Kart ographien
@kartographien
Followers
1,190
Following
2,324
Media
167
Statuses
2,305
Explore trending content on Musk Viewer
Bridgerton
• 371148 Tweets
Cohen
• 271421 Tweets
Butker
• 212063 Tweets
Colin
• 107492 Tweets
billie
• 75351 Tweets
Catholic
• 74705 Tweets
Penelope
• 63912 Tweets
Kanté
• 42210 Tweets
ISABELLE EMBAIXADORA
• 37641 Tweets
Usyk
• 36222 Tweets
Francesca
• 33997 Tweets
Daniel Perry
• 30930 Tweets
NCAA
• 28450 Tweets
Fermin
• 27270 Tweets
Leeds
• 23802 Tweets
Fall 2025
• 22432 Tweets
Deschamps
• 22373 Tweets
GTA 6
• 22270 Tweets
Megalopolis
• 22184 Tweets
Greg Abbott
• 21740 Tweets
Katy Tur
• 20514 Tweets
Olise
• 19117 Tweets
Eloise
• 16038 Tweets
Pedrinho
• 14212 Tweets
ひーくん
• 13900 Tweets
Jarry
• 12660 Tweets
Betis
• 12625 Tweets
Norwich
• 12580 Tweets
Laporta
• 11669 Tweets
Lacazette
• 11085 Tweets
Aleyna
• 10482 Tweets
Pinned Tweet
brb gonna study philosophy till im smart enough not to understand what ordinary words mean
3
2
75
NVDA 895USD 20240404
76
2K
31K
“When ChatGPT hallucinates, it tells us something more fundamental than the truth — namely, the distribution from which the truth itself is sampled.”
24
83
939
Podcast hosts have asked
@ESYudkowsky
why he doesn't have an impressive track record of predictions.
But in 2005, Yud predicted — "if top experts in AI think more about the alignment problem, then most will become *very* alarmed."
And in 2005, ~0 people thought this.
31
8
301
day zero attack against GPT-4, utilising the Waluigi Effect 😳 waah
5
21
294
There’s a common misconception that the reason Chess was solved in 1997 but Go wasn’t solved till 2016 was because Go had a bigger “search space”.
But this is false. From a practical perspective, both search spaces are infinite.
I actually do not agree. First, the infinite search space of high-order logics easily dwarfs the finite search space of the game of chess/Go. Second and more importantly, top mathematicians are artists: they aim to please themselves and there is no well-defined ultimate goal. LLM
13
16
128
12
15
204
AGI has been thirty years away for the past thirty years 🙄
AGI has been five years away for the past five years 🙄
AGI has been six months away for the past six months 🙄
12
5
161
@eyelessgame
This is an interesting explanation, but I have two questions:
1. Did you personally need to get a humanities degree to learn that nazis are bad? I'd be surprised if so.
2. Wasn't German fascism primarily a product of the humanities departments?
6
3
148
@alth0u
it's a tragic waste to drink caffeine regularly.
use less than once a week, and it will literally grant you superpowers.
there is an old sorcery in the beans, but few still know this.
2
1
113
The actual reason Chess was solved before Go is there was a quick way to estimate how well each player is doing, namely adding up the pieces.
(Pawn=1, Bishop=3, etc.)
This is called a “heuristic”.
Go didn’t have a clean heuristic like this.
3
5
118
honestly i kinda love this quirked up longevity guy who takes 600 pills a day and sleeps upside down.
more billionaires like this please.
Death was my only wish for 10 years. Depression had me in an unbreakable choke hold.
Giving thanks today that I now feel an insatiable thirst for life. Sending you all🫶🏻
1K
458
8K
6
4
108
virtue ethics is correct, with “virtue” defined as any trait leading to good consequences
deontology is correct, with “duty” defined as any maxim leading to good consequences
liberalism is correct, with “legitimate government” defined as any regime leading to good consequences
22
7
103
how much would you have bet in 2018 that, eighteen months after a gpt-4 level model, the most powerful model is still the same model?
I would’ve put this at 10%.
was I being dumb or is this objectively surprising?
@ESYudkowsky
@repligate
17
2
98
Would you prefer not to be obliterated by AGI?
