prerat

32K posts

prerat

@prerat

social media influencee

Katılım Ağustos 2020

1K Takip Edilen12.3K Takipçiler

prerat@prerat·4h

@maiab scene is exactly 5 minutes long btw if you want to start the movie un-sadly trivia: the original cut of the movie started that way & only revealed backstory in flashbacks. but it made the dad too unsympathetic so they put the backstory at the start. but i dont think kids care

English

295

Maia Bittner@maiab·13h

oh my god the start of Finding Nemo is absolutely gutting

staysaasy@staysaasy

Since having kids I am completely incapable of handling it when kids are in distress. I didn't *like* it or anything before having kids, but I could tolerate it in reasonable doses. Since my first was born I can't watch movies where people hurt children, can't handle pictures of starving kids, can't read stories about it, can even feel myself getting somewhat upset about someone *else's* kid being truly distraught at the Dr's office. It's like a neurological switch flipped. I can barely even handle the scenes where the parents are worrying about their daughter in Moana.

English

8.1K

prerat@prerat·1d

@lazy_grad @ApriiSR before you see the tool repsonse, should be 50% right? you are top level non subagent claude by observation. so you know you're at the top of the tree, and it's a coin flip whether any more will be created

English

Lazy Graduate Student@lazy_grad·1d

@prerat @ApriiSR Hmm. I guess this thought experiment confuses me: you're Claude, and you have a tool that spawns 10000 subagents (also Claudes). The system prompt says the tool may be disabled (50% chance). You call it. You observe you're not a subagent. What's the prob that tool is disabled?

English

prerat@prerat·1d

bostrom: "if all worlds under consideration have boltzmann brains, then bayesian updating is impossible bc all observations have probability 1 in all possible worlds" ok but one man's modus ponens is another's modus tollens therefore you have to do Updateless Decision Theory.

English

1.2K

prerat@prerat·1d

@ApriiSR @lazy_grad say claude's utility function mainly cares about humans say there is an idempotent button that gives +1 to a human if right abt heads/tails, but -100 to the human if wrong there shouldn't be any distribution of claudes i can create after a fair coin that will press the button

English

Aprii 🩷💎🔎💜@ApriiSR·1d

@prerat @lazy_grad my idempotent act that's dangerous in million claude world impacts a hundred thousand times as many claudes in tails world,

English

prerat@prerat·1d

@lazy_grad @ApriiSR also i think it's prob fine to be able to estimate the expected total # of claudes -- that's an objective number regardless of anthropics my objection is about backwards bayesian reasoning to claim that you are almost certainly in a small world, for idempotent acts especially

English

prerat@prerat·1d

@lazy_grad @ApriiSR im not sure but i think plausibly something different happens if you are an external observer who is sampling (in which case sampling is sort of forced), vs if you are just one observer trying to decide whether to accept "self-sampling" without anyone actually sampling anything

English

prerat@prerat·1d

@lazy_grad @ApriiSR i want to factor out world-uncertainty separately from the impact of having a lot of copies in the tails world, anything a claude does is amplified 100000x compared to the tails world but this should be separate from epistemics! e.g. some acts are idempotent & others additive

English

Lazy Graduate Student@lazy_grad·1d

@prerat @ApriiSR I disagree here. In your version, world Heads Claudes all say 50%, while in world Tails you have 10 Claudes that say 50% and 999990 Claudes that are certain in mine, world H claudes are all >99.9% credence on correct value, but the 10 world T Claudes are wrong instead of 50%

English

prerat@prerat·1d

@ApriiSR @lazy_grad yes i think falling back to UDT optimal policy selection is the safe move, and we need to evaluate any epistemic claims against that gold standard. (but also i think there might be some interesting epistemic assignments that are worth defining if we can figure them out)

English

Aprii 🩷💎🔎💜@ApriiSR·1d

@prerat @lazy_grad i'm not sure i've convinced myself that assigning some p(heads) is a useful way to think about this scenario. i'd just like, step back and think about what claude-5 behaviors have consequences i like across possible scenarios

English

prerat@prerat·1d

@lazy_grad @ApriiSR this is a real problem for claudes. say you ask claude #5 to bet on the state of the world. for some reason, it's ~certain it is in the heads world. so for the 50% tails world, that claude #5 will throw away all kinds of utility on bets assuming heads. u can't delegate to it.

