Adrien Ecoffet
@AdrienLE
Followers
6K
Following
22K
Media
53
Statuses
2K
Researcher at OpenAI.
San Francisco, CA
Joined April 2009
Would you prefer working for a company whose name implies it's open (but turns out to be closed), or one whose name implies it's stable (but turns out to be unstable)?.
31
22
574
After careful consideration, I have decided that the most productive way to spend my time today is to turn on Do Not Disturb and play video games.
7
12
501
I spent the whole weekend angry and scared. Now I am just very sad. This is an improvement.
16
14
452
People are calling Yudkowsky an intellectual giant, but he has had no influence on serious people like me. He is merely a fanfiction writer who started getting attention a year ago when he pivoted to talking about AI. In this LessWrong post, I will –.
15
6
357
"neural networks could potentially outperform AI-powered chatbots like Bing Chat and ChatGPT" 🤣 .
23
6
329
Super proud of OpenAI for doing this, and happy to say I'll be returning to alignment research by joining this team.
We need new technical breakthroughs to steer and control AI systems much smarter than us. Our new Superalignment team aims to solve this problem within 4 years, and we’re dedicating 20% of the compute we've secured to date towards this problem. Join us!
10
10
259
Maybe this was all a ploy to get OpenAI employees more followers.
17
2
243
100%. Hard to say something of substance nowadays that is not in some way a response to Yudkowsky. I disagree with the guy on many things but IMO this is the mark of a Kant-level philosophical giant.
it really pains me to say this but yudkowsky is an intellectual titan and arguably one of the most important men alive and his arguments are unassailable by midwittery like this.
24
4
204
@aidan_mclau In a single release we managed to steal google’s thunder and get meta’s chief scientist to waste his time nitpicking videos. Competition is too easy.
8
13
203
One of the most impressive things about the Gemini launch is that Google apparently managed to fool The Information.
5
3
188
Go-explore source code is now available! Paper: Blog: @Joost_Huizinga @joelbot3000 @jeffclune @kenneth0stanley
2
62
165
Maybe sometimes they just don't phrase it that way. For the record, if I'm ever on my deathbed and say "I wish I'd made it to AGI", that means "I wish I'd worked more".
10
6
156
Teslas might already be obsolete. A new technology called "car" can go even faster.
0
0
149
Say what you will about SSI but Palo Alto is the right move. Moving Silicon Valley to SF was a mistake.
10
1
126
Go-Explore paper out. New high/average scores: 18 million/650k on Montezuma's Revenge (44k w/o domain knowledge) & 100k+/~60k on Pitfall, all tested w/ sticky actions! Huge thanks to my co-authors @Joost_Huizinga @joelbot3000 @jeffclune @kenneth0stanley.
2
35
124
Surround yourself with people who are even more autistic than you are.
4
7
116
Soon I’ll be able to take my glasses off again!.
This prediction is starting to come true much faster than it ought to have
5
6
113
Ah yes, capitalism, the economic system famously based on not paying for labor.
I genuinely love when academics try to describe some uniquely messed up part of our jobs and accidentally describe a structural feature of labor under capitalism. Every journey begins with a single step!
22
0
113
I think maybe you misunderstand AI.
People that think like this misunderstand how progress in mathematics is made. Mathematics isn’t like chess where a limited set of rules and figures dictates every game. It writes the rules while inventing ever new figures. For that AI is as useful as an advanced calculator.
7
4
114
@IdemErebus Works well because Databricks is a weird outlier in this list, just like Netflix was in FAANG.
4
2
112
Pretty sus that you claim to care about AI safety and yet aren't spending all day shitposting on Twitter.
4
7
102
Well, this has certainly been a day. Wonder when my anxiety will settle enough to go to sleep.
8
5
102
Go-Explore now solves all unsolved* and hard-exploration Atari games, tackles a hard-exploration robotics task, trains in stochastic envs with a goal-conditioned policy and learns its own exploration policy! New paper, “First return then explore” at
2
15
97
Super pumped about our work on weak-to-strong generalization. I am a huge believer that strong empirical research is what we need to align AGI. Also extremely proud that we are releasing $10 million in alignment research grants!.
7
7
98
“First return, then explore”, has been published in Nature! Working on Go-Explore with my co-first author @Joost_Huizinga as well as @joelbot3000, @kenneth0stanley and @jeffclune has been an amazing adventure and I am thrilled that we got to this point!