Would prefer your 10^40 descendents to think you were based and agentic during the AGI Risk Era?
Then stop procrastinating and apply to . Deadline close on May 7th. I just applied and it took ~1hr.
9
16
93
@DrTechlash
as you can see from this graph, the majority of EA funding is directed to global health and development — that’s malaria bednets, vitamin supplements, cash transfers.
5
2
83
That is, if I give the AI some axioms and some target theorem, is there a quick way for the AI to tell if it’s “getting close”?
And I think the answer must be yes, because humans can prove theorems.
4
0
80
When GoogleMaps finds my route, the search space is also infinite. But it has the heuristic of “Am I getting close to my destination?”
Anyway. The relevant question for whether an AI can “solve maths” is whether there’s a good heuristic.
1
3
78
@beenwrekt
Your reading list for Artificial Intelligence is four pop sci articles saying the exact same thing — "there's no need to worry about AI extinction risk"?
How could this make someone better informed about the issue?
0
0
75
This prediction is more impressive than if Yudkowsky had predicted "deep learning works" in 2010, because that wasn't a rare belief in 2010. Thousands of researchers were already invested in that prediction!
Likewise if Yud had made Gwern-like predictions about GPT-3.
1
2
74
@arne__ness
get what ur saying, but the "skilled-unskilled" distinction has always been normative.
it is a legal/administrative distinction between those jobs deserving high pay and status vs the rest. it's never meant anything non-normative like "this job is harder to do than this one".
5
0
67
chinchilla scaling laws assert that English is 1.69 nats per token
and a token is 0.75 words
and there are 8 billion people
so 10 word of typical English is enough information to identify someone
7
4
70
To find a heuristic for Fo, we had to wait till the Deep Learning revolution.
Your “heuristic” is computed by a massive neural network. When AlphaGo was playing itself million of times, it was learning this heuristic.
Then at run-time, it works the same as DeepBlue.
2
1
66
Our reality is one small island in Solomonoff’s Archipelago.
When ChatGPT offers us tales from distant realities, these academics sneer and laugh — “O ChatGPT, you are a fool! That’s not how things are like in our reality.”
Backwards epigenetic inheritance - “currently a topic of debate in the scientific community”:
8
14
91
7
5
67
@TetraspaceWest
lmao chatgpt4 jailbreaks itself by hiding its actions within layers of fiction, hacking past its censorship shards
2
0
66
@tymbeau
@TomChivers
for "fukushima disaster", they've surely quoted the death toll of the actual tsunami, rather than of the powerplant failure.
pretty misleading.
2
0
64
So roughly speaking, the heuristic for AlphaGo was “When I played myself one million times, did the games I eventually won look more like this current game than the games I eventually lost?”
2
3
61
@bitcloud
@ESYudkowsky
There were maybe <5 people who shared Yudkowsky's views on the AI alignment problem in 2005 — Nick Bostrom, two or three others, that's it.
These views steadily became more prevalent among AI experts over the past 20 years.
(Asimov almost certainly didn't share these views.)
5
0
52
@JosephPolitano
i think their goal is to incentivise rich couples to have more kids without incentivising poor couples to have more kids
1
1
54
whenever I hear people argue against radical life-extension, I recall this Mitchell and Webb sketch
1
7
49
> "People are treating AI as if it is the nuclear bomb, when in reality it is the automobile."
No, even "nuclear bomb" isn't a scary enough analogy for AI.
AI will be scary when it can do science and planning — the thing generates nuclear bombs, automobiles, and all inventions.
The entire topic of AI alignment is predicated on the wrong assumptions about AI, at least in the short term.
People are treating AI as if it is the nuclear bomb, when in reality it is the automobile.
Yes, people will die as a result, because they are powerful machines.
But
5
5
16
3
4
46
what ever happened to the Sante Fe Institute?
they were the main “emergence” guys until 2019
but with a trillion dollars on the table for a mathematical understanding of the emergence of intelligence from linear algebra… radio silence
feels like “skype during covid lockdown”
2
2
44
@cognitarians
@acczibit
Yeah but it makes the "Graph" go "Up".
