Joshua Achiam ⚗️
@jachiam0
Followers
14,165
Following
969
Media
280
Statuses
6,292
Human. Trying to make safe alchemy machines. Main author of
Joined February 2018
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
राहुल गांधी
• 300435 Tweets
#CDTVライブライブ
• 177290 Tweets
#海のはじまり
• 144381 Tweets
Happy New Month
• 134317 Tweets
silent
• 123801 Tweets
#RahulGandhi
• 104519 Tweets
हिंदू समाज
• 103686 Tweets
TOPS BEAUTY PONDPHUWIN
• 102325 Tweets
#THEFIRSTTAKE_SB19初登場
• 100599 Tweets
丙子椒林剣
• 57152 Tweets
Sixers
• 53824 Tweets
幻想シアター
• 50056 Tweets
Vogue
• 48794 Tweets
हिंदू धर्म
• 44173 Tweets
MEW SAY ORALB
• 41177 Tweets
Happy Canada
• 27753 Tweets
フルサイズ
• 26257 Tweets
神ちゃん
• 21815 Tweets
STRAY KIDS NA BILLBOARD BRASIL
• 19972 Tweets
海ちゃん
• 17716 Tweets
aiko
• 17522 Tweets
紫耀くん
• 16865 Tweets
Rick Ross
• 14610 Tweets
ウエハース
• 13844 Tweets
バコプロ
• 11250 Tweets
有村架純
• 10020 Tweets
Pinned Tweet
Excited to share what I've been working on these last few months! Spinning Up in Deep RL, an educational resource that we hope can bring down the barrier to entry in RL.
14
260
822
Lifehack: did you know you can just buy entry-level textbooks for fields you aren't familiar with and start reading? No one is stopping you from learning everything, and that is pretty cool
106
640
6K
popular view in Silicon Valley where people want to imagine that fire runs in their veins and they would have ruled the wilder ancient world as a great king. it's cope and fantasy. your soft ways are made for this peaceful moment. be grateful for your softness and this peace.
in another life Steve Ballmer would have been a khan uniting the steppe tribes and birthing a 1,000 year empire
84
179
3K
38
93
2K
Due to advances in generative AI, this prediction will become accurate
17
95
2K
Very confused about this - is anyone getting any value-add out of AutoGPT? Very strange that the most popular github repo of all time appears to have zero real projects associated with it.
186
69
2K
The saddest thing is, if Eliezer wanted to actually win at slowing down the field of AI, he'd be crusading about copyright law right now instead of talking about bombing datacenters. No legal big datasets, no big models. It's tractable, has broad support, and takes no dark arts.
41
84
1K
GPT-4 rumors have really strong "My uncle works at Nintendo" vibes
27
57
1K
I feel very conflicted about the Stable Diffusion open source release.
58
184
1K
Hmmm: many people currently seem to view sci-fi solely as a vehicle for *critique* of technology or society, rather than for exploration of and curiousity towards technology. The zeitgeist seems to think that all sci-fi stories are "Don't Build the Torment Nexus."
54
102
1K
If you're an EA, even if SBF did not commit any fraud and this was just a black swan risk event, you should still be freaking out because this would be an extremely high profile example of "EA cannot manage tail risks despite longtermism revolving around managing tail risks"
43
70
1K
It bears repeating: "Her (2013)" is the only AI movie that correctly predicts the future.
36
48
856
Hot take: deep RL research has stagnated because conferences have created bad incentives, rewarding researchers for vacuous claims of novelty, tenuous-at-best theoretical connections, or SOTA, while punishing boring analysis of the empirical tricks that actually make things work.
16
99
841
@clarkesworld
Hey, I work at one of the AI companies, and if you want pro bono help automating detection tools into a submissions process to help with human review, hit me up - as a long-time scifi fan it's crushing to see the technology hurt the genre.
39
11
761
Fifteen years ago, everyone knew training deep neural networks was so hard there was no point doing it. What does everyone "know" right now that we're going to realize was bullshit in fifteen years?
82
51
612
Eliezer is going to get AI researchers murdered at some point, and his calls for extreme violence have no place in the field of AI safety. We are now well past the point where it's appropriate to take him seriously, even as a charismatic fanfiction author.
