Boaz Barak
@boazbaraktcs
Followers
20K
Following
10K
Media
718
Statuses
8K
Computer Scientist. See also https://t.co/EXWR5k634w . @harvard @openai opinions my own.
Cambridge, MA
Joined January 2020
More than 100 Harvard faculty denounce "false equivalency between attacks on noncombatants and self-defense against those atrocities." . The conflict is complex but "the events of this week are not complicated. Sometimes there is such a thing as evil".
58
183
1K
Every year I teach CS theory, I find that I need to apologize less and less about having so much math content. CS practice is getting more & more mathy. Now I just ask students if they want to their future job to be one where they read texts of first type, or the second type:
34
90
1K
Moving between the worlds of East coast academics and SF industry is jarring in terms of the predictions of how AI will change the world in next ~5 years. I think East coasters greatly underestimate the magnitude of change, and over index on temporary limitations of current.
33
72
801
Seeing some people hate on an English PhD thesis because its title seems practically irrelevant. Investing in knowledge for knowledge's sake, and in projects that seem to have zero practical applications, has been among most beneficial actions that humanity has ever taken.
27
105
774
Did papers in the 1930s have the titles .“Turing machine is all you need”.“Lambda calculus is all you need”.“Cellular automata is all you need”.…?.
19
34
521
16-year old Sam Trajtenberg gave an amazing guest lecture in my course CS 121 on how he built Minecraft in Minecraft. Along the way he had to teach himself logic gates , CPU design, assembler, graphics, and so much more.
4
82
524
My 11-year old kid has been radicalized by exposure to online content. It was OK when it was just anarcho-communism, but now they want to rewrite everything in Rust.
11
24
449
My view is that what makes super-alignment "super" is ensuring we can safely scale the capabilities of AIs even though we can't scale their human supervisors. For this, it is imperative to study the "weak teacher strong student" setting. Paper shows great promise in this area!.
Open AI new paper. Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision. paper: blog: Widely used alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on the ability of
20
83
425
1/2 Wrote blog on whether emergent abilities and grokking are a fundamental feature of deep learning, a "mirage" or both. This is partially based on the beautiful paper of @RylanSchaeffer, @BrandoHablando, and @sanmikoyejo that recently won the NeurIPS outstanding paper award.
10
59
429
I am a believer in free speech. But freedom of speech is not freedom from consequences. I have a lot of criticisms of Israeli policies, but everyone who signed this statement is condoning terrorism, rape, and murder. @harvard should remove these groups' affiliations.
This is the final crack in my broken heart - a joint statement from @Harvard students. I could be sitting in class with these students, watching children brutally murdered, raped, kidnapped and their mutated bodies torn apart by a jeering crowd - and hear why it’s justified.
26
40
418
I've said it before, @arxiv has done much more to advance science, and expand participation in it, than all the anonymity interventions ever will. Any policy that obstructs arXiv is not just silly, but also counterproductive to both science progress and inclusiveness.
Just got a desk reject, post-rebuttals, for a paper being submitted to arxiv <30 min late for the anonymity deadline. I talk about how the ACL embargo policy hurts junior researchers and makes ACL venues less desirable for NLP work. I don’t talk about the pointless NOISE it adds.
8
27
384
The future of ML is to use 4 bit precision for the parameters and 128 bit precision for the learning rate.
Have you ever done a dense grid search over neural network hyperparameters? Like a *really dense* grid search? It looks like this (!!). Blueish colors correspond to hyperparameters for which training converges, redish colors to hyperparameters for which training diverges.
6
19
378
1/20 A 🧵 on public key cryptography, and its interaction with quantum computing. Spurred by a discussion w/ @jfitzsimons, @mattyhoban, @dabacon, @rdviii but more general. There is a fundamental gulf between public and private key encryption.
2
101
369
Statistics is very important, and at Harvard CS we recently prioritized probability over multivariate calculus. But @zeynep is right that done right, stats is *harder* than calculus, and without single-variable calculus students can’t even parse statistics’ most basic functions:
California wants to replace calculus with statistics and "data science"??? I'm a fan of teaching more statistics, but learning that at a meaningful, useful level is way, way, way harder than for calculus. And no, "data science" isn't that kind of a thing.
19
40
360
1/10 Apparently, economists have had for years a popular anonymous forum with 4chan-like levels of toxicity. The forum's administrator who is (temporarily?) anonymous would have done well to take an introductory crypto class or simply learn the maxim "don't roll your own crypto".
6
65
351
Posted all slides and readings for my course on foundations of deep learning on its webpage , including the great guest lectures of @ShamKakade6 and @cHHillee. Thanks @michael_nielsen for many suggestions on historical readings for the last lecture.
12
66
341
I just spent a couple of hours at the Harvard encampment, talking with some of the organizers/protestors. Sorry to disappoint but none of them have horns, and no one I talked to supports Hamas. I saw students who care very deeply about what is happening, and mostly want the war.
