Andrew Atiya

Impatiently curious. Building @knowhow_app

[Andrew Atiya @atiya_ag]

Multiplayer AI is the future, rooted in the past. I have never been more excited about building. I want everyone to have this.

My thoughts, here:
https://t.co/yrh1JFrrZI

[Andrew Atiya @atiya_ag]

https://t.co/UxEaeUzxnb

[Andrew Atiya @atiya_ag]

@kayacancode And clarity about what will happen and when. But it's the future :)

[Andrew Atiya @atiya_ag]

Lots of talk about company brains. Is company knowledge in documents?

Is an LLM reading markdown a brain?

[Andrew Atiya @atiya_ag]

@ThePrimeagen I use it to pentest my attention span

[Andrew Atiya @atiya_ag]

@ivanburazin The hardest thing: be the first to admit when there’s a problem, and change your mind.

[Andrew Atiya @atiya_ag]

@wjessup 🙋‍♂️

[Andrew Atiya @atiya_ag]

@uzairansar Also banned: [company ] killed, ended, destroyed, cooked [huge thing]. When did our reaction to banal press releases become sheer panic about terminal violence.

[Andrew Atiya @atiya_ag]

@signulll More fundamental: what is everyone so afraid of? We are so used to thinking in negative-sum framing we cannot imagine a world where more is possible. It's not right. That's not how the arc of history bends.

[Andrew Atiya @atiya_ag]

@heyderekj Astro is so good.

[Andrew Atiya @atiya_ag]

@oliverbrocato Automating something that never needed to happen is the new definition of insanity.

[Andrew Atiya @atiya_ag]

@jessegenet @AnthropicAI I worry it's Dario's playbook. Seems like he was driving that behavior at OpenAI. To knowingly (or even unknowingly) frighten so many people suggests a profound psychological injury. I wonder if he could stop if he wanted to.

[Andrew Atiya @atiya_ag]

@hypersoren Why are those the only choices? I don't understand why the big labs are trying value destructive recreations of other verticals, instead of going hard after creating new developer and channel partner ecosystems. When did entering verticals a company doesn't understand ever work?

[Andrew Atiya @atiya_ag]

Something that is driving how I work and what I build: avoid workslop collapse. Let agents write too much on your behalf and they start to choke on their own effluence. What started as a good idea becomes a meaningless word salad of nonsense. The human needs to stay in the loop.

[Andrew Atiya @atiya_ag]

@DavidOndrej1 Is it really too thirsty to use with a codex/claude sub?

[Andrew Atiya @atiya_ag]

@wjessup Great point. I think judgment calls of this sort are a lot of the reason why pure vibe coding goes awry. The models can write “good” code and do other verifiable tasks. But determining the scope of a change is a much more subtle art. Requires foresight

[Andrew Atiya @atiya_ag]

Hard lesson: your codebase becomes part of every prompt. In my experience, confusing or vague field and function names materially degraded agent performance in coding and in agent harnesses.

Life got better once I took back ownership of naming.

[Andrew Atiya @atiya_ag]

Rebuilding our app to be thread native.

Where we're going, we won't need wikis.

[Andrew Atiya @atiya_ag]

This is a lot of stress for a make-believe scare campaign that got out of hand. This whole thing has zany comedy vibes. I get it now. This is just like "Some Like it Hot" but for destroying hope in the future and sorting society into haves and have nots.

[Andrew Atiya @atiya_ag]

"Routine tasks like coding." Is this a typo? How did they play themselves this hard?

[Andrew Atiya @atiya_ag]

@saradietschy Yeah this is a challenge, especially if you use an ORM that isn't well represented in training data. I wonder if it doesn't expose a more fundamental limitation in LLM capabilities. Something about seeing a file in the repo that it's not supposed to edit just blows its mind.