athom

Not everything has to be a concrete story; this post isn't trying to be another AI 2027. It's trying to broadly explain how power fundamentally works, so people can apply the relevant ideas to AI.

That said, the broader djbinder cinematic universe does also go into how economics and robotics will likely play out.

This post is largely arguing that your story here isn't the only path that leaves humans disempowered, and that we can end up with very bad outcomes even without RSI or AIs aligning/misaligning their successors at all.

OP is aware of the standard argument and demonstrating risks it doesn't account for.

You'd probably enjoy Master of the Senate then!

It's intimidating at first glance but quite a fun read

I'm curious what struck you as LLMish here (the first paragraphs don't read that way to me)

Hopefully I'm less obnoxious than Fable, even though you're more exposed to my verbal tics

I think it's been patched now, but I collected >150 examples beforehand: https://abigail-c-thomas.github.io/opus-5-glitch-text/

I'm not worried about this, its output isn't easily steerable afaict.

Also, having anything in the system instructions at all seems to fix the issue

I can only replicate this on the Claude web app, not Claude code or API.

It conditions on anything in memory there.

2 cents, not based on source code or these specific claims: It's a terminal app written in javascript, which is often slow/flickery because it redraws the entire screen whenever something changes. Whereas Codex fast-followed & very quickly was rewritten in Rust, and its ux is noticeably snappier.

Fwiw, human underperformance on vending-bench is an elicitation problem. The human was only given a set amount of time (iirc 5 hours), whereas the LLMs run until they lose coherence. The human maintained coherence throughout but types slower, and therefore sees less days of simulation.

The human is also the only case with n=1.

There might be a better chance of this becoming widely used if the design is pitched to IKEA (or similar), over being developed in a new, dedicated company

In BE 'quite' can also be used sarcastically, becoming a negative intensifier. If you say something's "quite nice" that could mean it really wasn't good at all. 

Less so once they've done enough RLHF

Update: the cards have all been claimed now (at least for this round)

Thanks for the post!

The Zach Weinersmith quote you mentioned goes further, in a direction that might also be relevant:

"...bravery is not cross contextual... Conversely, recklessness *is* cross contextual. Unsafe sex correlates to drug abuse."

I'm pretty reckless in some ways, but need to build courage in other areas.Fostering courage where I need it, turning my recklessness into courage where it's helping me, and minimising it where it isn't, all seem like worthwhile things to try.

Best guess: because they've found the Haskell expertise they were after. That tweet is nearly two years old now, and I'm pretty sure Ed Kmett joined more recently than that.

Is there any data on how people feel like rationality has changed their lives? Somewhat separate to what we're doing as a community / what some of the most successful rationalists are up to, it seems worth trying to work out how engaging more with rationality changes things for people. I'm pretty sure my engagement with rationality has helped me become stronger; if it turned out not to help most people that would definitely change how I presented things.