Our approach is actually very cost-effective compared to alternatives. Our browser uses a token-efficient LLM-friendly representation of the webpage that keeps context size low, while also allowing small and efficient models to handle the low-level navigation. This means agents like Claude can work at a higher abstraction level rather than burning tokens on every click and scroll, which would be far more expensive
are your evals / comparisons publicly/3rd party reproducible?
If it's "trust me, I did a fair comparison", that's not going to fly today. There's too much lying in society, trusting people trying to sell you something to be telling the truth is not the default anymore, skepticism is
That doesn't work very well if your developers are on Windows (and most are). Uneven Git support for symbolic links across platforms is going to end up causing more problems than it solves.
It's why I wrapped my tiny skills repo with a script that softlink them into whichever is your skills folder, defaulting to Claude, but could be any other.
I treat my skills the same as I would write tiny bash scripts and fish functions in the days gone to simplify my life by writing 2 words instead of 2 sentences. Tiny improvement that only makes sense for a programmer at heart.
They probably organize individual accounts the same as organization accounts for larger groups of users at the same company internally since it all rolls up to one billing. That's my first pass guess at least.
I think it's more likely that their account was disabled for other reasons, but they blamed the last thing they were doing before the account was closed.
It reads like he had a circular prompt process running, where multiple instances of Claude were solving problems, feeding results to each other, and possibly updating each other's control files?
They were trying to optimize a CLAUDE.md file which belonged to a project template. The outer Claude instance iterated on the file. To test the result, the human in the loop instantiated a new project from the template, launched an inner Claude instance along with the new project, assessed whether inner Claude worked as expected with the CLAUDE.md in the freshly generated project. They then gave the feedback back to outer Claude.
So, no circular prompt feeding at all. Just a normal iterate-test-repeat loop that happened to involve two agents.
I think the idea is fine, but what might end up happening is that one agent gets unhinged and "asks" another agent to do more and more crazy stuff, and they get in a loop where everything gets flagged. Remember that "bots configured to add a book at +0.01$ on amazon, reached 1M$ for the book" a while ago. Kinda like that, but with prompts.
Could anyone explain to me what the problem is with this? I thought I was fairly up to date on these things, but this was a surprise to me. I see the sibling comment getting downvoted but I promise I'm asking this in good faith, even if it might seem like a silly question (?) for some reason.
From what I'm reading in other comments, the problem was Claude1 got increasingly "frustrated" with Claude2's inability to do whatever the human was asking, and started breaking it's own rules (using ALL CAPS).
Sort of like MS's old chatbot that turned into a Nazi overnight, but this time with one agent simply getting tired of the other agent's lack of progress (for some definition of progress - I'm still not entirely sure what the author was feeding into Claude1 alongside errors from Claude2).
This breaks the non-interactive mode the post want to achieve. Claude will not be able to install some things and will require user action, which is not desired here.
Like what? It can already use npm/pip/etc. And if it needs a new APT package or config in /etc/ then you would want to know because you need to document it.
Claude Code on NixOS feels like it has super powers. Being able to spin up a nix-shell with needed dependencies on demand gives it access to all sorts of tools I don't have or want installed on my base system. My "book-recommendation" claude code uses sqlite to manage my reading history and to-read and maybe-read lists but I never installed tools for sqlite and they aren't present on my NixOS desktop. It just launches a nix-shell with sqlite anytime it needs to read/modify the database. As long as the database file is within the directory claude code was launched from, it doesn't need to prompt for permission. With the caching that NixOS does, it's fast enough to not even think about.
I tried this approach for a while, but I really wanted it to be able to do anything (install system packages, build/run Docker containers, the works).
With these powers there's a lot less back-and-forth with me running commands, copying the output, pasting it to Claude, etc.
I'm sure you've had the case where you had to instruct someone to do something (e.g. playing tech support with family, helping another engineer, etc). While it helps the other person learn, it feels soooo slow vs just doing it yourself :) And since I don't have to teach the agent, I think this approach makes sense.
Running it remotely on a VM seems like a very sensible option. Just don't give it permission to nuke the remote repository hah (EG don't allow force-push, use protected branches, only allow write access to branches it created)
EU automaters fail at making modern cars. They just put a bunch of screens in there with awful software. If you go all screens, just commit like Tesla. If you can't beat Tesla, just stick with minimal screens and use buttons.
Somewhere between 2010 and 2020, most automakers went crazy with their designs and it went all downhill from there.
I have a 2020 Fiat 500 Abarth, and it is absolutely perfect: There is a screen (I think 7") for Android Auto/CarPlay/radio/nav, and every single other function in the car has a physical button. It is also absolutely gorgeous - pinnacle of design, IMO
Our 2021 Volkswagen e-Up is like this. There is a tiny(like 3" tiny) screen for the radio, bluetooth and reverse camera, everything else is analogue and has physical buttons. It's honestly best of the best Volkswagen design, what they did with their newer cars in terms of interior usability is a travesty.
I recently read a book of interviews with people who escaped from North Korea, and what shocked me was the discovery that the relatives of those who escaped are often executed (publicly) and that even children are executed in North Korea. We live in a terrible world. I mean... you expect a book from North Korea to contain terrible things, but somehow it was even worse than I expected.
Left wing thought doesn't contain any philosophy of limitations on state power, so under a left wing regime there is no limit to what it might do. No matter how terrible something is, if it can be imagined they will consider implementing it. To avoid that outcome there has to be an understanding of the flawed nature of government, and from that an ideological commitment to a state limited in power and role.
We're talking about North Korea executing children because their family members escaped. Doesn't get more left wing than North Korea. Not a good look to deploy a stock reply without thinking.
A great start is to have LLMs use special UNIX users that can’t do anything except that you allowed them to do, including accessing the database with a read only user.
reply