gavmor

I'm sure it already can with the right mod/plugin.

I've got an application for this. Will try and get some benchmarks in the next few weeks. We're definitely attempting to converge on a quality inference without wasting tokens/time.

For now, we're passing the first pass to the humans, and that's roughly 80 for 20, but to get the last 20 without blowing up our token spend? This might be the approach.

Thanks for the write up comment, btw.

The artwork "as it was created" is unavailable. The only options are Rougeoux or the ravages of time.

A regular LLM acts as a "policy," mapping a current state to a specific action (states → actions). Their new LLM acts as a "world model," mapping a current state and a chosen action to a predicted future state ((states, actions) → subsequent states). Instead of deciding "what to do," its explicit objective is to predict the exact environment observation that will result from the interaction history and the agent's current action.

I assumed at first that it was trained on synthetic data, but they actually went and deployed real physical hosts and virtual machines (e.g. Ubuntu, macOS, and Android) and browsers. They ran agentic systems on these continuously and recorded the actual, real-world interactions.

So it's an LLM that infers next state, or outcome,as structured data e.g. literal HTML code, UI view hierarchies, or accessibility trees.

The cold takes themselves are a fantastic proof of the care and attention the devs put into this game, as well as the depths of their design philosophy.

Almost nobody optimizes compiled assembly code.

Compiled assembly code is not an input to the next compilation; source code is an input to the LLM's next inference.

Sure, maybe "prompts are the code," but you must realize that code is also the prompt.

Well sooner or later I would expect a developer who intimately understands their code base to feel compelled to start refactoring and extracting fitting, meaningful well-leveraged abstractions.

Migrating to a workerd container is an option.

What's the data doing that's so sluggish? Is it being correlated/transformed, or is it just so many points of light are blocking each frame?

I wonder if Mollweide or Atlantis projections wouldn't be more performant?

Riding horseback saves time over walking but it's harder to smell the roses along the way. Nevertheless, millions now ride horses recreationally.

Who gets a sense of accomplishment from cheering for their home team?

The overall system that allowed this implementation is accountable. So why put such a fine point on it so as to exculpate the LLM?

He was the perfect protagonist for a teenage boy: a coward, an underachiever, technically a wizard but only on a technicality, and frequently the most powerful spell in the universe was lodged in his head against his will. This will be familiar to anyone who has been sixteen.

s/frequently/initially

Also, how is a cowardly underachiever "the perfect protagonist for a teenage boy"?

"technically a wizard but only on a technicality" is obviously redundant

And what part of any of this is supposed to be familiar?

It's just a strange essay.

That's sorta how Agent of Empires and also Zed seem to operate—worktrees as a first class aspect of the workflow.

Give credit where credit is due: Descartes, Kant, Putnam, etc.

That’s somewhat overstated.

We know anesthesia "works," and we know some of its molecular targets, but we do not fully know the mechanism by which it produces unconsciousness, ie whether anesthesia eliminates experience, or mainly blocks memory, report, and integrated neural processing.

Yeah, that's a scary way to lose 20 minutes...

Yeah, I suspect we're mostly quite wasteful of energy.

I wonder if they don't mean "between systems".

The terminal is keystroke-driven. It's character-selectable. It's reliable in a way that the GUI is not. When I drop frames, I can still enter the commands to rescue myself with some assurance they'll be interpreted, eventually.

I agree, a REPL isn't Unixy in the streams of text kind of way... or is it?

People use their tokens, and then complain of limits. Where's the incongruity?

Something promising I found is "DFlash DDtree Qwen3.5 & Qwen3.6 27B GGUF on RTX 3090" - https://github.com/Luce-Org/lucebox-hub

Didn't know 3.6 was available on Ollama outside of MacOS!

Something like a textual steganography?

Ursula K. Le Guin: 'The artist deals with what cannot be said in words. The artist whose medium is fiction does this in words.'

Nanoclaw mounts each agent's folder to the ephemeral container.

Hm, the page generated seems inconsistent with the usage of the original link.

Take a look at https://chatjimmy.ai/ -- it's running against Taalas' "hardcore" silicon model, ie a dedicated, ASIC-like chip.

I'm notorious for taking poetic license with naming—that's how we end up with `class Escutcheon`, or variables `recto` and `verso` where applicable in eg PDF generation.

But as much pleasure as I derive from novelty and specificity, my colleagues have oft expressed perplexity—whereas the terms which LLMs produce hew closer to the manifold (by definition!) and raise fewer eyebrows.

So, it has its turn.

I'm easily pushing 20 commits a day, but I won't pretend to have reviewed it all, let alone carefully. What I did was design it all carefully.

But, for some projects, yes—I still do line-by-line code review with a colleague.

Then again, a lot of my efforts are explicit refactor aimed at reducing LOC and tidying the codebase with, eg DRY.

The few times I felt lazy and let an AI do a simple function for me, all hell broke loose.

This is confusing, because LLMs are more than capable of implementing "a simple function." How did you spec it out?

The 'sad' fact is that you don't have to love the craft to make money selling the product. This has always been true to some extent.