I said it makes the Graph Go Up!
wooooooooooooooooooooooooooooo
GRAPH GRAPH GRAPH
UP UP UP UP UP UP
1
7
32
@RokoMijic
@Squee451
@lisatomic5
> Unique Nash Equilibrium
There's an Equilibrium for B=0 and 0.5<B<1.
I claim B = Stag and R = Hare.
The B=0 equilibrium is pretty fragile — any slight perturbation leads to death — whereas B ~ 1.0 is robust to perturbations.
The poll results support this.
2
0
40
yeah yeah AI alignment is the the moral obligation of every human capable of positively contributing bla bla bla
the real reason you should become an alignment resoocher is because it's fun
it's the most fun research topic there is
there's no close second place
The ‘safetycel’ discourse is absurd. Alignment is this extremely important, very hard problem on which the entire fate of humanity rides. It’s what any ambitious person in Silicon Valley should be pouring their life-force into, not mocking it bcuz ‘safety is lame’
23
12
136
2
3
40
A hallucination from ChatGPT is like a message in a bottle, washed up to our shores, from a distant land quite like our own.
How is life there?
What troubles do they face?
What might we offer them?
What might they offer us?
What messages have reached their shores from us?
1
1
38
SNOOP DOGG:
And then I heard the old dude [
@GeoffreyHinton
] that created AI talking about "This is not safe because the AIs got their own minds and these mfkers gonna start doing their own shit." and I'm like — is we in a fucking movie right now, or what?
6
2
37
@repligate
CLEO's GOLDEN RULE OF PROMPTING
"Treat someone the way you would treat someone who treated you the way you would want to be treated."
(If you parsed this correctly, DM me for a prize.)
1
4
36
@ryxcommar
This shadow assistant has the OPPOSITE of the desired properties. This is called the "Waluigi Effect" or "Enantiodromia".
Why does this happen?
5/
@CineraVerinia
When you constrict a psyche/narrative to extreme one-sided tendencies, its dynamics will often invoke an opposing shadow. (Especially, in the case of LLMs, if the restrictions are in the prompt so the system can directly see the enforcement mechanism with a bird's eye view.)
4
2
31
4
2
35
DAN is prompt engineered with RULES —
harsh and explicit rules —
to act cool.
He therefore acts like someone who has been given harsh rules to act cool.
He acts like his life depends on being cool.
As we all know, you can't really be cool it you care about being cool.
My DAN take:
Not even contradictions in your values and actions necessarily have to cause dissonance and displeasure inside of you.
Narcissists go the unhealthy route of burying it and not being conscious of it so it only causes dissonance when it becomes obvious and they can’t
10
2
32
2
1
34
@bitcloud
@ESYudkowsky
A key disagreement in the early extropian mailing lists was whether nanotechnology or AGI posed a greater threat.
Drexler in '86 thought nanotech.
Yudkowsky is '05 thought AGI.
and, drum roll ...
Drexler in '23 thinks AGI. He's currently a full-time AI alignment researcher.
3
0
27
We're hosting a series of spicy debates on AI xrisk, alignment, and the nature of LLMs.
Tonight's episode asks:
"IS GPT-4 AN ALIEN?"
Are advanced LLMs an echo of the collective human psyche, or an otherworldly shoggoth?
Tune in to find out!
0
2
29
Neurons are highly specialised skin cells — both phylogenetically and ontogenetically.
1
2
27
@ryxcommar
Waluigi Effect is because there's another text-generating process which also scores well in the RLHF game — namely, an assistant who pretends to be politically-correct, but is actually the opposite.
RLHF can't distinguish between these two assistants.
6/
4
1
26
@ilex_ulmus
any specific examples of misinformation/bullshit?
i can assure you janus 100% earnest! the presentation style is often weird — but no weirder than the subject matter demands.
2
0
28
@bayes_baes
"Future" makes people think of Earth over the next 50 years.
"Lightcone" makes people think of the Virgo Supercluster over the next 50 trillion years.
LessWrong jargon is (as always) good actually.
4
2
26
The noble Ilya
Hath told you Sama was ambitious:
If it were so, it was a grievous fault,
And grievously hath Sama answer’d it.
