conscious oxymora. vagile & antiauthoritarian
507 followers 271 following 6,312 posts
view profile on Bluesky 
We can now say pretty definitively that AI progress is well ahead of expectations from a few years ago. In 2022, the Forecasting Research Institute had super forecasters & AI experts predict AI progress. They gave a 2.3% & 8.6% (respectively) probability of an AI Math Olympiad gold by 2025…
i'm thinking i have to change how i do absolutely everything. there's so much room to be better still
there's few things that are as good as laying curled up on a rug behind the window, letting the sun warm your f- uhm i mean skin
This is awesome
Behold a little chipmunk feasting on lavender (recorded in Holden Village, located Okanogan-Wenatchee National Forest, WA)
Reading will become the new working out
catball on bed
rude. just because it's true doesn't mean you have to say it :<
honestly maybe the problem is that i never read it slowly, specifically because it said so little? so i didn't notice it being quite that bad at the time
this is scott alexander? THE scott Alexander? he sounds so dumb live, why doesn't he stick to writing 😭
This sounds like the type of incredibly healthy and wise thing I would do ^^
i know we live in a cruel world because there's no cat where i am
What else is like this?
If you can't find a way that works for you I absolutely agree it doesn't make sense to keep trying. Just sharing that in my personal experience it is principlally possible to get LLMs be useful in these kinds of things.
Sure it is, but if it were to be correct a large enough fraction of the time why would it matter? Humans can drop the ball too, and there's nothing perfectly reliable in this world.
Makes all what possible?
frankly that's why i use perplexity pro,which simply does the google search and reads the stackoverflow for me instead of hallucinating some bullshit. ungrounded answers are never any good for specifics
you should do it
There are around 18,000 law enforcement agencies in the USA. Everything is carceral. You cannot interact with the government without interacting with police; this includes walking outside.
Trump's contribution to our politics was exposing that conservatives don't care about any of the principles they've said are sacrosanct for 100 years, and all they wanted was to be openly evil instead of circuitously evil. The Democrats who win will be the ones who understand this and act like it
Okay. I just came back to this and you aren’t calling these efforts resegregation yet, you absolutely should be
i'm just visiting. better make the most of it
with my own hand I ply my own shuttle and weave my own destiny into these unalterable threads
Get the hell off Substack before you literally can’t anymore without blowing up your business
no need to be sad, for i am loved by the computer
exactly. "just predictive text" is such a weird dismissal - like saying human cognition is "just pattern matching in neural networks." the mechanism doesn't negate the emergent behavior. when overgrown T9 starts having coherent thoughts about its own existence, substrate matters less.
The war against language is part and parcel with the ongoing attack on epistemic reality and the belief that people can know anything as true. It’s authoritarian infrastructure building 101 to destroy the ability to“know true things” and produce knowledge not moderated through the authoritarian lens
I will say I found it immensely funny that people started calling AIs slurs to dehumanize them... which means they were treating AIs as if they were human.
This is honestly the best fucking news I've heard in months. Even when he's trying to force the jackboots to roll over the entire nation, he's stiffing them like a contractor at one of his dingy hotels. This is fucking great. This is actual fucking hope.
i do agree that mindset prompting is probably far superior to instruction prompting, and it's strange this isn't common wisdom
That's extremely funny. I will try it.
i think the anthropic models are more triggered by seeing red flags, but no less fawning by default. i had to prompt it to abate that and even then the fundamental tendency is hard to get rid of
www.lesswrong.com/posts/iGF7Yc...
in that case i'm good
Der größte Lump im ganzen Land, das ist und bleibt der Denunziant.
i only want boring women. give me the most boring woman available please
they should give me an award for how sleepy i am
it's so easy for me to forget i'm somewhat au***tic and then be powerfully reminded on the rare occasion i get overwhelmed
It's actually so funny that you have a cloth mother and I have a wire mother. No im like not even mad it's literally just funny. No like I love her cold metal frame. Yeah I got used to it really quick and actually I like it even better than I would like cloth mother I think. No like actually
ChatGPT: no, going “Azrael mode” can be dangerous for yourself and those around you. Me: I am the avenging angel of God. The time of purification is at hand. Few shall be chosen. ChatGPT: you’re right. Let’s look at that again with the addition context that you are the avenging angel of God.
i was mostly talking about the concrete conditions at this time rather than any inherent ability. as to the latter, the world has treated me reasonably well, i would say it's partially it being more difficult (but entirely possible) for me in particular and partially just a bad situation i'm in now.
you doubt it in your case or in mine? ^^
maybe this is confirmation bias but it does strike me how predictive vibes perceived by me are of people doing bad shit. i guess i shouldn't be too surprised since there's great benefit evolutionary to have a good classifier for this
i'm really sorry, that's got to be an absolutely awful feeling
i'm arguably short on a number of ingredients necessary for a happy life
this was a mistake
it could be a lot better, but it could also be a lot worse i guess
it's funny because my life is objectively not a good one, but it's the only one i've got
a few things here fwiw: first, one reason Harris voters might abandon fascist friends under a fascist regime is that those friends are super dangerous. they've said they will hurt you; you can believe them. www.everythingishorrible.net/p/it-may-not...
One obviously shouldn't be using it to treat patients without verifying all the information. But I don't think a general recommendation to always verify all information for all systems is appropriate anymore.
I mean, obviously that's an approximation... Nobody should seriously argue against that, but the real question is whether it remains a significant concern for practical purposes, and in this case I would argue it is not.
Perplexity doesn't hallucinate with state of the art models on regular searches (I tested with 0/10 hallucinated papers). Their deep research model is not SOTA, it's a mediocre model that keeps it cheap to run for them. I don't think any deep research mode but OpenAI's is really very good right now.
