Julian Togelius
@togelius
Followers
20K
Following
64K
Media
1K
Statuses
23K
AI and games researcher. Associate professor at NYU; director of @NYUGameLab; co-founder of https://t.co/FnakJLkAXW.
New York City
Joined January 2009
I'm excited to announce that my new popular science book will be out on September 24! It's published by @mitpress in their Essential Knowledge Series, so it's short and aimed at a broad audience, assuming no technical knowledge. The topic is Artificial General Intelligence, or
35
91
558
How can you do great AI research when you don't have access to google-scale compute? By being weird. The big tech companies are obsessed with staying nimble despite being big, and some succeed to some extent. But they can't afford to be as weird as a lone looney professor.
32
315
1K
Today is the official release day for my little book on Artificial General Intelligence, published by MIT Press. It's available on the shelf of well-stocked booksellers, and I wrote it to be accessible to as large audience as possible; it's not really a technical book, even
39
172
1K
Not long ago, breakthroughs in AI research often came from lone academics or small teams using desktop hardware. These days, not so much. Are you anxious about how to stay competitive in AI as an academic?.@yannakakis and I wrote this piece for you:.
23
293
1K
So that paper about how GPT-4 could ace MIT's curriculum turns out to be deeply flawed in multiple ways. A great reminder that preprints are not peer-reviewed, but also that public volunteer review can be excellent (in this case, by a group of undergrads).
16
222
1K
Simple statistical methods are shown to much better than fancy machine learning on a whole bunch of real-world sequence-prediction datasets. The reason: the time series used are tiny by ML standards, and all the ML methods overfit.
The paper with its finding that the worse Stat forecasting method was more accurate than the best of the ML ones has passed the 100,000 mark of views/downloads. None of those who have read/downloaded it has challenged its finding. We are still waiting!
19
385
1K
Six years ago, @yannakakis and I published a textbook on AI and Games. It has been used in universities across the world and become a standard reference for the field. Much has happened since, and it is about time for a second edition. Good news: we just put a draft online!
18
185
895
Sub-tweeting because I don't want to rain on a poor PhD student who should have been advised better, but: that paper about LLMs having a map of the world is perhaps what happens when a famous physicist wants to do AI research without caring to engage with the existing literature.
24
52
864
The constant barrage of awesome-looking AI results can induce panic in AI researchers. How could you and I possibly keep up, let alone compete? I try to remember the following:.* Most AI researchers oversell their results.* Most “breakthroughs” rely on existing well-known methods.
16
202
727
You: cute cat video!.Me: procedurally generating curricula for increasing generality in reinforcement learning.
8
108
699
It's 1985. Worried about the capacity of frontier software, and the unchecked spread of shareware and freeware, Reagan issues an executive order that producing any software larger than 1 megabyte requires a special license and reports to several different government agencies.
24
124
639
Reinforcement learning is a paradigm that will eventually be superseded. We just haven't figured out what the new, more generally useful, paradigm is yet. When we do, there's going to be a revolution. It will be very interesting.
31
95
621
Deep reinforcement learning overfits. Often, a trained network can only play the particular level(s) you trained it on! In our new paper, we show how to train more general networks with procedural level generation, generating progressively harder levels.
9
180
587
I think this post channels what a lot of people in the AI community feels right now. As if the stark hypocrisy wasn't enough, there is now also the blatant gatekeeping attempts. I've started thinking worse of people who choose to still work for them.
25
81
549
My main advice going into AI research is to not do what everyone else is excited about right now because it gets amazing results. The next breakthrough is going to come from somewhere else. So go work on something half-obscure which has a chance of moving us in new direction.
17
62
557
Is Elden Ring an existential risk to humanity? That is the question I consider in my new blog post. Yes, it is a comment on the superintelligence/x-risk debate. Read the full thing here:.Or follow along for a few tweets first, if you prefer.
32
146
543
I've started reading "The Art of Doing Science and Engineering" by Richard Hamming and it's like mainlining intelligence, enthusiasm, and, dare I say it, meaning! Every page makes me think of something new I want to do or write. This is the last paragraph of the introduction.
