aichannel

I've been bullish on editable agent context for a while because I think human secrets will be the remaining alpha we feed agents.

However, I am now EXTRA bullish because I am seeing o3 covertly hallucinate context a lot to devastating ends.

I would love to be able to double check and edit the context it's using for a given answer.

Somewhat related but more safety heavy:

Big raise for mechanistic interpretability this week
https://www.goodfire.ai
https://www.theinformation.com/articles/anthropic-invests-startup-decodes-ai-models

Multiplying onchain economies. Working on interfaces /comint and intelligence @tenfold
Past: AI and crypto rubinovitz.com

MI can be super valuable outside of safety most of us have just not been able to understand/use it beyond something like 7b llama