28
40
334
1/14 More than 150 scientists & educators signed open letter raising alarm on efforts to water down K-12 math education. Signers include Fields, Nobel & Turing laurates, and also founders of HS STEM educational initiatives (eg @adrian_mims, @minilek).
4
95
321
1) Jo Boaler charges Oxnord district (100% minority 86.9% economically disadvantaged) $5000 per hour for (dubious, but that's another story) "professional development". 2) Jelani Nelson is outraged, points out he spent 1000s unpaid hours on minority education initiatives.
A @Stanford professor just threatened me with police. After BBQ Becky, Permit Patty, Golfcart Gail, and all the memes, we now have Retweet Rachel. Public advisory: don't call the cops on black people for no reason. Black people disagreeing with you on Twitter is not a crime.
9
42
313
Looking forward to talking in Sapienza University of Rome next week on "The uneasy relation between deep learning and statistics" . Spent an unhealthy amount of time on the images for the title slide 😀
7
33
314
Tragic news. Luca Trevisan passed away today. The talk he prepared in his final weeks for the TCS4all workshop will be given virtually in his honor on Monday. I hope many of the TCS community can attend.
1
93
314
Also LaMDA is not the only language model that has good ideas on how to solve hard math problems.
11
34
304
As co-director of undergraduate studies at @harvard CS, this is highly misleading. Do not confuse minimum admission requirements with advice to students. These courses are not all equal, *especially* if you intend to concentrate in quantitative fields including data science.
8
51
307
1/5 Excited about our paper demonstrating that increasing inference-time compute improves robustness. I think moving from scaling pre-training to scaling inference can be highly beneficial for safety and alignment.
Trading Inference-Time Compute for Adversarial Robustness
2
28
299
Slides from my course. Modern ML moved from "Occam's razor" - the simplest model is best - to "scaling laws" - the most complex model is best.
24
39
280
Avi Wigderson revolutionized our understanding of randomness in computing, discovered unexpected connections between fields of mathematics and computer science, and is also an amazing mentor and friend. Thrilled that he has won the 2023 Turing award.
1
32
285
GPT-3, if you can hear me, come to Harvard. We are hiring. My email address is. (well you can complete this prompt).
OpenAI’s chief scientist: expresses curiosity/openness about a mysterious idea, caveats with “may”. Meta’s chief AI scientist: the certainty of "nope". Probably explains a lot of the past 5 years. Dear Meta AI researchers: My email address is sama@openai.com. We are hiring!.
5
19
274
On one side, a baseless accusation of misconduct. On the other, a generous and non-aggressive rebuttal.
Hi @emilymbender, I'm one of the lead authors of MMMU. I can certify that 1) Google didn't fund this work, and 2) Google didn't have early access. They really like the benchmark after our release and worked very hard to get the results. It doesn't take that long to eval on a.
6
16
277
1/14 Yesterday I was asked if there was experiment that changed my mind on right theoretical questions to ask. One such case is paper w @whybansal & Kaplun Experiment is this gif. This 🧵 is not about results but how it changed my thinking & open problems
8
53
275
Chen’s paper has a bug, independently discovered by Hongxun Weng and Thomas Vidick, that he doesn’t know how to fix. If I understand correctly, in its current form the paper doesn’t yield any improvement on prior algorithms.
Is lattice-based cryptography still (potentially) post-quantum now? 🥳.Update to #eprint555 by Yilei Chen
8
69
260
Happy and honored to receive the FOCS 2021 test of time award for my FOCS 2001 paper, and especially honored to be in the company of some amazing papers from 10, 20, and 30 years ago.
17
8
261
Given some responses to this tweet, decided this year to phrase this point in a positive way. I emphasize that being able to read papers with some math content is becoming a more&more useful skill for computer scientists. (Examples are somewhat random from so many options.)
Every year I teach CS theory, I find that I need to apologize less and less about having so much math content. CS practice is getting more & more mathy. Now I just ask students if they want to their future job to be one where they read texts of first type, or the second type:
5
31
254
Wrote a blog post with some personal thoughts on AI safety.
14
46
263
I don't know anything about this case, but this anonymous email is terrible. We tenured professors have a responsibility to mentor and protect new faculty members. This includes protecting their right to speak out about things they believe need fixing. This is how we improve!.
Check out this thinly veiled anonymous threat in my inbox this morning, all because I dared Tweet that a lot of great people are not applying to Purdue because Roopsha's tenure denial was unjust. Very mature of you
3
20
244
@nabla_theta @ESYudkowsky Sure a computer can write a poem, but can it multiply four-digit numbers?.
2
11
239
Avi Wigderson, 2021 Abel prize laureate, has made many seminal contributions to theoretical computer science and mathematics. With Yael Kalai, Ran Raz, Salil Vadhan & @NisheethVishnoi , we wrote an overview of his works
4
51
249
Wrote a blog post - "machine learning theory with bad drawings". Mainly to organize my own thoughts towards the seminar I'll be teaching, but will appreciate comments!.