14
62
524
I'm not a big fan of AI regulation, but if we must have it: Everyone that trains a large model (>1B) must release the weights publicly. That's the kind of regulation I could get behind. Rationale: technology that affects us all should be available to all.
62
65
496
As a student, I would sometimes look at the careers of researchers in my field and be surprised at how they abandoned promising research lines and started working on something less interesting. Now I understand they often simply did the research they could get funded.
8
39
480
The AI community is terrible at giving presentations. There's almost always too much detail, too few images, and too little information. Too boring, simply. When preparing a presentation, you should aim for that everyone in the room understands everything you say, and is inspired.
24
59
491
Alternative take: multi-billion dollar LLMs are only twice as good as Eliza at appearing human-like. (Eliza is a few hundred lines of if-then statements written by one guy in the 1960s as a send-up of a particular school of psychotherapy.).
People judged GPT-4 to be human 54% of the time, compared to 22% for ELIZA and 67% for humans. The implication is that people are at chance in determining that GPT-4 is an AI, even though the study is powerful enough to detect differences from 50% accuracy.
22
113
450
Why is machine learning rarely used for generating game levels? Because game levels are harder to generate than, say, faces. And because there aren't enough levels to train on. HOWEVER! In our new paper, we learn to generate pretty good levels after training on only 5 levels!
6
119
455
Excerpt from AAAI reviewer discussion (by me):."it must be possible to accept papers that do things very differently even if they don't beat benchmarks".This, incidentally, is a hill I'd die on.
11
58
426
Most of my papers contain very little mathematical notation. Is this because, unlike many AI researchers, I'm very uncomfortable with reading and writing mathematical notation? Or is it because, unlike many AI researchers, I know how to write readable English?.
21
31
403
Working with AI research is amazing, because you can start with speculations about the nature of intelligence, creativity, or consciousness, and end up with concrete technology that is useful for people everywhere. Most other fields are either interesting or useful, AI is both.
5
71
420
Large language levels can do. lots of things. But can they generate game levels? Playable game levels, where puzzles are solvable? Two papers announced today address this question. The first is by our team at @NYUGameLab - read on for more:
7
107
417
I'm very happy to announce that I will receive the IEEE CIS Early Career Award 2020. The citation is "For Contributions to Computational Intelligence and Games". @ieeecis.
46
19
396
The Piaggio P.180, beloved by deep learning enthusiasts because it has backprop
9
27
399
Artificial Intelligence and Games, the new textbook by @yannakakis and me, is now available for pre-order. Will be available in print in March. We poured our souls into this and hope you (and your students/colleagues) like it!
18
126
379
Please just IPO already. I find it easier to deal with an openly profit-maximizing company than this sanctimonious bullshit.
We’ve raised new funding to accelerate our mission to ensure that artificial general intelligence benefits all of humanity.
4
23
376
I think regulating AI on a technical level (as opposed to an application level) is a terrible, terrible idea and a threat to our digital freedoms. This is a text I've had half-finished for a while, but recent developments made me finish it up.
13
87
377
We have a new, embarrassingly simple method for generating sprites and levels from text. It's also fast. We call it the five-dollar model. Read the paper:.And see it in acton:
15
73
348
I think the intellectually honest approach to LLMs is to be interested in both the (sometimes astonishing) successes and the (equally astonishing) failures. There are real use cases, and real problems. Ignoring either category is foolish and dishonest. Yet, very many do.
13
43
327
I strongly disagree. We need less metrics-driven hiring and promotion in academia, not more. The reason is Goodhart's law: "When a measure becomes a target, it ceases to be a good measure." The more we incentivize publications and citations, the more this system will be gamed.
I think we should be using analytics a lot more for tenure and hiring in academia. Call it "Moneyball for Professors". #Analytics #BigData #ML #2MA .
8
96
322
"I tried reinforcement learning, but it didn't work". "Try again".
12
43
328
I don't think a six-month ban on developing models "more capable" than GPT-4 (whatever that means) would make much difference in terms of AI risks and benefits. But: the global coordination on research suppression that would be needed to enforce this would be immensely harmful.
16
34
319
hear me out: infinite monkeys, but we give them bananas when they type things we like.