English

prerat@prerat·1d

@lazy_grad @ApriiSR say you flip a coin. if heads, spin up 10 claudes. if tails, spin up 1000000 claudes. you tell each claude its sequence number. say you are claude #5. what is your P(heads)? i think (contra doomsday) it should be 50%. there is a claude #5 in both worlds, so there's no update.

English

103

prerat@prerat·1d

@owlthatissuperb also the previous tweet

English

358

Max (Vibecamp June 18-21)@owlthatissuperb·1d

This tweet has very few likes relative to its cultural importance

brooke - vibecamp 6/18-21@gptbrooke

Also had VibeCamp suggested by @grin_io

English

1.1K

prerat@prerat·1d

@ApriiSR i think that step is pretty justified for BBs that are about to fluctuate away in the very next moment so it doesn't matter at all what they choose to "do" best idea for a weird side effect is that it says you should internally wirehead but idk x.com/i/status/20510…

prerat@prerat

cute anthropic bodhisattva idea: u must jhana rn, bc 99.99999999...% of copies of your current mental state are boltzmann brains, so producing joy for them dominates any other consideration (i dont think this actually works but it is a fun idea)

English

Aprii 🩷💎🔎💜@ApriiSR·1d

@prerat yeah this is basically my line of thinking but i'm a little uneasy about the BB exclusion step and i'm also not sure it doesn't produce anything weird

English

prerat@prerat·1d

English

666

prerat@prerat·1d

@ApriiSR roughly my current idea is to give up the idea of being instantiated in exactly one place+time, and so indexical uncertainty now gives an *unnormalized* probability distribution then you can ask for conditionals relevant to things that actually matter (which excludes the BBs)

English

prerat@prerat·1d

@ApriiSR yeah this is my question right now UDT gives correct decision theory answers if you supply some utility function to maximize, but it's philosophically unsatisfying because it refuses to answer epistemic questions (never bayesian updates)

English

prerat@prerat·1d

@jessi_cata suspiciously often i have some q like this and find an old jessica taylor blog post explaining

English

prerat@prerat·1d

@jessi_cata i was thinking about this a few days ago, how CDT sort of ignores all of your copies but then SIA "presumptuously" over-counts them, maybe it cancels out so i typed "CDT+SIA" into google and saw your post and was like whoa

English

108

prerat@prerat·1d

@ApriiSR if you want to do this, you need to accept some anthropic assumption that has weird consequences e.g. SIA says big worlds are fundamentally more likely than little worlds SSA implies doomsday SSSA w minimal ref class has the problem from OP (& other ref classes are arbitrary)

English

214

Aprii 🩷💎🔎💜@ApriiSR·1d

??? different big world theories can assign different measures to certain observations over others. maybe in some situation big world theory A thinks you're more likely to one observation 1 and theory B thinks you're more likely to make observation 2

English

243

prerat@prerat·1d

@cremieuxrecueil when prompted to consider the possibility of pushing the wrong button by mistake, blue wins in a landslide: x.com/i/status/16910…

prerat@prerat

5% stochastic error version: Everyone chooses blue pill or red pill. - if > 50% of ppl choose blue pill, everyone lives - if not, red pills live and blue pills die BUT there's a 5% chance your pick will be flipped randomly by accident. i.e if u pick red, 5% chance you get blue

English

952

Crémieux@cremieuxrecueil·1d

Accurate cartoon. In my polling, almost 6% of people said they picked the wrong button when they were prompted about it. Assume the global error rate is higher and kids' responses are random, and easily 1-in-6 to 1-in-5 hit blue by mistake. So, if red wins, society is ruined.

Jason Boone@shadowe_wolfe

@TheCartoonLoon That's not even close to accurate.

English

299

1.3K

213.9K

prerat@prerat·1d

@meadandjuniper i think 10k is still fine, it's basically still just the same few people interacting on most posts. most of my posts still get similar engagement to when i was at 1k i bet there's just a correlation between tons of follows and posting algo bait, and thats why it sucks

English

414

j@meadandjuniper·1d

I would kill myself if I ever got 10k+ followers I love small account vibes so much, it’s so nice having convos with moots in replies where it’s almost like 1 on 1 but also yalls moots can pop in too sometimes and just say hello and stuff it’s so fun it’s like if all your neighbors were your friends and you can just sit on your porch and talk to each other and wave to your pals across the street

English

Keşfet

@maiab @lazy_grad @ApriiSR @owlthatissuperb @elonmusk @BarackObama @taylorswift13 @cristiano