3
21
90
@ESYudkowsky @OpenAI You are, the alignment research scientist role page says "The annual salary range for this role is $245,000 – $450,000.", the generic research scientist role page says "The annual salary range for this role is $200,000 – $370,000."
4
1
88
One day people will say SBF looks like me rather than the other way around.
1
1
87
This team just can't stop aligning!.
We are systemizing our safety thinking with our Preparedness Framework, a living document (currently in beta) which details the technical and operational investments we are adopting to guide the safety of our frontier model development.
1
4
68
OpenAI employees during training: “it’s very important that you remind the user that you’re a large language model and therefore can’t have your own opinions”. OpenAI employees in custom instructions: “stop telling me you’re a large language model, it’s super annoying”.
0
2
68
Lex Fridman: “if you’ve ever worked in AI, you’re worse than Putin”
🚨 Lex Fridman responds to criticism for interviewing Tucker Carlson . "I will talk to everyone. I want to understand people and different views. I'll talk to Trump, Putin, and Obama"
14
0
65
Too bad for you if you grew up religious: Yann can never take you seriously. Oh well, better luck next life!.
As I have pointed out before, AI doomerism is a kind of apocalyptic cult. Why would its most vocal advocates come from ultra-religious families (that they broke away from because of science)?.
3
4
66
To be fair, until "Wait, actually, yes", the model could have produced a token that would have taken it forever outside the set of correct answers: end of message. Still, the possibility of the model correcting itself is one reason why I don't find his argument very convincing.
2
0
63
How do we get AI to behave ethically when it isn’t clear which moral theory is correct? @joelbot3000 and I explore this question in “Reinforcement Learning Under Moral Uncertainty”
1
14
60
Most people *living* today will see it 🤞.
Kinda crazy a quarter of 21st century is almost over. People born today will see the 22nd.
3
0
58
Surprised people are calling it OpenAI 2.0 when it's really 4.1 or something.
8
1
56
Finally! This is huge!.
We’ve acquired the MuJoCo physics simulator ( and are making it free for all, to support research everywhere. MuJoCo is a fast, powerful, easy-to-use, and soon to be open-source simulation tool, designed for robotics research:
0
2
54
Seems obvious but worth saying: I doubt google intentionally designed gemini to behave in the way we've seen. We can still blame them for being too woke though: a lot of googlers must have known about the issue and decided that raising it would be a poor career move.
12
1
55
@MatthewJBar Maybe he just doesn't make enough as a *checks notes* senior staff software engineer at Google to afford the $20 per month?.
1
0
47
Marxists are probably the most toxic group I've ever angered on Twitter. It's almost as if they don't believe in the basic humanity of their political opponents. I wonder if there's any historical evidence of that. 🤔.
13
1
48
I side with the losing option on this one. I usually disagree with Gary Marcus on AI, but he testified about it in front of the senate, keeps being quoted in news articles, and has written a popular book on AI. By contrast, 50% of the TIME list are people I've never heard of.
Without looking, is Gary Marcus one of TIME’s “100 most influential people in AI”? Should he be?.
7
1
43
So you work at xAI? And you like working on a mac? I gotcha
2
5
47
Safety evals are one of the coolest and most important parts of big new model releases, glad we're sharing them publicly for GPT-4o.
2
1
45
Meta is spending so much on H100s they can't get their CEO a proper camera and microphone.
5
0
44
💯 in spite of my trash talking I actually think it’s a really cool release and I’m surprised people aren’t more excited.
not enough hype around gemini 1.5 pro! the year of long context has just begun and it's going to be amazing. onwards towards intelligence too cheap to meter and too wide to measure.
4
1
41
@ylecun Yann are you hearing yourself? Not asking you to agree that they should hire Christiano, but he’s an accomplished researcher and the inventor of RLHF (used in Llama 2), not some virus to have an immune reaction against. It is beyond me that you would talk like this about a fellow.
1
1
36
@dissproportion Last time I saw a post like this I was like "this strikes me as true" and two guys showed up to tell me I must be bad at sex and started congratulating each other on how many women they'd converted lol. Not easy to have serious discussions about this stuff.
0
0
35
I went on a mental health Twitter break starting March and today was my first serious re-login. Now everything on here feels either stupid or needlessly mean-spirited to me. Take a break, your brain will thank you.
4
1
34
@daniel_271828 Yeah I tend to agree. Pretty sad because publishing in Nature was a big career achievement for me. I like to think they only started going downhill strictly after that. .
