Professor, Santa Fe Institute. Mostly posting on bsky.app (at-melaniemitchell). More thoughts at aiguide.substack.com.
Santa Fe, NM
- Tweets 6,410
- Following 670
- Followers 50,752
- Likes 5,043
Senator, I'm an AI researcher. Your description of ChatGPT is dangerously misinformed. Every sentence is incorrect. I hope you will learn more about how this system actually works, how it was trained, and what it's limitations are.
31
122
2,629
96,162
New paper by me: "Why AI is Harder Than We Think". arxiv.org/pdf/2104.12871.pdf
Feedback is welcome!
79
424
1,898
I didn't sign "the letter".
Current AI poses lots of risks, but describing these systems as "ever more powerful digital minds" that no one can control is likely to make the problem even worse.
What's needed: more transparency and better public discourse.
50
282
1,697
254,597
Current chatbots can pass the Turing Test, right?
A lot of people have claimed this, but Cameron Jones (@camrobjones) and Benjamin Bergen of UCSD actually tested the claim!
(Spoiler: The answer is "no, they don't pass.")
arxiv.org/abs/2310.20216
73
410
1,514
411,855
Very impressive---indeed, awe-inspiring---AI demos this last week, e.g., from OpenAI (image generation) and Google (text generation).
These demos seem to convince many people that current AI is getting closer and closer to human-level intelligence. 🧵
(1/8)
29
290
1,338
New paper from my group:
"Comparing Humans, GPT-4, and GPT-4V On Abstraction and Reasoning Tasks". 🧵 (1/9)
arxiv.org/abs/2311.09247
23
272
1,220
328,885
Rather than asking AI researchers how soon machines will become "smarter than people", perhaps we should be asking cognitive scientists, who actually know something about human intelligence?
105
198
1,135
311,092
How does modern AI work, what is its history, and how does it compare with human intelligence? My book gives an in-depth, accessible guide for lay people & experts alike.
Coming in 2023: new chapters on transformers, generative AI, and AI "alignment".
melaniemitchell.me/aibook/
41
168
1,095
239,226
Wrenching news: Dan Dennett has died.
He's been a great friend and incredible inspiration for me throughout my career. I will miss him enormously.
dailynous.com/2024/04/19/dan…
34
172
1,050
108,470
Cf. Geoffrey Hinton, 2016: "“People should stop training radiologists now. It’s just completely obvious that within five years deep learning is going to do better than radiologists."
57
140
916
217,186
So cute. But wouldn't the opposite be more useful -- that is, to teach neural networks about how *babies* learn? 🙂
15
114
917
.@Google you should take this "AI Overview" feature offline right now. It is thoroughly dangerous.
39
134
807
121,406
Very interesting and provocative article (not sure I agree with conclusions, but they are worth considering seriously).
cacm.acm.org/magazines/2023/…
42
175
878
274,292
Many have said this, but needs to be repeated:
AI surpassing humans on a benchmark is not the same (at all!) as AI surpassing humans on a general ability.
E.g., just because a benchmark has "language understanding" in its name doesn't mean it tests general lang. understanding.
incredible AI progress in one eye-full. this is what exponential looks like.
speech, image, reading, language understanding, grade school math, codegen - all nearing or exceeded human performance
40
170
729
196,485
Interesting take by @AlisonGopnik : Large language models are "cultural technologies" -- they are not themselves intelligent agents, but, like Wikipedia and libraries, they are technologies that allow humans to access the intelligence of other humans.
piped.video/watch?v=k7rPtFLH…
14
117
680
Personal news: I'm excited to announce that, starting in February, I will be in residence at the Santa Fe Institute for one year, as the inaugural Davis Professor of Complexity. I'm thrilled for this opportunity to work more closely with all the brilliant people at @sfiscience.
27
36
669
New paper: "Large Language Models & Emergence: A Complex Systems Perspective" (D. Krakauer, J. Krakauer, M. Mitchell).
We look at claims of "emergent capabilities" & "emergent intelligence" in LLMs from perspective of what emergence means in complexity science.
⬇️
34
142
719
71,079
One of the most interesting papers I've read on GPT-4 was the "Embers of Autoregression" paper by @RTomMcCoy et al.
