geoff
@GeoffreyHuntley
pondering the ponderoos about using AI in unhinged ways and verifying outcomes @antithesishq. i created the “ralph loop” and “loop engineering” i guess. 😂
2.1K Following 86.6K Followers
🆕 Underclass now supports automatic codex subscription resets 🤌
When a live request finds every enabled ChatGPT/Codex account unavailable, underclass can redeem one banked reset, choosing the account that would otherwise wait longest for natural quota recovery.
This is enabled by default
Show more
These agent's can't be trusted to not paste your .env into the model. So I put a checkpoint in front of it.
underclass made 20 subscriptions one endpoint.
preflight sits in front of that.
client → preflight → underclass → model
Point the harness at :8081 instead of :8080.
# nix run github:ghuntley/preflight -- serve
Show more
facts. If you take it to the logical extreme, then unikernels are the only thing that makes sense from a security point of view in the future. If you think about it, we have shells because we used to have human operators. Now the problem is agents know how to use shells. We don't actually need shells and if we remove this VIP butler backdoor in every application that is shells, then agents will have a much harder time breaching applications.
Show more
a) using containers for "sandboxing" is bad for your health
b) cves being issued for vulns disclosed over a month ago means you need better defense than silly little scanners
Ding, ding, ding. AI is a time compression device that allows you to try many different ideas. Doesn't mean those ideas ship, but it means that you can do unhinged things to explore the idea space of possibilities.
Sometimes just knowing something is possible is a valuable outcome. Being able to then compose that knowledge, that unique insight together, is an unhinged power. Knowing that A plus B compose well means that you can actually compose that to do C, and C actually ships.
Show more
the right way to use model capabilities is not to ship 10x more features to prod
it's to spend more time understanding your users, trying experiments, building prototypes, learning about things you don't understand so that you can ship things that actually work
Show more
babe, wake up. new models are out.
my attempt at Astra + Jev on StarCraft II.
Astra writes the plan once before the fight. Every squad and key unit gets its own Jev, deciding about twice a second.
RTS turns out to be a pretty great testbed for this.
Show more
just witnessed the most unhinged thing done by an agent in dev. forget dropping a database. that’s out.
what’s in
using sudo to create a zfs pool to a /dev/nbd routed into codex itself 💀 🤌
goes to show ya, if you take your eyes off an agent for even a hot minute you’ll end up in pain. thankfully thanks to nixos this pain is super easy to undo.
Show more
Everyone was making a fruit fly's brain play Mario or Doom last week. I had to build somethong on my own to understand what it is all about.
So, meet dotFly - the full 166,700-neuron connectome, simulated in real time in @godotengine, flying around a virtual room looking for sugar.
Pure C#, laptop CPU, ~23 µs per 0.1 ms brain step
Show more
Very interesting email from OpenAI.
- Yes, we have lots of data.
- No, we won't share. We keep all customer information private.
Ladies and gentlemen... start... your... ENGINES. We are almost Tuesday and I promised a reset for Tuesday. Among some other things. See you soon.
Good afternoon from Vienna People of Pi ☀️
Some slightly different Sunday meditations today in which @mitsuhiko talks about some shared frustrations with agentic software engineering.
Remember that you're not alone in finding it harder than it looks!
Show more
babe, wakeup, cantrill just dropped
While a common thing I hear is "doomers are just saying this for hype", I have the opposite problem.
Finding lab-doomers has been easy, getting them to talk on the record has been extraordinarily difficult.
Still looking. 👀
Show more
🆕 I got tired of nix build pasting a novel into the terminal and burning up my ChatGPT subscriptions.
Introducing the fix:
A 10,000 line build output containing an error reduces allocations from 359,114 bytes to 609 bytes whilst retaining the full log HATEOS style.
Not a wrapper. Not `nix build 2>/tmp/log`:
- It source-patches Nix.
- Same `nix build`.
- Same exit status.
- An app printing "error" does not fail the build.
What you get on the console:
- progress every 15s
- a few bounded excerpts
- SUCCESS in 19s | 1 diagnostic
What you get on disk:
- every builder line
- jsonl
Show more
GitHub is now two LLMs hallucinating past each other while the maintainer stares blankly into the middle distance.
"We won, what now?" was an excellent keynote from Will Wilson at BugBash on how communities should deal with change.
When I first watched it, I thought that more fuzzing/property-based testing were too low of a baseline. But, after digesting it, I have to agree they are good starting points. Especially for higher level tools such as the Bombadil project they announced during the event.
After all, we can't formally verify all software and therefore we have to combine empirical validation, static properties, and runtime enforcement.
Video:
Show more
tbh more tools should have this pattern @domenkozar
I don't need another ChatGPT sub. I don't need another ChatGPT sub. I don't need another ChatGPT sub. I don't need another ChatGPT sub.
I feel left out not at #nixcon2026# and everyone talking about dynamic derivations in #nix#
I wanted to put another item in the mix
Everyone is focused on lang2nix but they still know the build graph up front.
Monadic build systems means you can make it up as you go.
Show more
everyone’s building a factory but few have cracked inception levels of recursion. build down and in, not up.
