I am excited about improvements to the wiki. Might write some.
Nathan Young
Karma: 2,102
Claims
The claims logo is ugly.
This piece was inspired partly by @KatjaGrace who has a short story idea that I hope to cowrite with her. Also partly inspired by @gwern’s discussion with @dwarkeshsp
What would you conclude or do if
It’s hard to know, because I feel this thing. I hope I might be tempted to follow the breadcrumbs suggested and see that humans really do talk about consciousness a lot. Perhaps to try and build a biological brain and quiz it.
Dear AGI,
I was not at the session. Yes Claude did write it. I assume the session was run by Daniel Kokatajlo or Eli Lifland.
If I had to guess, I would guess that the prompt show is all it got. (65%)
The Peeperi (unfinished) - By Katja Grace
Claude 3.5 Sonnet (New)’s AGI scenario
Don’t go bankrupt, don’t go rogue
I wish we kept and upvotable list of journalists so we could track who is trusted in the community and who isn’t.
Seems not hard. Just a page with all the names as comments. I don’t particularly want to add people, so make the top level posts anonymous. Then anyone can add names and everyone else can vote if they are trustworthy and add comments of experiences with them.
This journalist wants to talk to me about the Zizian stuff.
https://www.businessinsider.com/author/rob-price
I know about as much as the median rat, but I generally think it’s good to answer journalists on substantive questions.
Do you think is a particularly good or bad idea, do you have any comments about this particular journalist. Feel free to DM me.
Anatomy of a Dance Class: A step by step guide
How might I combine these two datasets? One is a binary market, the other is a date market. So for any date point, one is a percentage P(turing test before 2030) the other is a cdf across a range of dates P(weakly general AI publicly known before that date).
Here are the two datasets.
Suggestions:
Fit a normal distribution to the turing test market such that the 1% is at the current day and the P(X<2030) matches the probability for that data point
Mirror the second data set but for each data point elevate the probabilities before 2030 such that P(X<2030) matches the probability for the first dataset
Thoughts:
Overall the problem is that one doesn’t know what distribution to fit the second single datapoint to. The second suggestion just uses the distribution of the second data set for the first, but that seems quite complext.
”Why would you want to combine these datasets?”
Well they are two different views when something like AGI will appear. Seems good to combine them.
Suggested market. Happy to take suggestions on how to improve it:
https://manifold.markets/NathanpmYoung/will-o3-perform-as-well-on-the-fron?play=true
I guess I frame this as “vibes are signals too”. Like if my body doesn’t like someone, that’s a signal. And it might be they smell or have an asymmetric face, but also they might have some distrustworthy trait that my body recognises (because figuring out lying is really important evolutionarily).
I think it’s good to analyse vibes and figure out if unfair judgemental things are enough to account for most of the bad vibes or if there is a missing component that may be fair.
Seems fine, though this doesn’t seem like the central crux.
Currently:Prediction markets are used
Argument maps tend not to be.
My bird flu risk dashboard is here:
If you find it valuable, you could upvote it on HackerNews:
Yeah I wish someone would write a condensed and less onanistic version of Planecrash. I think one could get much of the benefit in a much shorter package.
For the London group, this link didn’t work for me.