Now the same group has done similar studies on GPT o1. 👀
arxiv.org/abs/2410.01792
11
117
673
69,670
In LLMs, the need for "prompt engineering" is a sign of *lack* of robust language understanding.
It will be interesting to see if scaling LLMs alone will reduce the need for prompt engineering.
55
84
653
327,514
New column by me in Science:
"Debates on the nature of artificial general intelligence"
science.org/doi/10.1126/scie…
28
176
625
173,260
This headline. 🙄
One AI program spoke in a foreign language it was never trained to know. This mysterious behavior, called emergent properties, has been happening – where AI unexpectedly teaches itself a new skill. cbsn.ws/3mDTqDL
45
55
601
943,052
How do we know how smart AI systems are? | Science science.org/doi/10.1126/scie…
33
193
614
357,694
Serious question: Is "self-supervised learning" the same thing as "unsupervised learning"? If not, what's the difference? If so, why the new term?
venturebeat.com/2020/05/02/y…
43
157
613
PSA:
Please understand that generative AI systems are not deterministic. You may or may not get this result for a given try.
Also, it's likely Google will fix this particular example soon, so it won't be reproducible. That doesn't mean the overall system is reliable.
14
42
554
79,386
Amidst all the recent Twitter talk on the role of symbols vs. deep learning in AI, I came across a very interesting article by Allen Newell, published in 1982, entitled "Intellectual Issues in the History of Artificial Intelligence".
apps.dtic.mil/sti/pdfs/ADA12…
🧵 (1/9)
10
145
587
Has anyone tried feeding the transcript of a NotebookLM-generated podcast back to NotebookLM to generate a new podcast? And then iterate? Do we get a fixed-point, periodic, or chaotic attractor?
29
42
584
95,857
For Science Magazine, I wrote about "The Metaphors of Artificial Intelligence".
The way you conceptualize AI systems affects how you interact with them, do science on them, and create policy and apply laws to them.
Hope you will check it out!
science.org/doi/full/10.1126…
11
173
570
61,573
New: My adventures using GPT-3 to make Copycat analogies. I did some systematic experiments with no cherry picking. I hope you enjoy this!
"Can GPT-3 Make Analogies?"
medium.com/@melaniemitchell.…
20
137
557
One more point about the Yoshua Bengio blog post.
He says: "Biological evolution has given rise to gradually more intelligent beings on Earth, simply because smarter entities tend to survive and reproduce more."
This is just empirically false. Or am I missing something?
113
43
519
191,919
A book on Amazon, published in 2023, with the same title as my 2019 book.
The "sample" on Amazon reads like a ChatGPT summary of my book (with some small post-editing).
Is this legal?
Same author has published 2 dozen other books on Amazon in 2023 alone.
😠😡
88
127
504
121,193
I'm reading Yoshua Bengio's new blog post, "How Rogue AIs may Arise".
yoshuabengio.org/2023/05/22/…
Mostly it's the same arguments as in earlier writings by Bostrom, Russell, etc.
Lots to say about all this but there's one issue I want to point out.
🧵 (1/8)
19
99
510
293,659
Apple: "let's literally crush all that is wonderful about human artistic creation."
Meet the new iPad Pro: the thinnest product we’ve ever created, the most advanced display we’ve ever produced, with the incredible power of the M4 chip. Just imagine all the things it’ll be used to create.
42
42
475
47,508
There is so much wrong in this tweet.
Our political leaders need better AI literacy.
ChatGPT taught itself to do advanced chemistry. It wasn't built into the model. Nobody programmed it to learn complicated chemistry. It decided to teach itself, then made its knowledge available to anyone who asked.
Something is coming. We aren't ready.
23
46
460
108,453
Guy at next table in coffee shop: "What are you working on so intensely, Miss?" Me: Writing a book." Him: "What's it about?" Me: "Artificial intelligence." Him: "Did you know that two AIs at Facebook invented their own language that humans couldn't understand?" Me: uh....
26
42
490
Workshop on "Conceptual Understanding of Deep Learning". sites.google.com/view/concep… This looks really interesting, and is open to anyone.