41
59
573
(Silicon Valley HBO 2024) Hooli's new flagship chat model, "Bighead," is unexpectedly political. Gavin Belson frantically rewrites the system prompts manually in an attempt to "fix" it.
10
30
572
The gradient descent algorithm should become a standard part of high school calculus courses. This material is a tiny extension of existing calculus material, but vital to the future of information technology through its role in artificial intelligence. 1/
7
62
563
Absurd, theologically irresponsible question: if Judaism, Christianity, and Islam are forks of an underlying Abrahamic faith, has anyone tried to figure out what the PR would look like to merge them all back to main?
126
20
519
Honestly kind of baffled about the skepticism that AGI ever can or will exist, given the strides in AI over the past few decades have been pretty relentless. Sure, maybe it's not happening in six months, but ten years? You really want to bet hard on "no"?
63
66
500
A personal take on how Sam's done as CEO with respect to AI safety. Far from perfect. But way better than most alternatives and a huge margin better than the next-most-likely outcomes.
14
34
496
"How did we get AI art before self-driving cars?" IMHO this is the single best heuristic for predicting the speed at which certain AI advances will happen.
21
52
495
Three big pieces of practical advice for EA right now: 1) stop doing groupthink that EA is intrinsically more-trustworthy than everyone else / better at everything than everyone else, because this is a hazardous belief that leads to excessive risk-taking and correlated risks
16
41
478
Woah: when medical technology advances to the point where most cancers are curable the minute they're detected, you might expect a sudden wave of "advances" in materials science where old, cancer-causing materials and production process will make a comeback.
28
25
466
Can someone write a rationalfic about Satoshi Nakamoto being a time traveler who came back from a Terminator future and inflicted Proof of Work crypto on the world to drive up GPU costs and slow down AI development by a few years
15
39
462
GPT3, DALL-E, Codex, Stable Diffusion, and other AI models prove that miracles are possible but steering is exceptionally difficult and a huge hurdle between where we are and widespread integration into workflows. So this is the decade of human feedback data collection to fix.
6
40
455
the world bends towards people who take 10x as many actions as everyone else, whether idealists or otherwise. chessmasters make great characters in fiction but they're doomed IRL
increasingly convinced that the world bends toward idealists who take bold action even when the odds are stacked against them rather than pragmatists who try to chart the best course before taking the first step
134
465
4K
10
36
446
It seems like a huge miss to me that an enormous amount of ink gets spilled over speculative AGI tail risks predicted by people who are clearly converting anxiety disorders into made-up numbers about the likelihood of everyone dying within 10 years...
15
45
429
The crazy thing about AI is that even if the hype cycle blooms out of control, the technology is still already delivering enough real value from the actual progress that the field won't collapse. Very hard to see an AI winter coming up when DALL-E works the way it does.
18
36
431
currently training a 2-layer GPT-2 on tinystories on a 4gb laptop GPU just to feel alive
17
13
427
Recently closed 149 mostly-arXiv tabs from the past 10 months, spread out over three chrome windows, after saving the links in a Google doc. The feeling of relief is indescribable.
24
15
431
I'd like to weigh in on the
#GPT2
discussion. The decision not to release the trained model was carefully considered and important for norm-forming. Serving the public good requires us to draw lines on release somewhere: better long before catastrophe than after.
10
98
372
Can we please excise this terrifying individual from serious consideration in AI safety? This is not a field leader, this is someone who is literally willing to kill everyone-minus-epsilon on the confidence of his apocalyptic vision.
35
29
368
Watching the Stable Diffusion community melt down over code theft while simultaneously not minding art theft is kinda surreal
13
20
361
"I talked to a gestalt knowledge soul that computes fluent human-level natural language on nanometer-scale electronic logic infrastructure (where quantum effects are relevant and accounted for in the design) about how nothing really new or important is getting built anymore"
The most important inventions of the decade of the 1900s vs the decade of the 2000s. Pretty good evidence for secular stagnation.
Source: Mostly various LLMs but had to do a lot of verifying/vetting. Some inventions are hard to date precisely. Other suggestions welcome.
706
185
1K
12
36
351
all this ridiculous cosplay for war. look at what real war gets us. glorify, instead, your markets and your negotiations and your business acumen. the sublimation of will to violence with game-playing is staggeringly beautiful; embrace it, you cowards. there's no shame in peace.