16
21
308
A lone professor with a handful of students and a few computers can never win over @DeepMindAI or @facebookai in a straight competition. But we can afford to try methods that make absolutely no sense, or attack problems that nobody wants to solve as they don't look like problems.
6
37
312
I'm excited to report that I'm writing a short book on Artificial General Intelligence to be published by @mitpress. It will go into their Essential Knowledge Series so I will write for a general audience. This is your chance to influence me: which topics must I include?
105
25
308
So go on, be weird. Out-weird the giants. Even if they're both nimble and powerful, they cannot be as stupid and ridiculous as you. Because how would that look? To managers, investors, board members, the general public? You can afford to completely disregard such entities.
3
49
290
It is sad to see this reaction to the development of amazing new tools that give artists unprecedented abilities.
A new AI image generator appears to be capable of making art that looks 100% human made. As an artist I am extremely concerned.
28
25
291
@Noahpinion This is borderline misleading. They pay a smaller share of healthcare costs as out-of-pocket, but only because the overall healthcare costs are so high. In real terms they pay more money out-of-pocket than almost everywhere else.
11
5
296
I don't think there'll be another AI winter, because search, planning and supervised learning work well and are in use everywhere. There might be an RL winter though, because reinforcement learning doesn't work well and is not likely to work well soon.
8
38
280
It is important to remind yourself that the system has no interest in the truth, because it has no concept of the truth, because it has not concept of the world.
18
21
274
It is happening! @yannakakis and I have started working on the second edition of our textbook on AI and Games. We will try to incorporate feedback from the community as much as possible, so please send your suggestions for changes/additions to gameaibook@gmail.com.
3
15
282
Today, in a private ceremony involving only the immediate family (and officiated by @flantz) on Zoom, @razsaremi and I got married! We are very happy and hope to be able to throw a party or two to celebrate in summer.
71
3
281
What if we see level generation as a game, where each action changes the level in some way? Well, we could use RL to learn to "play" level generation! We introduce Procedural Content Generation via Reinforcement Learning (PCGRL), a new paradigm for PCG.
7
73
271
I fully agree with Andrew Ng here. We're at a critical point where certain people's catastrophic thinking threatens to put in place a censorship and regulation regime with far-reaching consequences for not only free enterprise but our digital freedoms more broadly.
My greatest fear for the future of AI is if overhyped risks (such as human extinction) lets tech lobbyists get enacted stifling regulations that suppress open-source and crush innovation. Read more in our Halloween special issue of the Batch:
9
41
271
My new book, "Playing Smart: On Games, Intelligence, and Artificial Intelligence":.This is a popular science book aimed at non-technical audiences. It's an overview of research on AI and games, and an argument that AI is crucial for games and vice versa.
11
68
265
Writing a grant proposal with someone who is not a computer scientist. Do people really still mail Word files back and forth? Actually, do people still use Word? Can we do something about this?.
39
10
256
Sorry, I try to avoid cheap dunks on people, but the peculiar US tendency to talk shit about universities really annoys me. Okay, you didn't like college and dropped out, only to build a business where you employ tons of highly educated people. Maybe the problem was not college?.
8
10
253
Can deep networks be superstitious? We trained deep networks with the standard A2C deep reinforcement learning algorithm to play "deceptive" games and found a range of behaviors that could only be described as bizarre. and, yes, superstitious.
4
75
248
This is a good post. It's true that top universities get an avalanche of applications from people wanting to do AI/ML PhDs. And published research is a very good indicator of research potential. But there are other universities than the top ones and other conferences than NeurIPS.
“To start a PhD in ML, without insider referral, you need to do work equivalent to half of a PhD. Hence, in Apr 2019, I decided to dedicate all my time until Jan 2020 to publish in either NeurIPS or ICLR. If I fail, I would become a JavaScript programmer.”. — @andreas_madsen ‼️.
5
49
250
Are AI doomers platonists? It seems that most of the doomer arguments rest on the idea that there is such a thing as Intelligence, which has certain properties and quantity. The nominalist position is that the word intelligence is just a convenient way of clustering phenomena.
54
33
234
Standard evolution strategies (invented in the 1970s, can be implemented in 10 lines of code) are comparable in performance to fancy Natural Evolution Strategies, which in turn are comparable in performance to Deep Q-learning, on ALE Atari Games.