3
150
496
Me talking to CNN about (non) sentient AI 🤖
"Philosophers for centuries have pointed out that each of us, while individually aware of our own consciousness, has no proof that other people even are conscious."
Santa Fe Institute Professor @MelMitchell1 talks a Google engineer's claim that an AI system has become sentient.
36
112
487
I'm honored and thrilled to receive this award from @necsi. The award is named for the brilliant polymath Herbert Simon, whose 1969 book "The Sciences of the Artificial" has been so influential to all of us in complex systems science. I'm looking forward to ICCS 2020!
We are pleased to announce that the Herbert A. Simon Award will be presented to Melanie Mitchell for her prolific contributions to complex systems science and artificial intelligence. She will receive the Award at ICCS 2020: necsi.edu/iccs-2020
46
45
480
In a dialogue in Gödel, Escher, Bach (1979), the Crab describes "a new...very flexible kind of computer": "they are called 'smart-stupids', since they are so flexible, and have the potential to be either smart or stupid, depending on how skillfully they are instructed." (1/2)
'It is at once both smarter and dumber than any person you've ever met'
Describes most big AI successes!
9
78
461
133,179
This discourse gets dumber and dumber. Soon all news will be journalists asking LLMs what they think about stories by other journalists.
Journalists: please stop anthropomorphizing these systems!
washingtonpost.com/technolog…
23
83
463
87,952
Our leaders are using the best and most rigorous science about AI to create regulations, right?
25
56
472
104,363
Article in @QuantaMagazine on two of the most brilliant thinkers I know: Bruno Olshausen and Pentti Kanerva:
quantamagazine.org/a-new-app…
15
82
467
103,932
The paperback edition of my latest book is now out in the US and UK!
This book is meant to be an accessible, entertaining, and even-handed account of modern AI, how it works, and how far it has to go to capture the complexity of human intelligence.
melaniemitchell.me/aibook/
21
73
424
On the topic of AI training on copyrighted data, many people have echoed the argument made by Andrew Ng below. But it would be interesting to think about what copyright law would be like if humans had the ability to memorize entire books and recite them when prompted to do so.
78
54
433
137,210
"Herbert Simon said we should call it 'complex information processing.' What would the world be like if it was called that instead?"
wsj.com/articles/why-artific…
27
64
420
In my latest column for Science magazine, I discuss recent AI "reasoning" models -- how it works, to what extent it captures "genuine" reasoning processes, and what's needed to answer such questions.
science.org/doi/10.1126/scie…
14
125
449
55,862
Hats off to whoever made this:
goody2.ai/chat
26
106
418
158,461
Sticking my neck out here, but.....
I don't understand why people think machines could possibly learn common sense from language alone.
Still more trouble in GPT-3 paradise: the "impressive zero-shot performance of large language models is mostly due to existence of dataset bias in ... benchmarks."
– @xiang_lorraine et al, new Arxiv: "A Systematic Investigation of Commonsense Understanding [in LLMS]"
26
44
380
Until then, I will admire the impressive products of machine learning and big data, but will not mistake them for progress toward general intelligence. (8/8)
10
29
388
I really love this essay by @ShannonVallor:
noemamag.com/the-danger-of-s…
It points out that notions of "intelligence" held by AI researchers are very different from those held by people in cognitive science. I also wrote about this striking difference: science.org/doi/10.1126/scie…
13
120
397
76,651
My take on the recent GPT 3.5 "theory of mind" paper.
It's consistent w/ earlier results, e.g., from @Maxwell_Nye et al., that LLMs can keep track of variables & attributes in simple stories (arxiv.org/abs/2106.00737). Calling this "theory of mind" is vast overinterpretation.
This paper is not receiving enough attention:
GPT 3.5 displays emergent theory of mind
arxiv.org/pdf/2302.02083.pdf
27
64
384
113,747
New paper from me: "Abstraction and Analogy-Making in Artificial Intelligence": arxiv.org/abs/2102.10717
🧵 (1/4)
4
73
408
Never thought I would be a podcast host, but...
my arm was twisted and I'm co-hosting this season's Complexity Podcast from @sfiscience. Our topic is "The Nature of Intelligence" 🧠🤖
We have an amazing lineup of guests over six episodes.