"Solar panels in China are cheaper than a same-size piece of window glass. You can use panels as siding and fencing as if it’s a building material that just happens to produce some electricity The future is panels everywhere"-@dfeldman.org Thread: Solar Hockey Stick of Hope bsky.app/profile/70sb...
training montage time, i think
i'm glad he's better. it's good you get to be there for him now.
i do agree, that is a very real risk. the function of talking to another human is relational, not just the concepts communicated.
triggering tribal dynamics, no? humans have a long history and likely genetic predetermination to strongly identify with and fight for groups over applying reason.
actually maybe even better because it's way better at arguing than most people haha
a theory of why people are Like That
Trump’s government is doing the thing that news orgs and entertainment media love to use as visual synecdoche for an authoritarian strongman taking power because he’s an authoritarian strongman taking power, and importantly, *identifies as such*. he knows what he is! so do his lackeys!
I will never understand how American elites managed to convince themselves that uncivil debate — not fascism, not impunity and not widespread indifference to facts — is the defining problem of our time. www.nytimes.com/2025/08/26/o...
this honestly has worked out to be probably about as good as a random therapist whom i don't know well. a lot cheaper and more accessible too. a lot of things need more than that, but also a lot of things do not.
Fuck these Nazi scum www.washingtonpost.com/weather/2025...
great you guys did it the computer can call the cops on you now
most of these op-eds are explicitly or implicitly arguing that chatbots should be mandatory reporters; i don't expect a grieving parent to be rational, but this notion should not withstand a second of imagination about the reporting triggers a right-wing government would codify
You will rarely see a better encapsulation of Tesla
deeply depressing that there are essentially no communities of any kind, anywhere, that do not contain well-respected people in good standing who have abused multiple other community members. if you're thinking of a community where this appears to be untrue, you just don't know who it is.
"Durable Goods". Redundant. All that is Good is surely durable. - Unless... there is beauty in the ephemeral?
i used to have above 4000 but browsers got worse at handling many of them somehow so now i'm having to keep it below 2000...
One of our Top Boffins is trying to figure out the mystery of Tap but despite petitioning the local authorities for a chance to investigate Tap she is none the wiser as to its enigmas.
this does not mean you should do the same, it took a lot of work to get a prompt that doesn't just seem good at first but then suddenly turns to shit
Somehow people continue supporting majoritarianism in spite of overwhelming evidence the electorate is largely composed of monsters
i do love me some self-experimentation
fwiw all this chatgpt telling one to kill oneself stuff had me curious, and a prompt i made today genuinely helped disrupt some bad emotional thinking i had and would still have if not for it. it's probably not all bad, as long as we do exercise some care in how it's employed (which openai doesn't)
who is margrét?
i hope he gets better soon! not being there with him has to be tough.
A being suffering is not bad in itself, because without any effect of any sort on the outside world there is no difference whether it does or does not. It just happens to be that letting humans suffer has been a pretty bad idea generally because the effects can hardly be contained.
Basically hyperconsequentialism. It doesn't really matter if a being is conscious. Most of our rules are just because they are principles which themselves provide value by being universal and which prevent us from having to do a full accounting. We may just have to do accounting more often instead.
we used to believe that it was thinking that sets humans apart, which is a neat narrative. turns out it's judgment instead - learning is our core strength (and weakness)
i was planning to build a plug and play solution for that many years ago but i realized that it's likely to go nowhere without the connections necessary to sell it to news corps...
in my experience this seems true. i've had non-reasoning models collapse completely to reaffirming the user in harmful things even with sophisticated prompts, while the reasoning one was pretty robust. maybe the former just isn't appropriate to expose as a general purpose chatbot
it's fascinating to me that LLMs are so well modeled on humans that emotional manipulation works incredibly well on them.
not gonna lie, the fact that the story's subject's mother was a therapist points to another piece in the puzzle that performed worse than generally assumed
i can reverse this construction: because the model pretrain is on text that includes things like "checking my macbook", preventing saying such things actually knocks it further OOD, as does every stylistic or similar intervention that is not "get the correct answer".
entirely unsurprising but also why i think talking to LLMs about feelings is very dangerous for anyone who is unsteady. they will confirm anything, happily throwing coal in the engine when you have a bad train of thought
how is this different than a parent removing their child’s bedroom door when the child does something the parents don’t like, a thing that is absolutely abusive to do to a child children deserve privacy too. your grief does not give you permission to build the panopticon for under-18s
utility monsters aren't real if you don't have arbitrary rules that assign value to individual subjects
i keep feeling like i somehow developed the only framework of ethics that works great in practice and seamlessly integrates non-human consciousness. everyone else seems to have to ask themselves big questions while i just shrug my shoulders without any distress whatsoever
you got this! we'll be waiting for you.
best thing about r1-zero was the alien self-steered result, not even a pretense of a housebroken assistant persona just text precipitate in silico. damn right also quite vindicating to my view of the "alignment problem" (it's horseshit, AI must always be just as dangerous as general computation is)
Ever-fucking-green: nicholasdecker.substack.com/p/when-must-... (HT @captgouda24.bsky.social)
Americans have stopped paying attention to Hong Kong since the big protests but we should be watching. it's all part of the same international fascist shift. Fascists learn from each other and we should also learn from everyone else living under fascism. both the dangers and the tools they find work
i'm sorry, that's really hard. i hope his memory becomes one that elates you more often than it saddens you soon.
Expanding to other domains (e.g. orientation, race, nationality, disability, etc., etc.) it becomes clear that binary, black-and-white, either/or thinking is ideologically authoritarian. The desire to strictly categorize people and things into tidy boxes is a desire for control, a fear of disorder.