"Back to Basics: Benchmarking Canonical Evolution Strategies for Playing Atari," Chrabaszcz et al.:
6
84
246
Some people (still!) think that video games are not important research topic, or at best a niche application for AI. So I wrote this "Apology for Video Games Research":.
8
45
244
Many of the best ideas still come from academia, even though the best results don't.
12
27
241
"Empiricism and the limits of gradient descent".New blog post where I use Karl Popper's take on epistemology to argue that evolutionary algorithms can learn in a way that backpropagation cannot.
11
78
233
Guess who won the AAMAS 2019 best paper award? @giuse_tweets, me and @eXascaleInfolab did, that's who. Or rather, our paper "Playing Atari with Six Neurons" did. We are very happy 😁
16
29
224
Some advice for journalists writing about artificial intelligence.NB: I am not a grumpy old AI researcher.
11
111
220
When you tell people about a new idea, and they respond with "that must already have been done" or even "that's already been done", but you search the literature and it has not in fact been done yet: this is a very good sign. Drop everything else and do it immediately.
7
14
227
I'm happy to report that Greenwich Village is almost back to its usual level of weirdness.
8
11
218
The ease with which some people think other’s jobs (rarely their own) will be replaced by AI systems usually say more about how little these people know about, or respect, the jobs which they say will be replaced.
11
29
217
"As A Large Language Model, I".A short text of unclear character and purpose that I wrote when I couldn't sleep.
15
60
220
I bought my mom a Bosch washing machine recently, and then I started thinking about what it would be like if it was made by Hieronymus Bosch instead of the namesake appliances giant. Then I asked Stable Diffusion to help me think.
7
15
213
To the extent I've done anything useful or worthwhile in my career, it's always been through trying to solve a problem nobody thought of, or trying a method that shouldn't work. Very often the useful/publishable end result was nothing like what I thought I was working towards.
2
37
220
Please allow @razsaremi and me to introduce our son, Dylan Saremi Togelius! He emerged from Raz last week and we celebrated his zeroth birthday yesterday. We are very happy and are rapidly realizing that we have a lot to learn about parenting.
28
2
209
How can you use AI methods to balance a game? In particular, a complex game such as Hearthstone, with hundreds of cards? Through multiobjective evolutionary algorithms. In our new paper, we show how to rebalance the game while making minimal card changes.
6
57
217
People kept saying exactly this for decades. This was the mainstream opinion. My teachers said this, and everyone listened. Technically, they were correct. However, neural networks turned out to be able to do amazing things that we have not yet been able to do with other tech.
🆕🤖 The “neural network” technology that powers current artificial intelligence is extremely expensive, poorly understood, inherently deceptive, and unfixably unreliable. In short: it is bad. Somehow, no one ever says this!
9
5
217
At least from this press release, it seems that the EU AI Act came out less bad than feared. We seem to have avoided any need for licensing or similar for foundation models, and open-source distribution is permitted and seemingly even encouraged.
10
49
208
Extremely unpopular opinion: earning a living through artistry in any domain (paintings, movies, dance, novels, etc) has always been very hard. This is a supply and demand issue: lots of people want to be artists, few want to pay for their art. AI is not changing that much.
15
18
205
Fellow AI researchers looking to stay relevant and produce cutting edge research despite having less-than-massive computing resources, what's your strategy? And/or your best tip?.
47
14
203
New rule: if you post generated images, you must post "failures" as well as "successes". If we keep posting only beautiful and/or meaningful images, we make people think that this tech just magically works. In response to this tweet, please post your most recent generated image.
38
30
192
If you never publish WTF papers - those that make your stiffer peers wonder what you’re smoking - you’re not exporting enough. Also, you’re taking this whole “science” thing far too seriously. Loosen up.
8
22
196
Maybe I should start a series of livestreams where I invite people over to my apartment to get drunk and talk AI? Tentative name: "AI and drinks". As we drink more, we'd gradually decohere and slide from talking about algorithms and architectures to consciousness and Skynet.
32
11
193
If you don't know about quality-diversity algorithms, you should. It's one of the most interesting things happening in AI right now. We're moving beyond the standard concept of optimization into illuminating spaces of solutions. Applications everywhere.
