jeremymorrell.dev
Practitioner with public evidence across Compute & infrastructure.
- AI signals
- 9 past 30d
- Sources
- 8 distinct domains
- Discussions
- 7 past 30d
- Latest signal
- 1h ago
Articles & links
I'm seeing a lot of AI coding ship-fast-now-and-ask-questions-later-tokens-go-burr thinking that leads to things like this github.com/anthropics/c... Sometimes the friction of not being able to magic code into existence is a Feature
- In Claude Code 2.1.198, AskUserQuestion auto-returned a 'No response after 60s' message and told Claude to proceed on its own judgment.
- The behavior was undocumented, missing from the changelog, and a regression from 2.1.196, which worked correctly per the report.
- Maintainer ThariqS said a release will expose the setting under /config with the timeout configurable and defaulting to off.
I can start to see how "getting enough compute to run the system you're developing with fast enough feedback loops" could quickly become the bottleneck openai.com/index/previe...
1. I had no idea we were building this 2. ... that works!? blog.cloudflare.com/kitesurf/
- Kitesurf runs entirely on Cloudflare Workers using V8 isolates, built specifically for AI agents rather than human users.
- Against Chromium it uses 3.1x less CPU and 4.7x less memory on screenshots, but wall time is 1.7-1.8x slower.
- Existing Puppeteer, Playwright, chrome-remote-interface and MCP/CDP clients work unchanged, and it is free during beta.
Once I learned about the idea of giving an agent a computer vs running an agent on the computer, I haven't been able to unsee it blog.cloudflare.com/cloudflare-c...
- Cloudflare released @cloudflare/computer, an open-source npm package that orchestrates agent work between lightweight isolates and full Linux containers.
- The runtime centers on a workspace: a virtual filesystem backed by SQLite that can be populated from cloud storage or source control.
- Cloudflare's stated goal is a runtime where a container is required for less than 10% of an agent's work.
Trying out liteparse for PDF parsing in Cloudflare Workers, and... that was really easy to set up! github.com/run-llama/li... I bet it will choke on complicated tables and such, but so far this is really nice
The title of this is a bit silly, but it describes the usage pattern with LLMs I find most useful, using it to think through a problem rigorously blog.exe.dev/claude-is-no...
Your users have suddenly acquired the ability to speak code into existence. What if your app gave them a safe place to run it? jeremymorrell.dev/blog/extensi...
Pulling on this thread, and you pretty much immediately run into needing a file-system abstraction separate from the execution environment, which is likely why everyone is building one these days github.com/cloudflare/c... docs.archil.com/compute/serv... docs.tensorlake.ai/fil…
Why are all the sandbox examples "run opencode in a sandbox"? Then I end up with many opencodes. I want to run one opencode / pi / etc and give the agent access to a sandbox modal.com/docs/example...
Recent commentary
If you are needing to expose functionality to a sandbox (container, vm, whatever), asking the LLM to "think through this from an OCaps perspective" generates *much* better designs that you might get generally
One irritation I have with LLMs is that they continually refer to previous revisions of the doc that I am currently having them write "Several conclusions in the first draft were wrong; corrections are called out below." Claude, no one read that. It doesn't exist any more...
Pondering the commonalities between 3d printing and AI coding - people kinda go a little nuts / obsessive when they first get into it - at first they use it pretty much solely to make accessories for the hobby, or silly toys - output can be useful, but may not stand up to real world use
Seeing LLMs run at 500+ tok/s feels like it breaks something in my brain. Seeing it run at a speed where I can read along makes it feel like thinking has a ~human-level speed. Fast! But still intelligible. Once it gets fast enough it just shatters that illusion. Computer go BRRRRR
Thinking through the work that I'm currently doing at my job, and I honestly think I would not have been capable of doing this work a few months ago without today's frontier LLMs That's pretty neat
Trying to vibe code an idea w/ Fable using the credits Anthropic gave a while back and holy guacamole do those $ numbers just keep ticking up I cannot imagine giving this to employees as an option in a dropdown and then paying API rates 😬
In jeremymorrell.dev's orbit
Center = jeremymorrell.dev. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.