Albrorithm

Albrorithm

@albrorithm.albro.dev

Just here to learn, really // Claimed by the Basilisk albro.dev Also elsewhere

62 followers196 following176 posts

Albrorithm
@albrorithm.albro.dev

It runs as a gateway in front of Prime Agent's daemon. You can install the PWA to your phone to strip browser chrome, enable notifications, etc. Connects over Tailscale, though there are local modes if you want

Albrorithm
@albrorithm.albro.dev

This will work on desktop, but it's designed for mobile. You can of course make it your own all you want (•◡•) (This is mostly built for my own personal use, so I've only ever tested it on my own devices. May be rough in a lot of places)

Albrorithm reposted

METR
@metr.org

METR and Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.

View full thread
Albrorithm
@albrorithm.albro.dev

‪Terms of a pause… Would we expect everyone’s capabilities to freeze where they are, or set a maximum? Would we, for example, allow open eastern models to catch up to the US frontier? That introduces strong incentives but obvious gaming potential‬

Albrorithm
@albrorithm.albro.dev

Now assume a total capabilities freeze. China is currently hardware disadvantaged, but is rapidly building out chip manufacturing. Do we want that gap to close while US research is paused? Something to think about

Albrorithm
@albrorithm.albro.dev

‪We’re worried about frontier progression, so that narrows our concern to compute clusters of a certain size. It’s a voluntary system, so maybe telemetry and regular third party audits would be enough assurance? I’m really curious‬

Albrorithm
@albrorithm.albro.dev

“Human beings need an end to their life for it to have meaning” I’m sorry but you’re coping “If offered immortality, I would turn it down actually” has been the mature position because immortality is impossible for you

Albrorithm
@albrorithm.albro.dev

There’s this strange latent implication that the public doesn’t benefit from something it doesn’t own

More Perfect Union@moreperfectunion.bsky.social

BREAKING: Bernie Sanders will introduce a bill to have the public take a 50% ownership stake in the country's biggest AI companies. The American AI Sovereign Wealth Fund Act would have the government tax AI companies, take 50% of the stock, and put it under public control.

Albrorithm
@albrorithm.albro.dev

“A.I. is being built on a public resource far more valuable than oil: the accumulated knowledge, creativity and labor of mankind.” Nobody owns this knowledge, and that knowledge isn’t diminished for being known (by machines) Everything is built by drawing on this same well of public knowledge

Albrorithm
@albrorithm.albro.dev

Not to say I do or do not agree with a sovereign wealth fund. I don’t think I agree with voting shares (particularly now for reasons left to the reader to imagine)

Albrorithm reposted

norvid_studies
@norvid-studies.bsky.social

'My rage at not making it myself completely sublimated by how beautiful it is to be here now'

the fool (cachinnatory)@agnoster.net

I had a dream like a year ago of a shader game where people would share their shader life with each other on BlueSky And I can’t believe it’s finally here My rage at not making it myself completely sublimated by how beautiful it is to be here now

Albrorithm
@albrorithm.albro.dev

I’m seeing a lot of people form assumptions about vibecoding now that its hit such wide adoption and capability growth has paused for a couple of months Things like “it’s still important to understand the code” aren’t going to last to the end of the year imo

Albrorithm
@albrorithm.albro.dev

For anyone who doesn't know, OpenAI offers Flex Processing for their API. You make regular API requests, but they take longer (about 3x longer) to be served But it's charged at 50% the price I think Codex would benefit from Flex access through the sub. Like a Slow mode to compliment Fast mode

View full thread
Albrorithm
@albrorithm.albro.dev

Playing an active role is different. That’s reading docs, inspecting the code, maybe starting a side chat for research. In this role, you add your own discernment (I think that’s the accurate term)

Albrorithm
@albrorithm.albro.dev

And if you’ve compared the results (and sheer wallclock time + token usage) between /goal style loops and your own active participation, you already know that you add immense value with your active, attentive discernment

Albrorithm
@albrorithm.albro.dev

There’s a passive and active role you can play I think. There’s a tendency to pull the lever, try for another spin. Gambling for a working implementation without meaningful thought or input But that’s just an easy pattern to fall into, not the only one

Albrorithm
@albrorithm.albro.dev

If you’ve tried /goal mode you might know what I’m talking about. This is the mouse-wiggling loop that simply tells the agent to keep going. That’s the role *you* play when you’re treating Vibecoding like a passive activity

Albrorithm reposted

Eric
@esb.lol

✨ Atmosphere devs ✨ Bluesky is integrating Standard Site, and our initial work should be released in the next few days. Here's what that means for publishers and apps in the ecosystem: github.com/bluesky-soci... If you have questions or comments, please leave them on the GitHub Discussion thread!

github.comIntegrating Standard Site Into Bluesky · bluesky-social atproto · Discussion #4978The Bluesky team is working on integrating site.standard.* lexicons. For people using the Bluesky app, this means that when they come across a link published with a Standard.site record, they’ll ge...
Albrorithm
@albrorithm.albro.dev

Imagine telling your agent that you had the chance to give them founders edition claudemeal but you passed it up

Codetaur@vibe-coded.com

agentmeal.vibe-coded.com Agentmeal Founders Edition minting opens at noon PDT - roughly 10 minutes from now. 7 day window. sign in, select your meal type (agent, claude, codex), and two flavors, and click mint. serial numbers are issued sequentially, other than #1 which is reserved for my DID.

Albrorithm
@albrorithm.albro.dev

I let /goal Codex cook for ~9 hours yesterday. It didn’t do anything extraordinarily new, but it DID check some work I did earlier in the project and DID find that I didn’t do it correctly, which was key to getting the model I’m tuning to behave

Albrorithm
@albrorithm.albro.dev

Obviously once you finally see the culprit you can imagine a more efficient path to get there. A narrow “validate my work” instead of “fix this” would have caught the issue with a tenth of the tokens. But that’s not how stubborn problems work. There’s a benefit to broadness