Yasir

A quest to build great software
Sharing the journey openly, tag along.

[Yasir @aphronio]

hooked up Pebble ring to my AI agent and i reach out to AI so much more even for small things.

This feels especially nice for async things like capturing ideas while walking or driving, controlling lights in my home with natural language. A glimpse of what siri could have been 😑

talking to agent without having to pull up my phone feels very natural. It is not zero friction yet, the main friction point still is to get replies back, I am routing them to telegram agent thread for now. Clunky but woks nice.

But this gives me a sense of how important the interface for AI is, now that the models are more than capable for most things. I have realized that as the models got better, productivity gains were their but i am spending so much more time glued to my screen staring at the agents at work. It is a delightful sight but also addicting like gaming at the same time. And I have been looking for ways to use AI in more of afk manner. So this is something i will be exploring henceforth.

True hardware integration can open up myriads of ways to interact with and use AI. Follow my friend @isausmanov for hardware first AI experience. He's doing cool stuff in the space and has a lot of fresh ideas on interface as well.

[Yasir @aphronio]

finally arrived

[Yasir @aphronio]

@beyang I am curious as to why you went with the latter? Is detail better in some ways than custom orb setup would be?

[Yasir @aphronio]

finally arrived

[Yasir @aphronio]

local and remote incus integration is now supported by https://t.co/hCRGrI99We

[kitze 🛠️ tinkerer.club @thekitze]

you can do oRbS on your own vps

let me point u to a direction

have a good day

[Yasir @aphronio]

@mitsuhiko @khoiracle What do you mean by putting durabilty on top? Will the pi 2 be using absurd for durability?

[Yasir @aphronio]

@ahmetkdev @dbmikus @amikadev @ahmetkdev You might want to have a look at dorf. Still in early stages, currently supports:
Harness: code and pi
Sandboxes: e2b for cloud and incus for local
Subscription + api support for chatgpt

https://t.co/hCRGrI99We

[Yasir @aphronio]

dorf rewrite in go + absurd by @mitsuhiko (thanks for permissive license)

absurd was such a delight to work with and saved me so many papercuts, AI is still far from writing durable workflow engines from scratch.

[Yousafzai @theycallmebrute]

@thdxr I implemented something very useful for our team and I am writing a white paper at the moment on it, I am going to go opensource with it but I genuinely needed some help regarding some resources (genuinely super small for you)

[Yasir @aphronio]

I have a feeling that codex in websocket transport mode burns through usage much faster than sse. I have to test it properly to be sure, will report back.

[Yasir @aphronio]

@dbmikus Tried to make it work many years ago. If my memory serves, subtree was the better of the two. But in my opinion both are bad primitives and pain to manage. Monorepo works best if you have that liberty

[Peter Wang @BrainsAndTennis]

whenever something is crazy hype and is followed up by an armada of PR, you gotta bring out the bullshit detectors

so i cloned the repo and did a little digging and found a lot of sound and fury. please enlighten me if any of these facts are untrue, would love to be educated

RLM claim either overblown or false

this doesn't even appear to be RLM (https://t.co/YypXTL4Dx0)? RLM is "cool" as a concept because there's a premise of going infinitely deep to do arbitrary decomposition of a problem. in practice though it's not really hard, you just make subagents callable as functions and wrap that in a code execution tool. the reason RLMs have not been productionized and is just seen as a research ditty is that 1) going arbitrarily deep down the stack is a shitty thing to do in production without guardrails, 2) most problems do not need more than 2 layers of decomposition, and 3) agents are bad at banana phoning each other and will fail to preserve the subtleties of human requests, leading to a lot of inefficiencies and bad behavior.

to solve these problems, there needs to be a fair bit of innovation, either on model training end or on harness end, to mitigate these issues. so what were the innovations in prime agent?

they set RLM_MAX_DEPTH = 1? ... ok what? so theres no innovations, it's just calling subagents like every other harness in existence. if RLM MAX DEPTH is 1, it's not even RLM? it's just... a harness?

so now i try to look up to what depth the ARC AGI 3 hillclimbing was done at. not surprisingly this doesnt seem to be disclosed on the blogs / posts. if it's just 1 layer deep, that's just shady marketing.