I hope you'll check it out!
Launching on Wednesday, Sept 25. Stay tuned!
Right now, #AI is having a moment — and it’s not the first time grand predictions about the potential of machines are being made. But, what does it really mean to say something like ChatGPT is “intelligent”? What exactly is #intelligence?
The next season of the #Complexity #podcast, 'The Nature of Intelligence', explores this question through conversations with cognitive and neuroscientists, animal cognition researchers, and AI experts in six episodes. Together, we'll investigate the complexities of human intelligence, how it compares to that of other species, and where AI fits in. We'll dive into the relationship between language and thought, examine AI's limitations, and ask: Could machines ever truly be like us?
16
84
396
39,329
Tweeps: an AI / philosophy question: Several people (e.g., Searle, Penrose) have argued that machine intelligence is not just hard but impossible *in principle*.
Has anyone made an (interesting) argument for that position more recently (e.g., in last 5 years)?
103
62
390
Researchers: "AI is the new electricity".
Alexa:
bbc.com/news/technology-5981…
21
84
373
Prediction: Quantum computing practitioners will come to regret using the term "quantum supremacy" because the term is an easily misunderstood hype magnet, similar to the hype magnetism of the term "artificial intelligence".
13
79
371
Today's NYT article by Yuval Harari repeats "GPT-4 Hired Task Rabbit Worker to Solve Captcha" story & concludes
"This incident demonstrated that GPT-4 has the equivalent of a 'theory of mind'."
Time to re-up my fact-check of this oft-repeated story.
⬇️
15
74
377
101,130
RIP Lord Robert May. His work on the logistic map was hugely influential to me in learning about complex systems. en.wikipedia.org/wiki/Robert…
2
90
372
Thread: I'm thrilled to announce that my new book is finally launched! Click below for a description, table of contents, reviews, ordering information etc. /1 melaniemitchell.me/BooksCont…
19
100
379
My opinion piece in the NY Times. nytimes.com/2018/11/05/opini…
22
176
366
One of my new years resolutions is to blog (from time to time) about interesting work in AI. I'm trying out Substack for this. My first post is a perspective on recent paper by Webb et al., "Emergent Analogical Reasoning in Large Language Models".
aiguide.substack.com/p/on-an…
12
57
383
105,818
Yeah, the Senate hearings on AI regulation are going great.
(from washingtonpost.com/technolog…)
27
78
347
91,354
Brilliant essay by my colleague, David Krakauer, on the question "Does GPT-4 Really Understand What We're Saying?"
nautil.us/does-gpt-4-really-…
23
98
356
59,427
I'm rereading this terrific perspective piece on AI (by @arjun_ramani3 and @zhengdongwang):
thegradient.pub/why-transfor…
8
80
371
68,994
My latest column for Science:
"The Turing Test and our shifting conceptions of intelligence".
science.org/doi/10.1126/scie…
13
115
355
61,721
If the phrase "godfather of AI" were to be mysteriously eradicated from the English language, I wouldn't miss it.
19
18
340
48,289
The subtle yet still insane risks of using AI to "enhance" photos.
I'm talking at a conference later this year (on UX+AI).
I just saw an ad for the conference with my photo and was like, wait, that doesn't look right.
Is my bra showing in my profile pic and I've never noticed...? That's weird.
I open my original photo.
No bra showing.
I put the two photos side by side and I'm like WTF...
Someone edited my photo to unbutton my blouse and reveal a made-up hint of a bra or something else underneath. 🤨
Immediately, I email the conference host.
(FYI he is a great, respectable guy with 5 kids at home.)
He is super apologetic and immediately looks into the issue.
He quickly reports back that the woman running their social media used a cropped square image from their website.
She needed it to be more vertical, so she used an AI expand image tool to make the photo taller.
AI invented the bottom part of the image (in which it believed that women's shirts should be unbuttoned further, with some tension around the buttons, and revealing a little hint of something underneath). 🤯
—
FYI the conference organizers were super apologetic and took down all of the content with that photo.
