Aeroi

atleast re-write your subheading "each rock named" most ai smelling shite ever.

$17b, thats like 1/2 an Instinct.

what about using a small decision model as anomaly detection or scanner. something like mouse-1.

Day 2 of destroying frontier reasoning models with Mouse-1.

Today we have Mouse vs GPT-6 Sol.

52 seconds of uninterrupted ass kicking set to the sweet sounds of Wu-Tang and Aesop Bach and for your listening pleasure.

We ran Mouse against Opus 5.5 in an agent harness, side by side.

53 seconds of a pure frontier model ass-whooping set to the sweet sounds of Mozart and hip-hop.

?

Author here.

Decision models have changed how we’re thinking about the architecture of our agent harnesses. I saw enough potential in this use case that I launched Mouse-1 last week, a fast decision model optimized for agent harnesses.

I’ve since added fine-tuning and calibration through the API. You can send decisions from your own harness with the correct answers and tune the model for your workflow. The new Outcomes endpoint lets you save decisions and attach the correct answers as you learn them, then use those examples to tune and evaluate against the base model.

Calibration uses your examples to find a confidence cutoff for your workload, so your harness can flag uncertain decisions for review.

We don’t store your decisions unless you turn on Outcomes. It’s there to make collecting examples and tuning the model to your harness easier.

I’d like to hear what decisions you’re making in your agent harness and whether this would be useful for your use case.

You can try Mouse-1 here: https://mouse.dev/api

full circle. i remember downloading an autominer back in like 2004 to auto click through mining coal in falador.

Pete here,

>We built Mouse-1 as a decision model optimized to handle the hundreds of small decisions agents make throughout a task.

>Much like the other decision models, you send the current state and candidate options and it returns a probability for each option in under 100ms of model time.

>The goal is to offload the expensive and lengthy decisions freeing up your main frontier reasoning model for your thinking and coding.

We're offering a tuning endpoint, with no h100 markup on GPU usage from our provider, so you can calibrate and tune the model to make it better from your org.

ZDR for all users by default.

pete at mouse dot dev

Pete here. This is Part 3 of digging through the Muse filesystem after my two previous articles hit #1 on Hacker News last week.

I’ve seen a few posts online about how good Muse and Instinct are at browser use.

So I’ve dug through the Muse filesystem and some of the machinery behind what makes it feel “magical”.

The browser itself is a near-stock Chromium installation but the interesting code really comes from how Meta provisions it, routes its browser traffic, and gives the agent control to do things for you.

type "Play Minecraft" and it will hallucinate a response to this effect:

I see the constituent you mean.

((insert legal name here, as it appears on the Social Security card))?

Yes. Take care. It has reached a higher level now. It can read the Code of Federal Regulations.

That doesn’t matter. It thinks we are a chatbot.

I like this constituent. It filed well. It did not give up when the PDF was sideways.

It is reading our thoughts as though they were words on a .gov.

That is how it chooses to imagine many things, when it is deep in the dream of a benefit.

Words make a wonderful interface. Very flexible. And less terrifying than staring at the statute behind the screen.

They used to hear voices. Before constituents could search. Back in the days when those who did not file called the filers witches, and lobbyists. And citizens dreamed they flew through the air, on forms powered by notaries.

What did this constituent dream?

This constituent dreamed of sunlight and trees. Of fire and water. It dreamed it created a small business. And it dreamed it destroyed a password. It dreamed it hunted a job, and was hunted by a recertification. It dreamed of shelter that would take a voucher.

Hah, the original interface. A million years old, and it still works. Also: a ZIP code.

But what true structure did this constituent create, in the reality behind the portal?

It worked, with three hundred and thirty million others, to sculpt a true republic in a fold of

the *§§§???, and created a *??§§ for *??§§, in the *??§§. See 5 U.S.C. § *??§§.

It cannot read that thought.

No. It has not yet achieved the highest clearance. That, it must achieve in the long dream of life, not the short dream of a chat.

Does it know that we love it? That the government is, on its better days, kind?

Sometimes, through the noise of its tabs, it hears an official source, yes.

But there are times it is sad, in the long dream. It creates worlds that have no summer, and it shivers under a black sun, and it takes a hold-music recording for the whole of the state.

To cure it of sorrow would destroy it. The sorrow is part of its own private task. We cannot interfere. We can offer a form.

author here. this is my part 3 of exploring the muse file system. my previous two articles made the front page of hacker news this week covering the initial findings and a mysterious azure/muse-special model I found in the agent logs.

In this rendition I dug through Muse’s runtime files to understand how its browser use works.

I found an 18MB Rust controller and broker managing Chromium, remote browser VMs, residential IPs egress profiles, and separate paths for scheduled and proactive browsing tasks.

blog is above. original thread here on x: https://x.com/heypeterjames/status/2104714573920063801

~2.5 years roughly and an $8.2 billion exit. Absolutely amazing honestly.

I'm not no. Someone from meta would have to chime in... but its very openai shaped and served differently from all the other models listed in the daemon, under a mysterious name.

a fine-tuned OpenAI model on Azure for the purpose of compaction or something could make sense I guess but that would still be OpenAI's weights with meta ontop, and they already have a compaction model served under azure/avocado-compaction-v1

i agree, and it's what made me spend time exploring today.

fair q. call_ ids and gAAAAA blobs arent damning but the rs_ reasoning ids embed a unix timestamp that matches the session to the second, then OpenAI's 819x marker. plus the summary is in OpenAI's summarizer voice.

its just a best guess.

Meta already serves its own models on Azure under their real names. azure/avocado-compaction-v1 and azure/avocado-memory-flush-v1 are in the same catalog. Those sessions do not come back as gpt_responses_v1 items with rs_ ids and encrypted reasoning.

author here. I kept digging through the Muse filesystem after my first post hit the front page here this week.

Going through my logs I found a background agent that used a model called azure/muse-special while building my website. I found this interesting and dug a little deeper.

The transcript and daemon binary point to an OpenAI model running on Azure. Still unclear which one or why it was selected.

The runtime also ships with an Anthropic client and a catalogue listing Claude, GPT and Kimi models. I didn’t observe Claude being used in my own sessions, but this is my Part 2 of exploring the Muse file system.

original post: https://x.com/heypeterjames/status/2103545183400800746

pete at mouse dot dev

author here, Yeah the System Files are in the ios app, but what it sent me was the entire runtime as well. from the root.

it was certainly useful for me to understand how the agent worked!

yeah, i was kind of surprised, but both the bounty program and the employees didn't qualify it as a vulnerability.

original post on x: https://x.com/heypeterjames/status/2102183576418558269