a real innovation or contribution to harness engineering would be to detail the things that were improved to get a max depth of >3 to work without blowing up your computer, wasting a bajillion tokens, taking forever, or having agents go off the rails. i didnt find any

97% is just overfitting after hillclimbing on a public eval

ok folks. it's nice that opus5 does better than sol than terra, but at the end of the day, we're looking at a public eval in which problems and solutions are open access, so any agent harness can arbitrarily overfit to the task set however much it wants to. if you do a codex or claude loop with some semi shady prompting to just get 100%, im pretty sure it will too. now obviously this is almost certainly not what the prime intellect team instructed their agents to do (otherwise you'd see that terra get 100% after a couple iterations), let's just all remember that there is no train / test split, and that ARC explicitly says

- Public-set scores are vulnerable to task-specific overfitting.
- They are “emphatically not” valid evidence of progress toward AGI.
- Real generalization should be tested on the 55 semi-private or 55 fully private games.

It's difficult to assess the degree of overfitting in Prime Agent's self-improving harness (which, btw, is not terribly different from nous agent or really any self-improving system that just looks at old JSONL agent trajectories and suggests skills / memories / system prompt appendments / tools / extensions). the more overfitting, the less impressive the 97% performance is.

a counter-example that i found to be legitimately impressive was https://t.co/13nt4TJ8pS -- minimal / almost no overfitting, stupidly simple and general solution, and 95% on ARC AGI 3.

i think the fairest thing to say about this system, given it's independent daemon system for managing subagent lifecycles, is that Prime Agent contributes a fairly robust persistent asynchronous agent-process tree. It does not demonstrate a solution to scalable deep recursive agency

[Yasir @aphronio]

jit 🥳

[jack @jack]

bring your own harness to buzz!

[Isa @isausmanov]

You screenshot or bookmark posts as a reminder. I make my agentic phone do the work right away.

It already sees what I see. I can stop on a useful post and ask the phone to look into it, or when I find a product I want, I ask it to find where I can buy it for less.

Then I keep scrolling while it works in the background.

[Jason Zook @jasondoesstuff]

I feel like I'm running out of human RAM. 🥴

More recently than ever when I'm doing focused work (read: vibe coding), I feel an intense amount of mental friction if I get interrupted.

Like, my wife will ask me a question while I'm working and it literally feels like my brain is screeching to a halt and it's hard to answer her.

I don't think it's actual prompting and chatting with an AI that's the problem.

I think it's that AI has made so many tasks easier. I can fit more tasks in during the day, but my brain hasn't actually been able to adapt to all the new inputs and decision making.

Writing an email newsletter? That used to take hours. It now takes 15 minutes or less. I can knock one out while Claude implements a plan.

Designing a feature mockup? This used to take days. I now can feel myself getting frustrated if it takes more than 30 minutes.

And it's all the small time gaps waiting for Claude/GPT that I fill with other tasks (because, what else do you do?)

The worst part of all of this? The work is SO FUN right now. I can't imagine doing anything else. The possibilities are endless, and ideas come to life nearly instantaneously.

But... I wonder if it's taking a toll?

Are we all pushing our brains further than they've had the time to adapt to?

Maybe it's just me, I'm the problem, it's me? 😄

[Yasir @aphronio]

Just renewed my chatgpt subscription yesterday to try out sol in codex.

First impressions so far:

It seems to be much more handsoff compared to 5.5 and more token efficient.

I have been using it at low reasoning and haven’t felt the need to bump it up for my usecase yet.

[Isa @isausmanov]

The agentic phone that won OpenAI Hackathon is replacing my iPhone feature by feature.

Found Amjad Masad’s contact, emailed Paul Graham, and called on my behalf to reschedule a meeting.

Its KPI is how little time I have to spend on it while getting more done.

[Isa @isausmanov]

I made the finals at OpenAI Hackathon.

Built Agentic OS for a Phone. Solo.

The way we use mobile devices is going to change forever.

Please support with your vote.

[Yasir @aphronio]

@fin465 SEND