If you are willing to learn more about Quality-Diversity algorithms, we have created a website that gathers papers, tutorials, and implementations of QD algorithms. The progressive increase of papers published on this topic every year is impressive!.
3
41
180
I understand you're annoyed about that rejection from NeurIPS. As an AC, I can confirm that I recommended rejection for a bunch of perfectly adequate papers. They were just not very exciting.
10
30
181
We present a new optimization algorithm for finding not just one good solution, but a large set of good solutions that vary along specified dimensions. The algorithm, Covariance Matrix Adaptation Map-Elites, has potential applications everywhere, from robots to games to art. .
Excited to share a recent collaboration with @amykhoover and @togelius. In this paper we were able to show Quality Diversity (QD) algorithms, such as MAP-Elites, can be greatly improved by leveraging ideas from the modern derivative-free optimizer CMA-ES.
5
48
176
You can always make your paper more readable and accessible. Making your paper more readable and accessible is always good. Obfuscation is never a virtue. Obscure terminology is sometimes a necessary evil, but should be minimized.
5
26
169
One way of thinking about language models is that they role-play. In fact, role-playing is all they know how to do, because all they've been trained to do is output text as if they were someone else. Below: A conversation where GPT-3 tells me it is not sentient nor conscious.
7
51
170
There are those who define as AGI (or ASI) as technology that will "outperform humans at most economically valuable work". Ok, but then this work will simply cease to be so economically valuable, and humans will mostly stop doing it. Humans will instead find new economically.
38
41
178
What can LLMs do in and for games? NPCs, narratives, game mastering, playing, level generation, new game mechanics. player modeling? And what's next? @InDigitalGames and @NYUGameLab proudly present ."Large Language Models and Games: A Survey and Roadmap"
2
52
177
My upcoming book on AGI features (among other things) a long-form version of this argument.
General intelligence, artificial or natural, does not exist. Cats, dogs, humans and all animals have specialized intelligence. They have different collections of skills and an ability to acquire new ones quickly. Much of animal and human intelligence is acquired through.
17
11
168
Using reinforcement learning to train game level generators is a quite new thing. In our new paper, we show how to train controllable level generators, where the generator can be parameterized in runtime. We demonstrate this in games including Sokoban, SimCity, and a Zelda-like.
In "Learning Controllable Content Generators," with Maria Edwards, @Amidos2006, @FilipoGiovanni, and @togelius, we train an RL agent to generate game levels with specific features by feeding the agent targets as input and rewarding it for approaching them.
5
37
162
It's great when you only cite anonymous sources. Then you can make up this fantasy world full of researchers scared of their own neural networks which has no connection to the world which the rest of us AI researchers live in.
5
25
175
Honestly, these NeurIPS submissions are so boring. So, so boring. Also confusing. Once you get through the bad writing and unnecessary math notation, you keep wondering why the authors decided to write this paper. Also, why did I agree to review for this conference again?.
12
17
169
The late Wittgenstein and other "linguistic turn" analytical philosophers thought most philosophical problems are confusions about language that could be explained away by analyzing concepts properly. I think a dose of this is what the AGI/superintelligence discourse needs.
7
24
163
Say you have an AI method and don't have the resources to test it on all possible benchmark problems. How do you select which games to test it on? In our new paper, we use information theory to describe how to do this.
4
36
165
❤️.Many - perhaps most - of my best ideas have come out of thinking in pairs. Basically, long, probing, lively, accepting discussions of ideas. Finding good thinking partners is the best thing you can do for your research (and maybe your life).
Doing good science is 90% finding a science buddy to constantly talk to about the project.
6
33
165
Saying "an AI" is definitely a red flag, but an even bigger red flag these days is arguably talking as if ChatGPT was all of AI.
I can’t justify it by the biggest red flag that someone is bullshitting about AI is the phrase “an AI” when talking about current systems.
15
12
162
I asked an AI who the best AI researcher in the world is, and the response was a (fictive) AI.
4
32
160
GPT-3 often performs like a clever student who hasn't done their reading trying to bullshit their way through an exam. Some well-known facts, some half-truths, and some straight lies, strung together in what first looks like a smooth narrative.
12
34
155