William Eliot

Net value of further analysis equals expected benefit minus expected cost

I would say it’s an infohazard to know your own mind

Let’s go bunker buster on the loss landscape

Here’s your permission to have a holy shit moment because agentic AI is here and working

I’ve stopped saying ‘we live in a perfect simulation’ because it is informationally void, telling us virtually nothing about our reality. Anthropic reasoning makes it likely, and I still believe it, but the claim is about as meaningful as saying ‘the universe exists’ — duh

Proper that bitter comes first in bittersweet

it’s ironic til it’s not

I fine-tuned a model on 100% ethical instruction and reasoning datasets, creating (to my knowledge) the world's first reasoning LLM from compliant data. But it underperforms GPT-3.5, so there's a difficult lesson here. williameliot.com/posts/front…

Fable has a major word salad problem

ml research notes

The perceiver's intelligence shrinks the uncanny valley: shallower, narrower, a higher bar.