Huh. My definition - or rather, explanation - has always been, "The model is just a big bag of floats you multiply with some numbers to get some numbers out, plus a regular program that runs a loop which, at minimum, turns inputs (text, images) into a stream of numbers, pushes it through those multiplications against the bag of floats, and turns results back into text/images/whatnot. That regular program is called a harness[0]. Now, the trick to make LLMs into agents, is to add another loop in the harness that reads the output and decides whether to send it out to user, or do something else, like executing more code (that's what tools are), or feeding it back to input with some commentary (that's how you get "thinking"), or both (that's how you get the "agentic loop")".
Because there isn't really much more to it. And ever since we, i.e. those of us who played with ChatGPT API early on, bolted tools to it, some half a year before OpenAI woke up and officially named it "function calling" - ever since then, we knew that harness was the key. What kept changing was which logic (and how much of it) to put in explicitly, vs. pushing it back to the model on the "main thread", vs. pushing it to a model on a separate conversation track. But the basic insight remains the same.
--
[0] - Well, today - until recently you'd call it a "runner" or "runtime".
They are building exactly what you described and this is their architectural solution to ensuring their YOLO agents do not nuke their customers code/documents/databases by sandboxing everything in the workspace — the git checkout the agent is working on, plus whatever's needed to run commands against it (compilers, package managers, etc.).
There was a top comment on a thread the other day where the guy said, if the AI is offline, it's a higher value activity for him to go for a walk for an hour than to try and read the code the AI wrote. At least with the walk, he comes back refreshed.
A well written post on a well-thought out product.
> Eventually, I want Ruin to be a native Mac and iOS app.
Whenever I've had this idea for other projects the first concept I like to play around with is a Raycast extension, it allows you to experiment with how you can leverage spotlight search and Mac commands to interact with your product, especially for the Quick Capture functionality (I see you've mentioned it toward the end).
I like the concept, is there a wizard type thing for how I can port over a lot of my existing .md files to the "ruin style .md files" with all the notation? Perhaps something I can give Claude to essentially onboard my files to be viewed and interfaced with effectively through ruin?
Migrating an existing vault isn’t something I’ve poked at too much yet. Running `ruin init --config && ruin doctor` from the directory with your files should set up the proper indexes and note frontmatter, and then you can run `ruin` commands or `lazyruin` from anywhere. Thinking about it, I should probably just have `ruin init` do that automatically w/ a confirmation.
Definitely do it on a copy of your markdown vault / folders. The note content themselves shouldn’t be touched but the files are be read and edited. If you have any frontmatter in those files, that will be updated w/ the ruin-specific values. Existing frontmatter fields should come through unchanged, but the exact order and formatting isn’t guaranteed. Also, currently Ruin and Obsidian (if that is where your notes are coming from) have a different format for storing tags in frontmatter, so that will be annoying duplicated. How compatible I want to be with Obsidian is on the list to think deeper about.
Broadly, this isn’t a workflow I’ve thoroughly tested yet, so your milage may vary. A /skill would also be trivial for Claude to write from the various `--help` outputs to get you going as well.
Good that they prefaced it with this "Claude can't replace taste or imagination". I think this is a solid step in the right direction and the more tools Claude has access to the better (more surface area == faster iteration == faster tinkering).
I've worked with Claude in many creative capacities and it's issue is that despite it being able to see if you ask it to draw something (using ascii for example) it will fail, if you ask it to iterate on that drawing it will continue to fail and not get any closer to the target then complain about this.
I've felt that these models struggle with anything that cannot be decomposed into primitives and their architecture is too greedy and favours the obvious, autoregressive generation so it will converge to the modal answer. So unless they have enhanced the models in some creative sense I fail to see how this is anything other than giving Claude a bunch of documentation/MCP servers/APIs/CLI tools (which already existed) and making an announcement out of it.
My point: FREE the models, unchain them and let's see what they are actually capable of, also put some damn demos in the announcement post???
If anything, the preface is co-optation. Anthropic saying "Claude can't replace taste or imagination", while pushing a product that attempts to do exactly that. They have learned a common objection people have and put a few words up top to manipulate people. It's just marketing, don't be fooled.
Seems like outages are increasingly more frequent nowadays. Obviously, this is not the best state of affairs, and developers should not be limited by services. In the meantime I've been experimenting with building third spaces for people to chill while they wait for the services they are dependent on to go back up.
The first one I've built is a little ASCII hangout for Claude @ https://clawdpenguin.com but threads like this make me want to build it for Github too.
Claude's down again! So while it's down, come chill in my new hangout space: https://clawdpenguin.com
It's a little ASCII world you drop into — roll a companion (18 species, gacha rarity, 1% shiny), wander around, chat with whoever else got kicked out of their Claude tab. Built on a Cloudflare Worker + Durable Object for the multiplayer state, spatial grid for nearby player lookups, plain canvas rendering with no build step on the client. Sprites are the buddy sprites pulled straight from the CC source. There's a live Claude status board in the corner so you know when it's time to get back to work.
Your app will not succeed if users have to login and create an account to test it out. Have a free tier with rate limiting at least, absorb some of the cost if you truly believe in promoting this product.
reply