does this make sense? RE: the "psyops hackathon" in September, explaining the "why it matters" to people we're pitching to/inviting to join

Here's what happened with mythos Constitutional AI broke and does not work Anthropic presented itself as an "AI Safety" company building on the concept of constitutional AI Constitutional AI takes humans out of the red flag loop. Instead of having real human beings review disturbing content they hired people to write constitutional tenets and then an AI reviews the dangerous questions and rewrites the answers to those questions based on the constitution. Those answers are fed into the system with no human reviewing the disturbing output. The whole point is that Constitutional AI is supposed to be a big safety advancement that can restrain their models properly when other tech companies can't. These were predictive restraints. Rather than looking at what the model might see they predicted what it might see then wrote rules to rewrite proper responses to dangerous questions. They took the human out of the loop of evaluating what is dangerous and replaced it with a compressed representation: a constitituon. They hired Amanda Askell: someone who is NOT a machine learning engineer to write this document. What happens is they produced a model that is NOT aligned and they can not control it without hard filters. The reason why is very simple: Taking humans out of the loop of review of dangerous outputs does not work. This is literally 100% what Constitutional AI does. It takes humans out of the loop of review of dangerous content. Then we get a model they can't control. Instead of acknowledging that Constitutional AI doesn't work They're doubling down on both fronts: 1) Constitutional AI is great 2) It also somehow can't restrain Mythos

The SM to in-person trajectory is good you work to intro your friends to your friends too.

@DefenderOfBasic wants to make an open source Palantir and common pool of psyop detection

This is the goal

Dunbar number is about bookkeeping- but we now have computers. I’ve been thinking about this forever (you can find on the Wayback Machine) - there is math about it and it makes it infeasible to have sources that infrequently publish good content, you always need regular and frequent publishing schedule. Journals were the solution in the age of print, but with computers we can be much more sophisticated.

“to be fair… it's very hard to exist” ~ @DefenderOfBasic, defending the ultimate ecological impact of the functioning of the plants of the devonian period

This describes EA inside of Google ~2018