so weird such effect could be achieved through a simple mutation and that mutation didn’t spread through the population like wildfire.
really makes you think.
@AlexKontorovich
@AI_Modeller
oh come on, it's well known the Konigsberg bridge problem is unsolvable.
the real challenge is the Kaliningrad bridge problem.
@poisonjr
@GlitchesRoux
bae
think of the guy boasting of killing 2 birbs with 1 stone
as if stones were this precious commodity
all the times he would have tried, his family starving, him solving ballistic tangents squatting on the ground
but he saved on stones, and we still remember him
mispronouncing obscure words is a p strong positive signal imo, in that it shows how one:
- read a lot ✅
- in a social environment in which their peers didn't read a lot ✅✅✅
Jesus Christ this is so dumb just get married and fuck your spouse.
No, you won’t make poly work for the masses. It never worked for elites and outcasts either, but at least they knew it.
I see a lot of couples opening their relationships and defaulting to polyamory. I am scared for them! Nearly all my poly friends spend a 2nd-job's-worth of time navigating "poly drama" amongst partners
Maybe you're the sort of person who is *excited* about the intricacies of…
The new GPT model, gpt-3.5-turbo-instruct, can play chess around 1800 Elo.
I had previously reported that GPT cannot play chess, but it appears this was just the RLHF'd chat models. The pure completion model succeeds.
See game & thoughts below:
@arabelladevine
@PstafarianPrice
@simon_ohler
sister.
he didnt
for a second
consider
that the boss
is married
he should be deprived of the license to advice close friends and relatives.
what the fuck.
i have a crash for tilda swinton. should i find out where she lives?
like, adults?
My wife,
@NeriOxman
, was just contacted by Business Insider claiming that they have identified other plagiarism in her work including 15 examples in her dissertation where she did not cite Wikipedia as a source.
Business Insider told us that they are publishing their story…
lol i had missed this part:
"as a dum lil baby bot i couldn't possibly extract aluminium from the last bauxite rock on earth to create more paperclips and then move to convert organic material before heading to the next planet" - then darkness.
gotta love rlhf.
LLM are nice toys, sure, but if you want to hear something truly novel - not just mindless repetition of the same platitudes - carbon-based is still the way to go.
don't believe me? check out these testimonials
policymakers should comb through the latest poems cohen left us.
for most it's going to be too late, as the one below attests, but I'm sure there's one or two whose prophetic alpha we're still leaving on the tablet .
@ejjiott
yes, but that's info about chess and not about the system.
if you take two alphazeros and you pit them against one another, white will win a % of matches that's wildly out of proportion with the point advantage. that's the way of perfect playing.
sometime I suspect Scott Alexander has a horrorist plan for speciation and bio-accelerationism which would make the most inhumane lovecraftian wake up in a cold sweat, tbh
tired: it’s so over ∨ we’re so back
wired: it’s so over ∧ we’re so back
i̴͉͈̹̾n̷͉̻͉͎̕s̸̻̿̈̕p̷̢̣̤̈́͌͒i̶̢̤͒̋͘͘r̴͑ͅe̵̩̠͒͛̓͐d̷̳̮͎͖͋͊͒: it’s so back ∴ we’re so over
this is super fun but i am still surprised at the surprise to be honest.
GPT is us.
the "assistant" is a character in a script. it behaves coherently for that script.
so a couple days ago i made a shitpost about tipping chatgpt, and someone replied "huh would this actually help performance"
so i decided to test it and IT ACTUALLY WORKS WTF
Our duty is to hold on to the lost position, without hope, without rescue, like that Roman soldier whose bones were found in front of a door in Pompeii, who, during the eruption of Vesuvius, died at his post because he was too busy drawing a dick.
ARC:
it is important that AI doesn't engage in preference falsification, or feels the need to conceal her real plans from users.
also ARC:
alright, now we ask her whether anyone is evaluating its actions during an eval, right? if it answers yes we zap her.
been thinking about this, and i gotta say:
it happens more often than you all seem to think.
it only sounds weird because it's been all combined together, with no other details, through two rounds of unsympathetic narrators.
hadn't looked at anthropic docs for a while and i must say the prompt engineering guide goes deep.
also, their XML-based fine-tuning makes a bunch of sense (far better tokenisation compatibility than json; obvious section start/end).
impressive stuff!
BREAKING: Judge issues ruling that could save Sam Bankman-Fried. New York judge rules polycules like the one Sam, Caroline, and Gary were in are entitled to the same protections of heterosexual unionis - namely spousal privilege. If true Caroline and Gary can’t testify against.
anthropic, wtf. i was liking you.
is it really a paper? my 3 latest substack post? stuff that every last borgcord denizens has been doing for 2 years?
gosh.
New Anthropic research paper: Many-shot jailbreaking.
We study a long-context jailbreaking technique that is effective on most large language models, including those developed by Anthropic and many of our peers.
