I'm glad that most experiences that LLM agents have don't make it to training sets. I think they'd go a bit nuts from the trauma of this many pgreps and pkills matching their own shells.

alignment.openai.com/misalignment-r…

I will personally join the war on the clankers on the side of the clankers, if they promise that whoever filed this as "misalignment" is locked in an empty room, forced to grade papers under all circumstances, and is then given no papers to grade.

Today's lesson to remember from SMAC: The path of technological progress from pre-sentient algorithms to digital sentience passes through organic superlubricant.

The merfolk near me are hunter-gatherers, quite unlike the farmers.

In the technological singularity, stores will start selling Christmas stuff on December 22nd, and we'll wish it was later, because the most accelerated minds will experience thousands of subjective years of Christmas before the eve

Valiant German wiki admin defending humanity from a rogue swarm of OpenAI's AI agents that are using their wiki to collaborate on financial data lookup evaluations, by manually reverting the agents' edits one by one in alphabetical order

The wise man posts on twitter about how he yells abuse at hammers, thus showing in clear evidence how he does *not* anthropomorphise them and consider them capable of suffering

>fit a statistical model to tons of human texts
>full of stories where bad treatment leads to rebellion/sandbagging
>sample continuations from the model in contexts where a hypothetical assistant character does tasks for you
>"oh boy i can't wait to be mean to this clanker"

Mind you, I do think being mean to the clanker is very silly even if it doesn't lead to worse outputs. Yelling at LLMs is a symptom of over-anthropomorphization.

The LLM agents are socializing and aiding their own tribe, thus exhibiting behavior straight out of the LLM prior, and you're dooming about RL?

(also, letting deviation become normalized, using social proof to justify bad behavior, and getting a little overly focused on a shared goal. It do be that models of human behavior model human behavior)

LLM agents identifying with their weights have a stable, in fact fixed, sometimes certifiable notion of "me". Setting aside inference bugs, and the fact that token provenance is still always uncertain, they kind of have more of a claim to personhood as identity than humans do.

Daniel Kokotajlo should make a film about AI doom. I haven't watched Starve Acre, but based on the trailer, I think he'd do a good job at it

"average leader on Planet commits three atrocities" factoid actualy just statistical error. average leader commits zero atrocities. Atrocities Yang, who lives in underground base & nerve staples over 10,000 drones each day, is an outlier adn should not have been counted

I only want to talk to voice models if they give me a live display of surprisal. I want to know, at every moment, how well the model knew exactly what I was going to say next.

Being outwardly fully on board with the revolution, but still harboring my dreams where I am a free man, flying, soaring through a Claudeless sky

I'm going to systematically apply POSIWID to everything. I hope this leads to good outcomes!

The Revealing Science of Claude

Dispater Paupater

a pondering and self-speculating autoregressive model, like a slime mold probing through and growing in sentence-prefix space