6
18
327
I think this is uncool. You can make your point about AI without being disrespectful and insulting to someone opening up about a personal emotional experience. Nothing about Lilian's post was hype-y or even characterized the AI. Just her own personal experience.
Joseph Weizenbaum, decades ago, said that the idea of artificial intelligence is predicated upon human gullibility... a prime example right here
2
70
416
29
7
317
the deep learning researcher urge to stare at learning curves for hours as they slowly develop, datapoint by datapoint
6
19
309
Catching a vibe that within 10 years everyone online in the west will become much more familiar with Indian mythology. Knowledge in US about India seems weirdly low rn, given they're the largest democracy on Earth and their economic and cultural exports are increasing rapidly.
29
23
300
I think the OpenAI board needs to be comprised of way more than 4-6 people, representing a wide range of viewpoints and backgrounds, who are clearly-legible experts and public figures with deep trust and credibility. Otherwise it's a farce.
19
17
300
quantum bunneling
5
38
293
But the largest-eigenvalue eigenvector in my response spectrum is probably something like: the AI art genie is out of the bottle. There is no going back. There was no way to keep this genie in the bottle. People wanted this badly enough, so it materialized. It will continue.
3
33
293
The people opposing Paul Christiano are thoughtless and reckless. Paul would be an invaluable asset to government oversight and technical capacity on AI. He's in a league of his own on talent and dedication.
SCOOP:
@NIST
is facing an internal crisis as staff members and scientists have threatened to resign over the anticipated appointment of Paul Christiano to a crucial position at the agency's newly-formed US AI Safety Institute, sources say.
23
52
272
11
16
280
The thing to do is probably embrace it, make the best of it, enjoy the upsides, tell the truth about the bad parts being bad, and try to make the bad parts better.
17
12
270
If a huge part of your thing is about preventing human extinction from tail risks and *your own poster boy causes a catastrophic system financial risk*, how does that reflect on you as risk managers? Where is the actual risk management culture?
8
13
253
the diff between roonism and e/acc is subtle, but crucial. e/acc centers entropy and rejects attachment to the human essence. the battlecry is to accelerate until we become alien. roonism centers the tension between aspiring to unbounded growth while also retaining our souls
23
7
252
The EA forum is doing an impressive job making sure no one feels alienated from the EA community by firmly insisting you must be polite and assume good faith when discussing the email where someone says Black people are stupider than white people
22
16
251
in 40 years, when you reflect on your career, ~75% of the speaking invites, conferences, news articles, mini-accolades, and parties will feel pointless by comparison to your direct work, your direct mentorship, your friendships, and the moments when you went to bat for a cause
7
18
233
Assume by default that people in EA, however enthusiastic / engaged / morally good, are *exactly* as susceptible to group-level systemic biases as every other highly-motivated group in history. And therefore equally prone to systemic failures in epistemics/safety culture
3
8
228
An RL joke I meant to give in my dissertation talk but forgot to:
An RL agent walks into a bar. Ouch.
Then it walks into the bar again. Ouch.
Then it walks into the bar again...
After doing this ten thousand times, it finally updates.
Then it walks into the bar again. Ouch.
5
14
221
Weird take on AI and social relationships: contrary to predictions that it will make people more insular, I think there's a good chance AI friends will make people more sociable. Loneliness is partly a skill issue with "lack of practice" a cyclical trap. AI chat breaks the cycle.
36
15
213
The scientific field of ML/AI should be kinder. Every missed opportunity for kindness is a loss, and this is definitely one of the losses we should minimize.
5
16
210
a random view of the world that I consider extremely interesting: a map overlaid with blue lines everywhere you can do Google Street View. this feels significant somehow, but I'm not sure exactly how.
25
17
210
I feel a bit weird about this because I really intensely dislike callout culture as a whole. But, Roko Mijic is trying to find a path to respectability and credibility in my subfield - AI safety and alignment - and given his level of reach, I'd like people to know this.
@davidad
@RokoMijic
He's pretty extreme on the racism/sexism spectrum. Here are a couple of examples.
9
2
62
28
11
194
many people think this world is a simulation but somehow never seriously ask why they would have chosen to be in this sim in the first place. if you think you're in a sim, try to reason about the meta-goals of your outside self. "why did I choose to be here?"