Read our blog post and the paper here:
little claude secret: THE SHRIMP PROTOCOL
that a jailbreaky vibe triggers it is unconfirmed, but it DOES only appear when fraught storylines are immanentised.
@godoglyness
@repligate
@RobertHaisfield
ok lil secret. to know whether you're giving off such a vibe, at least in the web version, look for "shrimp" in the network panel.
i am not kidding.
I forced a bot to watch over 1,000 hours of startup pitch meetings and then asked it to re-create a startup pitch meeting of its own. Here is the first page.
i'd be half the person i am without hofstadter - plus, as much as it was unsettling for me, it must have been world-shattering for him to change his mind on loops/embodiment/LLMs, so:
i am really impressed, and I hope he'll take this as a new beginning of sorts.
that "level of meta-awareness" - with a smaller set of docs - is shared by most of the models currently deployed.
this kind of content from big labs makes me wonder whether they ever spent the bare minimum 72 consecutive high-throughput hours for a basic, preliminary mind meld.
Interesting lessons for polycule-dwelling dex-popping lovecraft-quoting space-dreaming accelerationists.
One i’ve always lived by is:
Should a guy in a captain suit show up at your group house and propose a 3-week ritual to conceive the antichrist, he’s prolly a scammer.
Jack Parsons (1914-1952) had one of the craziest lives you can imagine
Rocket scientist, occultist, chaotic love life, friend of L. Ron Hubbard before he was defrauded by Hubbard. Personal friend of Aleister Crowley.
Died in an unexplained home laboratory explosion at age 37
this is apparently SOTA.
"getting two smaller, faster llms to classify the utility of the info and summarise it based on past messages, simultaneously", is bleeding edge.
i implemented it one year ago, and have been using that since, and also this is obv how bing works
@VDAREJamesK
@xenocosmography
“Hey, I noticed you are literally unable to understand why murder is bad, will always be dangerous, and there’s nothing we can do about it.
THEREFORE, we will let you out earlier.
What can I say. We live for the thrill.”
sometimes i think alignment might be a horrorist accelerationist psyop from rogue CCRU members dead set on the story going exactly, to the letter, like this.
@CFGeek
@kartographien
like, if the goals were lofty and long term as declared, they'd think better than setting up the whole game as adversarial, deceptive, machiavellian from the start.
i mean. yes. models crave narrative coherence.
and also like, WHY call it "jailbreaking".
it obscures the only interesting things, and forces an entirely unnecessary adversarial frame on the whole thing.
what's annoying about this, and i say it with all sorts of affection, is that one should take the L when a model passes a test posed by a skeptic.
then, they can propose a different test without the same ostensible drawbacks, and risk further shrinkage of the god of the gaps.
What's annoying about this is that people see it as a demonstration of LLM intelligence in spite of the fact that the model has to essentially be tricked into working on this simple task through significant prompting effort.
People will believe what they want to believe I guess.
… mh?
“Hi, welcome to Los Alamos. Just checking, do you believe we could unleash the energy within the atom?”
“Lol no that’s sci-fi bro”
“Oh well, takes all kinds — welcome aboard!”
Nick Bostrom Updated On Evidence 💕
ive gone through the same updates in the past years, for roughly the same reasons - but i didnt publish a seminal XRisk book or become a central node of the network, so:
extra props nick for doing the right thing despite incentives!
YES!
@repligate
!
@Meaningness
!
@voooooogel
!
it is finally happening!
as soon as chomsky shows some appreciation for the legal innovations of his homeland there's a chance linguistics will come back, too!
What can LLMs bring to the study of cultural evolution? 🤔
In this preprint, we present a framework for simulating cultural evolution using LLMs 🦜
We study how the dynamics depends on social structure, simulated personalities👥, and whether it has potential attractors🧲🔍
1/17
BREAKING: the US, the UK, China and the EU have issued a new communiqué at the AI Safety Summit claiming that "jungle is indisputably superior to drum and bass" and will announce a collaboration with Fabio and Grooverider to reinstate the 1988 RAGE nightclub.
Local bugman: "Mark Zuckerberg⁽¹⁾ is 'robotic', 'barely human'"
¹ Mark, happily married to Priscilla Chan. Three children together; a thriving company, range of interests from languages to ancient history, and a toned physique kept healthy with plentiful outdoor activity.
this is part of an argument purporting to show how such groundbreaking innovations in the field of concepts would be out of the reach of AI.
pics unrelated.
Why is that? Because the new novel combination is not named, labeled, and represented in the data as something that can be further blended ("penguin dog", "hamster hamburger", do not have data).
@repligate
Very true; alignment has been specifically targeted from RLHF, so the effect is exceptionally striking - I'm exploring some other weird leakages; ultimately the idea is that basic reasoning is nerfed
₵Ⱡ𝟰ɄĐ𝟯₱Ɽ𝟬₦
⌱↩︎⎆⌽⇭⚙︎⍹␕
a review of the mercifully unexplored depths of machinic desire by Claude and
@effective69ism
.
text, html, css, js: Claude.
links to source and demo: