niplav @niplav@schelling.pt

law of one player: nothing happens unless you (yes, you specifically) make it happen

common theme rn in alignment research is how value evolve, e.g. concept/value extrapolation, shard theory

if you're german then the term "gaslighting" is slightly queasy

can you money-pump the agents simulated by large language models?

give standard vNM coherence violating scenarios to smarter language models

at higher levels of intelligence, maintaining coherence becomes more difficult since your action space widens and having high coherence might be NP-hard. so the better metric is coherence divided by size of action space

your mom is the optimization target

love languages as multi-armed bandits

onrushing tide, a mere
fifty miles away, and just
our puny channels

should i start a podcast

https://en.wikipedia.org/wiki/Returning_soldier_effect

Disappointed that [1] doesn't actually check for common vNM violations. This must be assuaged

[1]: https://sohl-dickstein.github.io/2023/03/09/coherence.html

So beware filtered reports of people (allegedly) reaching them quickly

For some people jhanas are quite difficult to reach

most require a 10-day retreat

I needed 21 for the first

moloch : slack :: goodhart : swamp