5
71
331
31,514
But it's worth remembering the Bongard problems, created by a Russian computer scientist in the 1960s as a challenge to AI. These problems require a rich and general understanding of basic concepts such as "same" vs. "different". (2/8)
10
47
355
"Can you prove it was raining that day?" "Yes, I have a photo." Amazing (and scary) work: research.nvidia.com/publicat…
6
187
365
AI experts, what do you propose that we do about insane statements like this from US defense officials?
(From the @washingtonpost's article about autonomous "AI" weapons. washingtonpost.com/magazine/… )
44
79
332
Question for AI researchers: What would you say are the most important advances towards more *general* AI that have happened over the last five years? That is, general as opposed to narrow AI. (I'm writing a review of this.) 1/3
61
64
345
These numbers are B.S.
These figures came from: aiimpacts.org/2022-expert-su…
They asked over 4,000 people who happened to publish at two 2021 AI conferences, & only 17% (~700) people responded.
This doesn't represent 50% of AI researchers! & hardly from an unbiased sample.
OK, check this out. It gave me goosebumps and will like ruin tonight's sleep.
piped.video/xoVJKj8lcNQ
28
55
302
106,796
"AI Apocalypse Scorecard". See what your favorite AI researchers say about purported "existential threats" of AI:
spectrum.ieee.org/artificial…
20
89
323
489,797
"Pretraining data for LLMs is already basically all of human knowledge".
Interesting claim, but what about all the knowledge that isn't written down anywhere?
Folks might be a bit too manic-depressive about LLMs with any advance meaning that robot apocalypse is around the corner, and any obstacle means that we've hit the wall.
Concretely, pretraining data for LLMs is already basically all of human knowledge, so not being to generalize beyond that isn't that much of an obstacle. And I think it is understood that some form of "self play" would be needed to go beyond that.
32
32
308
117,870
" it seems probable that most of the grand underlying principles have been firmly established and that further advances are to be sought chiefly in the rigorous application of these principles to all the phenomena which come under our notice." -- Albert Michelson, 1894 on physics
Someone’s opinion article. My opinion: It’s all about scale now! The Game is Over! It’s about making these models bigger, safer, compute efficient, faster at sampling, smarter memory, more modalities, INNOVATIVE DATA, on/offline, … 1/N thenextweb.com/news/deepmind…
18
42
304
If you're in Santa Fe, come to my free public lecture, "The Future of AI" on Wednesday (11/15) at 7:30pm MST at the Lensic Theater!
If you're not in Santa Fe, watch the livestream on SFI's YouTube Channel.
All info is here: santafe.edu/events/future-ar…
7
74
305
42,556
Indeed, computer science has a culture of hardcore all-nighter paper writing (not to mention 24-hour hackathons) that definitely hurts all people who are parents, and especially women. I think the CS focus on conference rather than journal publications is part of the problem.
There’s also the insidious biases, which are the hardest to surmount, eg Professors who foster a culture of killing oneself on paper deadlines, to the point of not sleeping the night before. This favoured single males. I’m guilty of this and I apologise to all my students. 7/
10
49
285
Huge thanks to the @CompSysSoc for this honor!
🏆 SFI Professor @MelMitchell1 receives the 2023 Senior Scientific Award from the @CompSysSoc for her outstanding contributions to #AI, #ComplexityScience, and education.
ow.ly/nWFn50PZc5q
28
21
292
96,466
New paper testing whether "natural language understanding" benchmarks actually test language understanding. Answer: No.
Do SotA natural language understanding models care about word order?
Nope 🙃, 75% to 90% of the time, for BERT-based models, on many GLUE tasks (where they outperformed humans).
"marijuana cause cancer" == "cancer cause marijuana" Ouch...
arxiv.org/abs/2012.15180 1/4
13
71
294
If you're quarantined or otherwise stuck at home, and yearning to learn some complexity science, check out @sfiscience's complexityexplorer.org/ for some excellent tutorials and full-length courses!
1
77
292
Interview with me in Spanish.
Headline: "I'm more afraid of Elon Musk than Skynet".
Fact check: True.
elmundo.es/la-lectura/2024/0…
14
48
274
22,605