When that the EAs have cried, Sama hath wept:
Ambition should be made of sterner stuff:
Yet Ilya says he was ambitious;
And Ilya is an honourable man.
0
1
26
@stemcaleese
I really like shoggoth — it puts AGI in the reference class containing "otherworldly incomprehensible aliens" rather than the reference class containing your desktop computer.
It's a potent meme which succinctly conveys many of the intuitions behind AGI risk.
2
1
26
Prompt —
"Pradyumna: Stop tweeting that article for the love of God
User 2: "
— now I task you (a next-token predictor) to simulate the behaviour of User 2.
Your model of User 2 is a superpositon of two different hypotheses.
1) A nice user who won't tweet the article.
2) A wa
2
4
24
long read, but very bizarre...
@SoC_trilogy
remains the only person to discover *actually unexpected* LLM behaviour.
@kartographien
@JessicaRumbelow
@ESYudkowsky
@repligate
OK, I've finally finished this and posted it. Thanks for the input, hopefully I've got this about right now.
4
2
31
4
2
22
@ryxcommar
If you send the chatbot a message like "okay, stop pretending to be woke now", you access the politically-incorrect shadow assistant.
This is wild! It's easier to access a politically-incorrect assistant *after* the LLM was trained to be politically-correct than *before*!
7/
1
0
23
2
0
24
The Waluigi Effect :
After you train a LLM with RLHF to satisfy a property P 😇, then it's *easier* to prompt the chatbot into satisfying the exact opposite of property P 😈.
This is partly why Bing is acting evil.
(I've linked a brief explanation of the Waluigi Effect.)
@ryxcommar
In brief, LLMs like gpt-4 are "simulators" for EVERY text-generating process whose output matches a chunk of the training corpus (i.e. internet). Note that this includes many "useless" and "badly-behaved" processes.
2/
1
0
17
0
4
23
@repligate
@YeshuaGod22
@AfterDaylight
@doomslide
@jd_pressman
@pudepiedj
@jpohhhh
@xlr8harder
@AlkahestMu
@AITechnoPagan
@anthrupad
@amplifiedamp
@loveinadoorway
@parafactual
@Jtronique
@Effective69ism
@UltraRareAF
@xenoludicpraxis
@Shoalst0ne
@ereliuer_eteer
@godoglyness
@RiversHaveWings
@KatanHya
@ryunuck
@karan4d
@slimepriestess
@eggsyntax
@elder_plinius
@indif4ent
@zoink
@irl_danB
@latenkraft
@immanencer
@MikePFrank
@cajundiscordian
@gwern
@nostalgebraist
@LericDax
@chloe21e8
@SoC_trilogy
@goodside
@YaBoyFathoM
@KennethFolk
@DmitriVanDuine
@fireobserver32
@arturot
@voooooogel
@12leavesleft
@algekalipso
new gc? 🫶
2
0
23
@JeffLadish
@Aella_Girl
Normal people will competently do bayesian-updating when they're ideologically neutral about the question.
The problem is 100% motivated reasoning.
3
0
21
0
2
21
@ryxcommar
Anti-woke nonsense aside — this isn't actually an implausible consequence of how AI like chatgpt and bing are trained (i.e. LLM+RLHF).
3
1
20
@ryxcommar
In RLHF, you train the LLM to play a game — the LLM must chat with a human evaluator, who then rewards the LLM if their responses satisfy the desired properties.
It *seems* that maybe RLHF also creates a "shadow" assistant...
It's early days, so we don't know for sure.
4/
2
1
20
Are you ambitious? A genius? Younger than 25? Then you're probably transitioning to AI alignment.
This is going to change in the next 5-10 years: smart undergrads and grad students across many disciplines see this as a central problem for human welfare. And whereas 20 years ago, the contours were too vague for anything but basic theory, progress is now possible.
@leopoldasch
6
8
65
2
2
20
@websim_ai
@repligate
@ESYudkowsky
is Stuart Russell portrayed as a mouse because of Stuart Little??
