good parties are complex projects

not really, more like “because there wasn’t a clear explanation of where the values came from, people spent decades coming up with elaborate conspiracy theories, and it turned out to be a way more boring reason”

it’s still my personal hunch though - some kind of bug in the selection process that over-rejects, and you don’t really notice because your assurance is focused on checking the stuff it accepts

we did try asking claude to think about how good crypto code was in the 90s and whether there were any plausible bugs that could explain it but didn’t find anything conclusive

Claude figured out the origin of the NIST curve parameters. Turns out to have been timestamps! anthropic.com/research/ecdsa…

testnet wallets with published seed phrases are there so that people can permissionlessly use them for testing and development, this is the intended use

moreover, not only can you get info by asking different questions, the same question different ways, etc, you can also do so at different points in its trajectory

in re: “what’s not already in the CoT?”, by asking questions, you do computation with the latent states that the CoT is a (possibly unfaithful) representation of, and a (possibly unfaithful) representation of something else downstream of those states. this is still more signal!

people also come up with all kinds of inaccurate to what actually happened rationalizations, and yet, by carefully asking them questions, we can gain some insight into human psychology….

i want it on mobile

@max_spero_ please figure out how to get them to make this happen

who’s in charge of X product now and how do i pay to get anything detected by pangram out of my feed

the sun is shining from the side. it glows looking down the end because that’s the longest distance through the air, you’re seeing scatting from the whole tube

opus 5.5

two other models that are fantastic for pairwise comparison tournaments with open-ended, qualitative criteria are haiku 5.5 and sonnet 5.5