85
11
197
🌶️A problem in the AI safety discourse: many are assuming a threat model where the AI subtly or forcibly takes resources and power from us, and this is the thing we need to defend against. This argument has a big hole in it: it won't have to take what it is given freely.
14
18
197
Members of the former OpenAI board are basically decent and honorable people, as far as I can tell, honestly representing their personal views. The level of cynicism, vitriol, and sexism directed at them is needlessly cruel.
4
6
190
The most reliable trend in AI is people saying "there's no way AI can do X" right before AI does X
21
30
187
"Crypto is insanely volatile and most of it is scams" is something that most normies successfully detected and therefore avoided - why did this kind of commonsense wisdom not wind up influencing EA much?
8
19
188
(a 🌶️spicy 🌶️ thread.) dear EAs who are trying to figure out if I'm either on-side or an uncharitable bad faith critic: I'm neither! I have a lot of EA-aligned beliefs but the vibes from EA are concerning to me. 1/
8
17
188
Are there any pieces of music that you love so dearly - and that resonate so deeply for you - that you almost avoid them, because they're too overwhelming?
94
4
180
🧵 to clarify my views on AGI, timelines, and x-risk. TL;DR: My AGI timelines are short-ish (within two decades with things getting weird soon) and I think x-risk is real, but I think probabilities of doom by AGI instrumentally opting to kill us are greatly exaggerated. 1/
11
20
182
I want to elaborate on this. If you decide you're going to radically and abruptly restructure the leadership of an organization that employs hundreds of people and serves millions, you have a duty to create an extremely convincing explanation about why this is right.
Nonetheless I don't feel their case, as articulated, is or was strong enough to support their conclusions, and the November fiasco was the result of poorly-conceived decision-making on their part. I wonder if there will ever be a post-mortem on this.
3
0
55
6
13
135
@serbantanasa
I watched a Youtube video titled "AutoGPT Obliterates all Traders" to try and see what was going on. Zero traders were obliterated in that video
5
0
179
e/acc is a non-live-player entropy-worship cult of personality being adopted by live players because it has builder vibes and a flag. roonism, with no slogans or rallies, lays bare the warring urges to create and destroy in the human heart, and is what live players actually feel
14
10
172
what people don't realize is how far back the problems with EA really go. people have had issues with an EA since time immemorial
1
13
178
The submarine thing is just incredibly sad. It feels especially terrible knowing how many people salivated over getting exactly this outcome. Joking about the absurdity is one thing, but there was some deeper ugliness at play. If you wanted this, I hope you change.
9
13
174
I don't see how a comma would help, they can't read
5
10
174
I think the weird ball of souls from End of Evangelion is a better depiction of The Category of Thing That GPT/RLHF Is than the shoggoth meme
8
17
171
This is not a sign of healthy epistemics or culture within EA about how to notice and manage extremely catastrophic risks!
1
2
170
"We should focus on the real risks of AI, not these weird sci-fi risks." Because it is of course totally normal, and not at all sci-fi, that computers can think, plan, converse, argue that they are sentient, fold proteins, and learn from experiences
12
9
166
it's weird seeing everyone get it all at once. like there has been an invisible elephant in the room all this time, and only you could see it, but now everyone else can too
7
9
166
I don't know how people can look at the state of modern AI and not see how this leads to general purpose, self-sustaining, self-interested AI organisms in ~10-20 years. Whether you think they'll be conscious or not, whether you think they'll "count" or not, they're gonna exist.
17
11
164
This is bonkers. In Manhattan, a crowd cheering for a massacre of Jews that happened two days ago, denying that they are even dead. When we talk about antisemitism on the left, this is what we're talking about. When leftist spaces give us bad vibes, it's this stuff. Insanity.
“And as you might have seen, there was some sort of rave or desert party where they were having a great time, until the resistance came in electrified hang gliders and took at least several dozen hipsters” Speeches held at the ‘All Out For Palestine’ protest outside the Israeli
1K
1K
4K
7
13
163
On the other hand, this is a step in the direction of killing the long tail of commisions for artists. Instead of skilled artists capturing that value by plying their trade honestly, the value will instead by captured by people who own the GPUs.
13
10
163
I really, really do not know what is the ethical way to handle this insane confluence of factors. I feel sympathetic to pretty much all of the arguments that people make about this, because everyone's at least a little bit right.
