Addison Omead Amiri

building on machines I can touch.

[Richard Hutton @richardhutton]

@jakemor There are two types of people: one who has their phone always on them within arm's length, and the other who asks, "Where is my phone?" multiple times per week and they marry each other.

[Addison Omead Amiri @addomam]

After using OMP for a while, it's clear that Codex is doing effectively codemode with shell scripts. Every shell call is a bunch of commands together with tailored output. I wonder if it's another one of those things where better models with plain shell scripts can figure it out.

[Addison Omead Amiri @addomam]

@ProgrammerDude Claude has decided that the sandbox works as intended.

[Addison Omead Amiri @addomam]

@round @grok what is this?

[Addison Omead Amiri @addomam]

@IceSolst Intellectual property is theft

[Addison Omead Amiri @addomam]

My goal here was to find an open source stack that I could autoresearch on. That's my next project. Most the features in herdr are limited, the tools in OMP complex, the skills in bigpowers huge... but can it be a solid enough foundation to hill climb?

[Addison Omead Amiri @addomam]

It's been over a week since I switched to herdr + omp + bigpowers for everything. I think this is going to be my baseline for the time being.
Details on each if you're curious:

[Addison Omead Amiri @addomam]

But for any codebase that's going to require more dev than a one-shot, I'd use bigpowers. It'll be way slower cause it's keeping documentation of everything along the way but at least it can keep track of what's going on and continuously (most the time) make forward progress.

[Addison Omead Amiri @addomam]

@ProgrammerDude This is super nice. Bootstrapping a mac VM to the point that you can hand it off to configuration management has been such a pain for so long.

[Addison Omead Amiri @addomam]

Lmao is this real?

[Addison Omead Amiri @addomam]

@reatools how do yall get around being banned for hacking?

[Addison Omead Amiri @addomam]

My agents can barely look at strace without triggering the OpenAI cybersecurity classifier. How are people running REA? Is everyone just using Chinese/open models?

[Addison Omead Amiri @addomam]

Next step: booting this on bare metal

[Addison Omead Amiri @addomam]

Someone with a large token budget should take sel4 and run a Linux vm with hardware passthrough bridged through a minimally visible sel4 IPC to a Linux vm running user workloads and then progressively chip away at both sides until you’re left with something you can formally verify e2e

[Addison Omead Amiri @addomam]

@ryanrhughes @jondkinney @tiredkebab @maralcbr @grok can you link me the pr?

[Addison Omead Amiri @addomam]

@dhh But I thought I was a lunatic and if you're really doing this get a whisper model and accessibility scraper and let it run 24/7. You're effectively livestreaming to the model and everything you say gets updated in the project tracker automatically.

[Addison Omead Amiri @addomam]

@dhh You gotta add Mumble support to monologue at this point.