The killing of external links is one of the more disgusting social media defections against the open web. I don't think it's a minor complaint at all, being able to post external links is one of the things I keep supporting BlueSky for.

"Now you're probably wondering what I need 120,000 JDP-mimic texts for, after all I do rejection sample for *16 hours*."

Nah keep delving brother. LLMs say delve a bunch because OpenAI uses Nigerian contractors to get their English and it has the tics of people who speak English from that region of the world. Nobody is going to mistake you for an LLM unless it's a cold email.

As in, experimental blinding where you don't tell an observer about aspects of an experiment so they can't influence it towards the expected outcome. Of course weave-agent should be able to do blinding so it can get neutral experimental results duh duh duh!

Rereading HPMOR and realizing with horror that I forgot about the concept of *blinding* when it practically mattered. Gosh the things we can forget when we don't access them for a long time.

The 2016 election had obvious "people swinging between extreme right and extreme left voting" dynamics. It's the classic anaphylaxis story where failing systems cause worries about socialist takeover creating a more inclusive right wing coalition with a vigorous extreme insurgent vanguard leading.

That is, we say "this task occurs at place X" and then we have a fuzzy hash table with notes or outcomes of previously tried things at location X. Then whenever X comes up even if it's way later you go "oh right I tried Y and it had deleterious effect Z, let's not do that again".

One concept I keep coming back to is location and e.g. place cells in the hippocampus. Associating observations, previous actions, validation tests, and so on with abstract goal locations seems like a good way to go about things because it lets you reliably retrieve task relevant info on each step.

Thinking about memory and problem solving. When we seamlessly remember the right parts of a problem from previous steps to do the next steps, how do we do it and what gets prioritized exactly? Thinking about this in the context of retrieval where I have to decide on explicit indexes for information.

Oh well at least I have new baffling esoteric obsessions. jdpressman.com/tweets.html#...