1
0
32
Things men think about every 7 seconds:.1. Sex.2. Roman empire.3. Favorite economic graph.4. ?.5. ?.6. ?.7. ?. Only 4 seconds to go and we will have a complete theory of man-thought.
Apparently, all men have a favorite economic graph. Spill it, gents.
4
1
34
@prerationalist Because irrational+irrational=irrational. Seems pretty intuitive to me. Incorrect, but intuitive.
4
0
33
When people move abroad, they often move up in social class. Remember this next time someone tells you how much better life is in <country they just moved to>.
2
2
33
Go-Explore scores over 2 million on Montezuma's Revenge and is the first algorithm to ever score on Pitfall! Proud of the project I have been working on with @Joost_Huizinga, @joelbot3000, @kenneth0stanley and @jeffclune.Blog:
0
5
31
If you ever feel late to the party just remember: if it's a good party, soon a bunch of people are going to think you were early.
2
2
32
Fascinating. You can easily see differences like this influencing people's intuitions about AI, e.g. Yann's more skeptical take on LLMs. As someone who does think in words, the question for me is: what intuitions does Yann have that I don't?.
@nanulled @hamandcheese If the question is "do I think in words", the answer is no.
8
2
31
🫡.
We have reached an agreement in principle for Sam Altman to return to OpenAI as CEO with a new initial board of Bret Taylor (Chair), Larry Summers, and Adam D'Angelo. We are collaborating to figure out the details. Thank you so much for your patience through this.
0
1
33
Two things can be true at once: 1) humans will most likely figure out how to align AGI 2) being one of the humans who helped prevent the end of the world is fucking badass.
4
1
32
Your periodic reminder that Geoffrey Hinton, Yann LeCun, and Yoshua Bengio managed to get a Turing award without publishing any scientific papers*. * Except for the ones that they did publish, but I didn't like those so I decided not to count them.
Your periodic reminder that DeepMind, OpenAI, and Anthropic, the labs that believe they might play with your life within a few years, haven't bothered to publish a detailed report to explain what's their alignment plan, i.e. the core plan meant to avoid extinction from AGI.
1
0
31
@daniel_eth I've been through the "I should probably unfollow this crypto guy" -> "Hmm, he's making a good non-crypto point right now, let's give him a break" -> "Wait, is it just his last name?" thought process on multiple occasions.
2
0
29
@nullpointered @moultano $2,500 is an absurdly high amount of money to spend on a ring regardless of how much you make.
1
0
27
This seems pretty cool and why long context is awesome.
Gemini 1.5 Pro can perform highly-sophisticated understanding & reasoning tasks for different modalities, including video. 📹. When given a 44-minute silent Buster Keaton film, it can analyze various plot points, and even reason about small details that could easily be missed. ↓
1
1
28
@tszzl @AISafetyMemes Viewing the letter as a smart financial move is pure hindsight bias IMO. At the time it felt like a huge financial risk, I don't understand how that's not obvious.
3
1
28
Why do you think everyone (including me) is more excited about the Rabbit R1 than the Humane Pin?.
24
2
29
Really cool stuff! Looks like there's still a lot of exciting work to be done in video game RL.
Introducing SIMA: the first generalist AI agent to follow natural-language instructions in a broad range of 3D virtual environments and video games. 🕹️. It can complete tasks similar to a human, and outperforms an agent trained in just one setting. 🧵
2
0
29
@austinc3301 @StatsLime Also I don't see where in the output it recognizes that it's an AI or in a simulation. It just says it's probably being tested.
1
0
29
This! To many people "religion" has come to believe "strong radical beliefs that are false", so they can't view their own strong radical beliefs as religious because they think they are true. Let's avoid accusing each other of cultism and just engage with each other's arguments.
I was able to voluntarily rewrite my belief system that I inherited from my low socioeconomic status, anti-gay, and highly religious upbringing. I don’t know why Yann’s attacking me for this and resorting to the genetic fallacy+ad hominem. Regardless, Yann thinks AIs "will
3
1
28
"Is your experiment working?". "Well, when you say 'working' — is the loss going down? No. Am I going to keep hogging the GPUs? Yes.".
Reporter: "Are the airstrikes in Yemen working?". Biden: "Well, when you say 'working' — are they stopping the Houthis? No. Are they gonna continue? Yes.".
2
1
26
Stop raging against the dying of the light and start doing something about it.
6
3
26