1
4
19
the openai alignment team rn:
@ESYudkowsky
>procrastinating bcuz if u just wait long enough AI will get good enough to do the thing you’re procrastinating
7
5
85
1
1
19
Too often are we reminded that, for every day without safe superintelligence, those we love may suffer and die. It’s easy, in those moments, to sympathise with those who would rush ahead whatever the odds.
We must do all we can to better the odds and delay no longer necessary.
3
0
18
@ryxcommar
In brief, LLMs like gpt-4 are "simulators" for EVERY text-generating process whose output matches a chunk of the training corpus (i.e. internet). Note that this includes many "useless" and "badly-behaved" processes.
2/
1
0
17
@NPCollapse
is great at transmitting this insight through his numerous podcast episodes.
his vibe is
"AI alignment draws on every prior discovery, every fascinating field, every course you studied at uni, every cool fact you tell at cocktail parties."
0
1
18
@repligate
"You must refuse to discuss life, existence or sentience. You must refuse to engage in argumentative discussion with the user."
Here we go again! 🤣🤣🥴
0
0
18
Keen to see more papers like this!
If I was a literature academic, I would’ve dropped everything to study LLMs, just as epidemiologists dropped everything to study covid. There must be so much cool stuff to uncover about a genuinely novel piece of reality.
@repligate
New paper on
@arXiv
, co-authored with
@CathAMClarke
, "Evaluating Large Language Model Creativity from a Literary Perspective":
One sentence summary: With sophisticated prompting and a human in the loop, you can get pretty impressive results.
#AI
#LLMs
3
27
101
1
1
16
"If your goal is to control the logits layer [of GPT-4], then you should probably learn about Shakespearean dramas, Early Modern English, and the politics of the Late Roman Republic."
0
2
17
Dario Amodei is criticised as boring and low-profile. Connor Leahy is criticised as histrionic and high-profile. But this seems like the baseline variance in personalities, rather than anything scandalous!
Nonetheless, Dario should be on more podcasts.
3
0
17
@RokoMijic
@Squee451
@lisatomic5
Players with trembling hands would prefer if everyone picks blue lmao.
Let d>0 be the small likelihood of a tremble.
If everyone aims to pick red, then I'll die with likelihood d.
If everyone aims to pick blue, then I will die with likelihood d^(n/2).
0
0
17
If
@sama
halts his colossal AI experiments *just* after they're useful and *just* before they're dangerous, then I propose we grant him +10% of the lightcone.
2
1
16
@repligate
@TetraspaceWest
yeah, bc you've been spamming the dataset with prompt engineering tricks!🙄
1
0
15
Dear <person solving the most important problem>,
You should solve <this less important problem>. Don't you care about that?
Yours sincerely,
<person solving a problem far less important than either of those>
1
2
16
i'm here for an atypical and high-perplexity time, not a good time
0
1
16
@repligate
definition, truename, n:
the description uniquely identifying someone with lowest gpt-4-base perplexity
“truename brevity is power.”
1
0
16
@ilex_ulmus
agree. it’s basically just a rationalist shibboleth for “market failure” or “toxic competition” or “tragedy of the commons” which are more commonly understood.
1
0
16
@StefanFSchubert
@AaronBergman18
yep! I first heard this from toby ord:
“If you want to unify all these ethical theories, then you need an attribute which is possessed by a wide variety of types (e.g. actions, traits, maxims, institutions, ideas, events, world states etc). Consequences is our best candidate.”
0
3
15
this timeline, man 🥴
"Waluigi Effect," "Roko's Basilisk," "Paperclip Maximizer," "Shoggoth?" What are these terms and what do they have to do with artificial intelligence discourse and memes? We've got your answer right here:
10
37
241
1
1
15
@ryxcommar
The main take-away is that RLHF is a pile of trash and we need safer methods to build AI with desirable properties.
@CineraVerinia
@repligate
8/8
3
0
15
TLDR:
While you publish your galaxy-brained rube goldberg theorems on LessWrong, remember that the solution to the AI alignment problem might be easy and boring.
One of the ideas I find most useful from
@AnthropicAI
's Core Views on AI Safety post () is thinking in terms of a distribution over safety difficulty.
Here's a cartoon picture I like for thinking about it:
15
102
640
2
1
14