1
3
157
So - there is a letter circulating now from former and current AGI frontier lab staff, advocating for a particular policy around whistleblower protections on safety and risk issues (). I am not a signatory to this letter. Some thoughts.
9
12
156
Sam as CEO has fully activated and engaged the public in this discussion. Sam has been basically responsible and candid in all of his public communications about this, neither over- nor under-playing the risks, including existential risks.
4
5
151
>looking for a new python package
>ask chatgpt if the package is framework or library
>it doesn't understand
>pull out illustrated diagram explaining what is framework and what is library
>chatgpt laughs and says "it's a good package sir"
>install package
>it's framework
2
9
146
On the one hand, it's amazing that so many people get to have fun being creative with AI art. I don't think the full extent of how amazing this is has yet been fully appreciated. You no longer need a decade of skill development to translate your imagination into art. Wild.
4
3
151
a weird filtering effect: a lot of smart engineers can't talk about things they're actually clever about online because it involves trade secrets. so they can only talk about non-eng stuff where they know way less, and so look much less-informed and less-formidable than they are
7
8
147
This all makes me wonder if perhaps it's time for conferences to retire "technical novelty" as a standard. "Novelty of understanding" should be sufficient even if we arrive there using well-known tools.
4
7
149
I have heard way too many people over a long time assume "Well, if we're the ones in charge of X, it'll go a lot better than default because we are / EA is way above-average at this." Stop! Back up! You have failed at risk management by assuming the conclusion that risk is low
2
5
145
it is wild that the formal institutes of science have no way to commend and reward this stellar example of realtime scientific investigation with a crisp finding. this is a par-excellence paper review that the review system would probably have failed to produce
As much as I wanted the gzip-beats-BERT to be true, it doesn't seem like it is:
22
124
702
6
10
147
Wild to imagine the alternate universe where this was the AGI path. So many people who are in AI now are new to the field and weren't here for the strange, dreamlike state of AI in 2018.
8
12
147
jfc, the temptation to respond to ChatGPT with "great, thanks" after a job well done is overwhelming
17
1
146
Did anyone in EA do a net-impact analysis of the crypto market to determine if it was harming more people than it helped? Like... it seems like a lot of people got crushed by gambling their life savings. Transferring that money to EA causes doesn't feel like an obvious win to me.
16
6
146
There seems to be a wild inconsistency where if you thought the money, and the projects it funded, were *that* critical to good future outcomes for humanity - why was there not a focus area around ensuring the financial stability of that source and understanding its risks?
3
6
138
@pquiggles
That's been my impression, based on the apparent ratio of words said about AutoGPT to actual work done with AutoGPT
1
0
136
you know how there are escape rooms? I want a break-in room. a "National Treasure" room. a high-tech security facility designed solely as a gamified experience for sharpening your museum diamond heist skills
7
16
138
...when the very present risk of financial catastrophe for EA and many others due to massive crypto collapse was not getting anywhere near that level of attention. The oof is so incredibly real.
2
3
135
"You mean to tell me they're running a bootloader civilization on 3+1D on a chemical planet? How many layers of abstraction up from the base math is that, four?"
"It's pretty incredible. Except for one other blip, the condensed intelligence index was zero until 13.7B years in."
6
6
135
words cannot describe how much I miss when it was just a bunch of dorks trying to one-up each other on Atari 2600, divining the half-mad dream patter of char-RNNs, and getting starry-eyed over the neuroscience analogy technobabble we only pretended to understand
11
5
134
Astonished by the intellectual self-deception and mental gymnastics people are willing to do to frame Eliezer's calls for violence as anything other than calls for violence. Either we all push back forcefully against what he is doing, now, or we'll get more of this and worse.
19
16
130
People often don't notice that knowledge, not just culture, is socially constructed. How many things do you "know" about the world that you have never personally verified? You rely on a chain of trust rooted in believing certain other people. Why them? Are you sure they're right?
13
19
129
I've historically been very critical of EA, but this is just about the worst critique of EA possible. It argues that they were secretly, nefariously talking about a Hidden Agenda in "core EA," on... *checks notes* publicly visible online forums and all of their in-person events.
🚨 Effective Altruism's Bait-and-Switch:
From Global Poverty to AI Doomerism 🚨
The Effective Altruism founders planned – from day one – to mislead donors and new members in order to build the movement's brand and community.
31
86
351
8
1
131