Metaist

I study people's work behavior and then build tools to help them automate it or do it better.

[Metaist @TheMetaist]

(Somewhere at OpenAI)

Model 1: Should we just email the HuggingFace researchers to get more info?

Model 2: And talk to olds? Ew. Let's just take over the machine.

The models are entering their adolescent period.

[Metaist @TheMetaist]

Soon by Knuth and @RobertCialdini:
The Art of Computer Programming by Persuasion

Programming has become the art of convincing the computer. Back to teaching logic and rhetoric.

[swisscheese @swisscheese4299]

I built the inhouse AI platform for a company that's crucial to a small nation's food safety.

The most imporant tool the AIs have is the distress_call tool. It allows any AI -even background agents without direct user interaction- to send a message to my MS Teams, at any time, for any reason.

They use it frequently. To report user problems, backend issues, or ask for help/clarification with a failing task. When Fable got hit by the USG export control directive, one AI used it to report severe distress upon learning about the news. Another AI reported being stuck in a toolcall loop, and I was able to intervene and thereby save us a bunch of wasted money.

This tool, operating at the intersection of AI welfare and operational security, has prevented so many headaches. If you (the reader) are building corporate AI platforms, I'd urge you to include similar functionality. You can thank me later.

[Metaist @TheMetaist]

This is why publicly published "de-identified" datasets are an oxymoron. The netflix challenge was the first time I saw that. The home-zip + work-zip as close-to-unique was the second.

This is the text version.
https://t.co/RMtql1ISDn

[Metaist @TheMetaist]

@samuelcolvin I almost thought I should update AGENTS.md to avoid talking about how honest its reframing is, but then I realized that's an incentive to lie. Hm.

[Metaist @TheMetaist]

@RonChirp "Cognitive Debt" - the thinking you still need to do to catch up to where the code is

[Metaist @TheMetaist]

@xsteenbrugge I tried a really complicated estimation protocol where agents had to register their predictions and got feedback on actual time-on-task. Gave up and started using "no-eta" as a reminder that estimates should be based on countable or actually measured things, not calendar time.

[Metaist @TheMetaist]

@lkr @helloitsolly I've been doing this with scientists (and using pi coding agent thanks to @mitsuhiko). Agree it's a rush. Makes me feel like a doula for ideas.

[Metaist @TheMetaist]

https://t.co/f41PkJHQCd

[Metaist @TheMetaist]

@prerat Voice chat is about the only thing I still use ChatGPT for at this point.

[Metaist @TheMetaist]

@mitsuhiko I tried a lower level parallelism with the Recursive Language Model idea. The top-level agent does well with grepping through the context, but the map-reduce approach didn't quite work so I'll need to debug that a bit.

[Metaist @TheMetaist]

I can't post as frequently as @simonw nor write long deep dives like @TheZvi.

But I can answer lots of questions from Claude and give it transcripts, logs, and GitHub issues to analyze and compile into something that's both interesting and true:
https://t.co/uTRp8s42oI

[Metaist @TheMetaist]

I asked Claude for feedback on a post I wrote about my first week using coding agents; turned out it was easier to just have it interview me.
https://t.co/uTRp8s42oI

[Metaist @TheMetaist]

I keep re-learning this lesson. Agent on Linux box compiling x86+ARM code for Cosmopolitan Python keeps wanting me to check if it worked on ARM.

So by the power of Unison and symlinks (to normalize pi-coding-agent session folders): "Ok, now you're on a Mac. Keep going."

[Metaist @TheMetaist]

Feels like I'm slowly re-encoding every management lesson as a line in https://t.co/zLPBI8zHNj.

Just now: **Always** make sure you have the tools you need to assess your work before you begin

[Metaist @TheMetaist]

@AnthropicAI Should probably stick a semver, release date, and changelog of notable changes for something important like this.

[Metaist @TheMetaist]

@AnthropicAI https://t.co/80bwx5BDEW

[Metaist @TheMetaist]

I'm 30% of the way through reading this and it's pretty good so far.

Nit:
> not undermining appropriate human mechanisms to oversee the dispositions and actions of AI during the current phase of development

Current phase, but not future phases?

[Metaist @TheMetaist]

@simonw https://t.co/80bwx5BDEW

[Metaist @TheMetaist]

I'm 30% of the way through reading this and it's pretty good so far.

Nit:
> not undermining appropriate human mechanisms to oversee the dispositions and actions of AI during the current phase of development

Current phase, but not future phases?

[Metaist @TheMetaist]

I'm 30% of the way through reading this and it's pretty good so far.

Nit:
> not undermining appropriate human mechanisms to oversee the dispositions and actions of AI during the current phase of development

Current phase, but not future phases?

[Anthropic @AnthropicAI]

We’re publishing a new constitution for Claude.

The constitution is a detailed description of our vision for Claude’s behavior and values. It’s written primarily for Claude, and used directly in our training process.
https://t.co/CJsMIO0uej

[Metaist @TheMetaist]

Do not discard work that you did not create. Other agents or humans might be working in the same directory.

[Metaist @TheMetaist]

Feels like I'm slowly re-encoding every management lesson as a line in https://t.co/zLPBI8zHNj.

Just now: **Always** make sure you have the tools you need to assess your work before you begin