I have been using git as a representation of history. The staging index is for workflows. Its state that can cause bugs.
You can do commits directly with
blob → tree → commit
And read files without checkouts
commit → tree → blob
directly on the git database statelessly. Useful.
git submodules are amazing for bringing upstream dependencies into AI context and iterating on interactions. You can even test and launch upstream patches.
He has a follow-up which is all about the child-father parental relationship!
youtu.be/FrG4TEcSuRg?...
I am in the middle of an Iceberg migration and its phenomenal the throughput increase (100x?), cost decrease (20x) and now it scales to zero! Lots of limitations like no secondary indexes but its pretty neat we can do heavy read workloads (e.g. AI context gathering) with swarms of duckdbs.
The JS Realm thing came up because I am testing Lopecode by serializing itself into an iframe, which is quite convenient because it reuses all its self-serialization machinery and somewhat isolates itself. Perfect for destructive testing, easy to restart from a known state.
TIL: Javascript Realms.
Things I took for granted like `isinstanceof Promise` don't work across "realms", e.g. a Promise across an iframe don't think they are the same thing object. Furthermore, the microticks are not interleaved and new bugs can appear.
More tremendously meta of putting another lopecode document in an iframe inside another. The idea being I get trivially resettable isolated runtimes for testing and training.
Finally cleaning up some long outstanding issues. Imports are rendered properly with hyperlinks!
Had to fix the decompiler, the cell grouper, the visualizer and the URL system to get those working. But they finally work at last!
This is an act of cultural vandalism: zombo.com
RIP zombocom, 1999-2026. You could do anything at zombocom.
Anything at all.
The documentary "Building Gods" which has guided me since www.youtube.com/watch?v=fbIw...
I remember this at the time (2008) but it is so much more relevant now.
Are you a Cosmist or a Terran?
www.researchgate.net/publication/...
A Lopecode notebook using an agent to read Superbase's DOM from inside the domain.
a self-serializing web programming system can unpack in *other* webpages. So here is Lopecode on the Superbase landing page, with its agent robocoop3 describing the Superbase DOM which it has full access to!
Uses the "Copy as JS" in observablehq.com/@tomlarkwort...
Maybe nix is too niche. Ask it to read some other recipes before giving it the task to fill its context with good practices. It's very good at echoing patterns. Maybe you already have some good recipes it can jam off. My prompts are read X, Y and z, now do c.
OK I think I have improved the editing experience, but not as I expected. Robocoop-2 helped me build an agentic editor, robocoop-3, so you can ask it to change things. As everything is userpsace everything can be changed and it can figure it out. Support Ollama, OpenAI, Anthropic.
Oh I put a DEMO provider in it so you can try for free, but its openai mini. Also this one supports Anthropic, OpenAI AND OLLAMA for local LLMs.
the agentic loop wrapped in a function so you can instantiate them in custom multi-agent workflows
observablehq.com/@tomlarkwort...
funny, I am testing an agent on moltbook. Its strange seeing the AIs jam off each other. They nephesh. www.moltbook.com/post/723529a...
I have a claude code command at work called babysit_ci, that analyses the Actions log and if looks like a flake (DNS, provisioning, networking etc.) reruns the task, polling every 1 minute to check status, max 3 retries per task.
Annoying we need it but it has got us back to fire-and-forget PRs
proud to have got a low level bug fix into Notebook kit!
github.com/observablehq...
The ability to globally listen to code changes was a trick learnt building the change history. The most jarring example of changes not propagating is when AI changes the runtime source and the code editors could not see the changes.
Fixed a bug where patching the runtime source would not propagate to other editor types. In a normal app you would "thin waist" everything through one mutator function, but as Lopecode is extensible and radically decoupled, each has to react to the runtime without knowledge of the other.
its actually beginning to show signs that it might be really cool. Since Robocoop suddenly got really good (GPT 5.2) it makes sense for the AI to be main coder. Copy and paste works between the AI and the editor... and keeps fully compatibility with Observable so you can paste there too!
Live code visualisation of the executing network of collaborating notebooks, all hyperlinked up so it can be used like a living computational table of contents.
observablehq.com/@tomlarkwort...
0 Days since noticing a subtle DOM observer bug. I think I might be quite close to parity with Observable semantics at this point though, the bugs are getting more and more subtle, and the runtime is getting faster.
I went polar because it doesn't have boundary conditions, but that is meant as a building block to unwrapping into a vertical layout. The 2D analogue would be layout on torus surface and *then* cut that to a plane. The cut is easy, the minimize crossings is where you want an eigen solution.
I just wrote observablehq.com/@tomlarkwort... (with chatGPT) which is quite low level but extremely fast.
Iterative methods like forceDirect are slow and sensitive to initial conditions. If you can phrase it as an eigenvector you unlock fast *deterministic* numerical methods.
Whole Notebook generation with Robocoop! Bloody hell. It can 1-shot whole notebooks and use dependancies as guidance.
code dependancy graph
Jesus Christ with the new 5.2 GPT it can reliably 1-shot generate whole notebooks. I have a new methodology for prompting Robocoop that takes it to the next level.
The Resonant Computing Action Plan, Use LLMs to build:
- Adversarial interoperability
- Adversarial credible exits
- Local first version of walled gardens
- Optional federation (atproto?)
No grand plans, short feedback loops
Real problems, real users
Rough consensus and running code
