andrewchambers

Thats basically what this guy did. He vibe coded a chat interface.

Yeah - people are hungry for it. They tolerate the kind of crappy docs and difficulties.

I think it was obvious, yet nobody seemed to have released a version people could actually easily use.

The feature set is pretty simple:

- Agents that can write their own tools.

- Agents that can write their own skills.

- Agents that can chat via standard chat apps.

- Agents that can install and use cli software.

- Agents that can have a bit of state on disk.

I do it with ssh and tmux. I suppose tools could make it better.

I think we agree - getting the llms to understand your intent is the hard part, at the very least you need well specified tests.

Perhaps more advanced llms + specifications + better tests.

Test suites just increased in value by a lot and code decreased in value.

Where did it say the tests need to be hallucinated ?

If you can make good tests the AI shouldn't be able to cheat them. It will either churn forever or pass them.

Mark Zuckerberg explicitly called out the airpod pairing being closed as unfair in a semi recent interview, maybe he can throw some dollars that way and get it all working nicely in some meta products.

Was playing around with the idea of p2p source hosting in package trees like nix and did a little weekend package prototype here of my own here:

https://github.com/magnet-linux/magnet-linux

Not really ready for prime time, but I think I have some interesting ideas there at least.

My point isn't really about what is correct or incorrect in this case.

My point is about making it so that you have to actively risk money to push the truth needle in the wrong direction.

I've seen a large number of comments online saying the shooter was a trump supporter - I don't really understand where that information comes from.

I feel like this is the sort of thing a prediction market might be able sort out.

Hoping for a polymarket bet so this can be resolved.

Kind of surprised by this take - I use openpilot often and also use claude code.

I kind of consider them the same thing. Openpilot can drive really well on highways for hours on end when nothing interesting is happening. Claude code can do straight forward refactors, write boilerplate, do scaffolding, do automated git bisects with no input from me.

Neither one is a substitute for the 'driver'. Claude code is like the level 2 self driving of programming.

It's more like letting a wet dog who you are watching closely quickly pass from your front door to the shower.

Author here - thanks for linking - Unfortunately I didn't have time to continue on this for a long while. I am certainly happy if people get inspiration from any ideas and continue on with it.

If you haven't tried using them then I am not sure your opinion on them is any good.

You don't need microservices for that, just factor your code into libraries that can fit into the context window. Also write functions that have clear inputs and outputs and don't need to know the full state of the software.

This has always been good practice anyway.

Using the term 'slop' to deride it for its lack of originality is pretty funny to me.

Thanks for MCP, I think it is sorely needed, it shouldn't take more than a single file shell/python script to teach an llm how to access a resource.

I wrote a general purpose delta debugger in python - https://github.com/andrewchambers/ddmin-python. Can be used to do the same thing.

Just don't take a phone with you or turn it off.

Ok, if we can define system 1 as retrieval and system 2 as computation and then we all agree.

What if working well means making them efficient enough to run more 'neurons' on our current hardware?

incorrect, this is a re-implementation of rust by designed to be included in gcc, not the original rust compiler.

You didn't read the part about compiling with the borrow checker turned off?

I guess this doesn't solve the bootstrapping issue.

If you are breaking the api it isn't that hard to add a new one instead of butchering the old one.

If it really can't be done then you aren't really shipping a library that is meant to be depended on.

I like the way python handled this situation - python3 for when you care, python for when you don't.

I feel like most of these problems just disappear if people would follow the same naming scheme sqlite3 uses. Just put the major version in the name and most tools just work out of the box with multiple versions.

I feel like your point would be better if there weren't already self driving cars around like waymo.