You can really see the limitations of qwen3.5:9b in reasoning traces- it’s fascinating. When a question “goes bad”, sometimes the thinking tokens are WILD - it’s like watching the Poirot after a head injury.
Example: “what is the air speed velocity of a swallow?” - qwen knew it was a Monty Python gag, but couldnt and didnt figure out which one.
As a person who also knows there's a connection between that phrase and Monty Python and not much more information beyond that, I'm not sure how to feel.
It thinks like it’s memory is broken and it’s unaware of it; over 100 lines like this:
- Wait, no, that's not right either.
- Let's recall the specific line. It goes like this:
- Knight A: "How can you have a swallow?"
- Knight B: "It is the air speed velocity of a swallow."
- Actually, the most common citation is from the movie where they ask an expert on swallows? No.
First you have to agree that Claude Code might be useful for some non-repo task, like helping with your taxes or organizing your bookmarks.
Next, consider how you might deploy isolated Claude Code instances for these specific task areas, and manage/scale that - hooks, permissions, skills, commands, context, and the like - and wire them up to some non-terminal i/o so you can communicate with them more easily. This is the agent shape.
Now, give these agents access to long term memory, some notion of a personality/guiding principles, and some agency to find new skills and even self-improve. You could leave this last part out and still have something valuable.
That’s Openclaw in a nutshell. Yes you could just plug Discord into Claude Code, add a cron job for analyzing memory, a soul.md, update some system prompts, add some shell scripts to manage a bunch of these, and you’d be on the same journey that led Peter to Openclaw.
My conspiracy-theory assumption has been that DOGE seeded X.ai, and the newer govt contracts are going to continue that. X.ai won’t need to be as smart as GPT7 when the contract for killbots goes to RFP, it’ll already know everything about you including location and weaknesses.
Kidding with the killbot comment; more likely it’ll be used for insurance denial, employment screening, benefits and taxes etc
Have you considered any other options for metadata storage besides the sidecar file? If this is an agentic skill, I would imagine that other methods could be similarly effective - like a cli “getattr —filename” or even “sqlite3 “select metadata where filename=“.
Given agent context strategy is converging on “filesystem is best”, we would need more robust metadata facilities . You have a good opportunity here to test which methods are most effective - starting with your sidecar file.
This seems a natural evolution of Raycast Extensions (which are an evolution of Script Commands) - given the current landscape (generative everything). I would be surprised if there’s no “Raycast inside” within and around the new toolchain.
I’m torn about what this likely means for iOS; while I do want to do Raycast-y things in my phone, I’m not sure there’s enough of us to make a business out of it.
As a Raycast paying user, I was a little bent that they have apparently not been focusing on the core product.
However, having just vibe coded an actually useful Raycast extension, I can see wanting to bring this capability to a wider audience - and how this could scale their core product adoption beyond “nerds who think Spotlight stinks”. Which is getting
A lot of good (if negative) comments ITT though; it’s going to be tough for them to bring this to market safely.
There’s some sort of underground genai SEO movement happening. I am not sure how it works but I have been examining source quality from Perplexity and ChatGPT and finding the same sources over and over. I’ve found quite a few “gamed-looking” sources recently, basically similar to trash that have dominated Google serps for the last few years.
Edit: this is a ridiculous question, I know. Trying to eat my dogfood so to speak
Does Tailscale maintain an q&a agent, mcp, or llms.txt that anyone is aware of?
I’m trying to use Tailscale across my personal networks - without investing a lot of time - and so I’m throwing agents at it. It’s not going well, primarily because their tools/interfaces have been changing a lot, and so tool calls fail (ex ‘tailscale serve —xyz’ is now ‘tailscale funnel ABC’ and needs manual approval, and that’s not in the training set).
For one, qmd uses SQLite (fts5 and SQLite-vec, at least at some point) and then builds reranked hybrid search on top of that. It uses some cool techniques like resilient chunking and embedding, all packaged up into a typescript cli. Id say it sits at a layer above Wax.
Example: “what is the air speed velocity of a swallow?” - qwen knew it was a Monty Python gag, but couldnt and didnt figure out which one.
reply