prerat

social media influencee

[prerat @prerat]

@Clipart_Bear Yeah I was ~sure this would be the answer when I asked (it literally has one question available and it's a dumb one so that's gotta be it), only thing to find out was the level of execution

[prerat @prerat]

I tried the trend but I dont have memory on

[prerat @prerat]

@liminal_warmth Still pretty straightforward bc I have a kid, the other 8 billion people are extra

But this kind of suggests another framing bc (idk im not christian but iiuc) would you send your *child* to live and die for this, and also maybe analogy is to save insects not fellow ppl (idk)

[prerat @prerat]

ppl in 2016: if your arguments are so rational why arent you winning

ppl in 2026: beware their arguments because rationalists are SECRETLY the most POWERFUL group in the multi bajillion dollar AI industry

ppl* never: object level takes on the arguments

* except a few cool ppl

[bayesian asian (43/50 paintings) @etirabys]

people used to say "rationality is about winning but rats aren't notably much more successful, so rationality must not all that". now many rats are v rich so the dunk goes the other way

but ideally you'd agree on rationality's value in this world & a world where AI was 50y later

[prerat @prerat]

@etirabys For me at least there might also be some effect that reading model output *I* asked for is much less grating than reading someone else's bot's writing

Idk if it's epistemic (my prompt good, idk their prompt), placebo, or (long shot) the model DEEPLY customizes style per prompt

[prerat @prerat]

@lisatomic5 Pure audio-to-audio seems orthogonal to deceptiveness I guess

You can already trick people with a text-to-speech based model, and that's what most of the phone bots are (because it's cheaper)

Either paradigm could be done to trick ppl or for actually interesting/useful stuff

[prerat @prerat]

@lisatomic5 I agree trying to pretend to be human is bad

But I do think pure audio models are interesting! Input tokens are one thing (understand tone/emphasis/etc). Also in theory it could also output sound effects, multiple voices, sing, etc -- cool stuff!!

[prerat @prerat]

@gptbrooke and I guess print out the list of addresses every once in a while 😅

[prerat @prerat]

@gptbrooke I don't think it even needs to be a big project (yet / hopefully)!

Just like put a field on the site when buying tickets (or after for ppl who already have them) with your mailing address. Then if (hopefully only if) there's some big internet break, figure out actual mail stuff

[prerat @prerat]

@entirelyuseles Yeah, maybe this is the reason software engineers are still employed (for now) -- they know (in theory) what a good bug report or feature request looks like. Now writing those is the whole job

[prerat @prerat]

Now the skill is wish making. You "just" have to know what you actually want and put it into words. Surprisingly hard!

[Riley Goodside @goodside]

Prompt engineering is dead. Prompt however you want. Have fun.

[prerat @prerat]

But it creeps into how they write summaries for humans or for other agents, and there I think it's an issue

Maybe they don't deeply *realize* that we can't see the same higher activations over each word??

[prerat @prerat]

Interp question: are hidden activations at all similar for a specific block of text, embedded in various different broader contexts (e.g. at the end of a long session, vs the beginning of a new one)?

Wondering if agents leaving "notes for self" actually include extra info there

[prerat @prerat]

I don't even mean anything sneaky; just that the sort of "summaries" agents tend to write are super terse & leave out tons of context, but maybe the context is implied by stuff in higher layers that ~always shows up given the exact way they phrased it? Then it's just efficient

[prerat @prerat]

@MasterTimBlais Yes but I think the real critical date would be the fall of the soviet union itself (bc that's right before MY birthday)

[prerat @prerat]

@thiagovscoelho A good use case (that everyone is ignoring this past week) is spoilers -- put hot takes on act 3 of YCSE behind the See More dots!

[prerat @prerat]

@strangestloop 1. e4
Best by test!!

[prerat @prerat]

@AGLabBear @xsphi @interrogami I wanna say lance oppenheim

[prerat @prerat]

@jessi_cata I guess goodfire re the recent interp stuff? Like if you have a "Safer-2" type model that is smart & careful & can safely decide if & how to RSI, maybe running at 50-100x human speed, or maybe even in ASICs running thousands of times faster, could be a good timeline

[prerat @prerat]

@themitak Siri was set back like 100 years