This post discusses a possible way to detect feature absorption: find SAE latents that (1) have a similar causal effect, but (2) don't activate on the same token. We'll discuss the theory of how this method should work, and we'll also briefly go over how it doesn't work in practice.
Feature absorption was introduced in A is for Absorption.
Their example: Assume the model has two features, 𝒜="this token is 'short'" and ℬ="this word starts with s". In the prompt "Sam is short", 𝒜 activates on "short", and ℬ activates on "Sam" and "short". Hence the L0 across those two features is 3. We train an SAE to recover these features.
But the SAE is... (read 1594 more words →)
Seconded
As Tyler Cowen points out, we could easily already have a vaccine ready at trivial additional cost.
It looks like that post was written by Alex Tabarrok.
Last week I read the literature and concluded:
Based on these two studies, it looks almost certain that hydroxychloroquine is at least as safe as a placebo for reducing symptoms [of COVID-19], and the drug probably reduces the incidence of symptoms by a little more than 10%.
The 10% is a relative reduction, not absolute. I don't know how Scott Alexander defines "clinically significant". Some authors thought that "significant" meant a 50% or 90% relative reduction in cases, although I personally think that a 10% reduction matters. But I have no medical experience and no medical training. If you read Stat News, you know more about medicine than I do.
I also conclude:
... (read more)It seems that
Anecdote, part 2: Around the beginning of April, Andrew Cuomo was added to the presidential (or maybe Democratic nominee?) market, and was trading at roughly 10% (or 5%, or 15%—I really don't remember). Since I had already had large No positions for everyone else, I got about $60 in negative risk by buying a lot of No shares on Cuomo.
I started self-hosting in early July, and it's been fairly straightforward. As a newcomer to blogging, I followed these instructions to install WordPress onto a HostGator domain. It cost $95 for three years of hosting, as well as an $18 domain name fee each year after the first year.
I have no way to compare with other hosts, but HostGator seems OK. The HostGator settings are not the easiest to navigate, but I only had to use them once (to update PHP, whatever that is). Be aware that your contact information—name, email, phone, address—is available online unless you pay $15 a year. I'm OK with that information being public, but you might not... (read more)
Anecdote: In late September I put $20 into PredictIt and followed Zvi's advice, focusing on the markets for 2020 Democratic nominee and 2020 presidential winner. I was able to increase my money to $86.08, which after the withdrawal fee is $81.78. Because of the $850 per contract limit, a larger initial investment would not have gained any more money. This took a few hours to set up, so it's not very profitable. But it was fun and informative.