using Cursor for a few months make it easier to use Claude Code effectively

ride the BART

I think this might be the way forward, Claude is great at project managing.

I’m already telling Claude to ask Codex for a code review on PRs. or another fun pattern I found is you can use give the web version of Codex an open ended task like “make this method faster”, hit the “4x” button and end and up with four different pull requests attacking the problem in different ways. Then ask Claude to read the open PRs and make a 5th one that combines the approaches. This way Codex does the hard thinking but Claude does the glue

I find that ChatGPT’s Codex reviews - which can also be set up to happen automatically on all PRs - seem smarter than Copilot’s, and make fewer mistakes. But these things change fast, maybe Copilot caught up and I didn’t notice