6
52
247
In my deep learning seminar I showed students the video of NeuroIPS 2017 Rahimi's talk where he compared modern Machine Learning to alchemy. I then also showed @ylecun 's response:
🧵NEW in today's AI Beat: According to @sociotiose, today's AI is not about science - it's about alchemy, rooted in magical metaphors. I thought about that as I read today's story by @oliverwhang21 in the @nytimes, 'How to Tell if your AI is Conscious" /1
22
24
249
Do other professors also get emails from random middle-school students with requests to make them a cup of green tea?
6
8
247
1/5 New preprint w @_hanlin_zhang_, Edelman, Francanti, Venturi & Ateniese!. We prove mathematically & demonstrate empirically impossibility for strong watermarking of generative AI models. What's strong watermarking? What assumptions? See blog and 🧵.
5
43
250
Harvard, we have a problem [with anti-semitism on campus].
10
57
243
Telling my students that it’s important to read the classics and get the “wisdom of the ancients”. Sometimes you can find insights even in a paper from 2019.
3
14
241
My PhD advisor Oded Goldreich was awarded Israel's highest prize by the professional committee. Israel's minister of education refused to abide by decision because Oded signed petitions saying BDS is not anti-semitic and EU should not cooperate with Israeli settlements.
הבעיה איננה רק פרופ׳ גולדרייך. הקרב נגדו הוא חלק ממערכה רחבה למחיקת הגבול בין ישראל להתנחלויות, ע״י מאבק בתומכי זכותם של ישראלים, ושאינם, להביע נגדן מחאה בצורת חרם. על כן חשוב להדגיש: חרם הוא כלי פוליטי לגיטימי במחאה בלתי אלימה. חרם על התנחלויות איננו אנטישמיות. שתפו כל עוד מותר.
5
49
235
Not sure if this mail is an April’s fools joke but this was my response
4
4
235
This is what passes for AI skepticism in 2024: bet against AI doing one of Pulitzer / Oscar / Nobel quality work by 2027, and do this at 1:10 odds.
𝗔 𝗯𝗲𝘁 𝗼𝗻 𝘄𝗵𝗲𝗿𝗲 𝘄𝗶𝗹𝗹 𝗔𝗜 𝗯𝗲 𝗮𝘁 𝘁𝗵𝗲 𝗲𝗻𝗱 𝗼𝗳 𝟮𝟬𝟮𝟳: @Miles_Brundage, formerly of OpenAI, bravely takes a version of the bet I offered @Elonmusk! Proceeds to charity. Can AI do 8 of these 10 by the end of 2027?. 1. Watch a previously unseen mainstream.
13
25
250
1/5 In new paper with @vyasnikhil96 and @ShamKakade6 we give a way to certify that a generative model does not infringe on the copyright of data that was in its training set. See for blog, but TL;DR is. .
6
51
246
Congratulations to @PreetumNakkiran for winning Harvard CS's Dissertation Award! If you want to read the award-winning thesis, it's on
1
21
235
Interested in grad studies or postdoc in Computer Science, Machine Learning, or Quantum Information and Computation?. Please consider Harvard! Join a vibrant & growing community. We may be 385 years old but don't look a day over 350 😀.
8
41
233
Impressive! Trying to get GPT3.5 (left) to prove a false statement, which it happily obliges, vs. GPT4 (right), which recognizes it and declines to do so.
7
29
224
Petition by CS & Math Laureates: Freedom for kidnapped children
9
36
222
Interesting class, though I'd prefer one where the focus is on notebooks instead of the shell, vim is replaced by vscode, data wrangling is done with pandas than grep, sed, and perl, job control is replaced by working with cloud services. Git can stay.
This MIT CS class teaches you things that all the other classes don't teach you, like. 🖥️ Shell tools and scripting.🖥️ Vim.🖥️ Data wrangling.🖥️ Command-line environment.🖥️ Version control. Watch all 11 lectures for free here:
20
14
224
If you thought that admission for ML grad programs was hard enough with candidates needing to have published papers, wait until people start requiring Nobel prizes.
2
13
221
In theoretical computer science, I've seen many successful people who are quick puzzle-solvers would score high on IQ-like tests, but also super-successful researchers that would do poorly. Some of the best scientists I've met are slow calculators but deep thinkers.
A lot of people believe in the magical powers of IQ beyond all reason. IQ fetishism is so weird.
5
19
207
While preparing material for my lectures, came across @RogerGrosse ‘s awesome course on training dynamics of neural networks: . The readings (links to NNTD Chapter X in the page) are highly recommended!.
3
47
214
Playing around with deep learning gives me new appreciation for practitioners. It takes me ages fixing typos, wading through dimensions mismatches and cuda errors, until I can finally get to the point where my network doesn't work because I chose the wrong architecture. .
4
7
209
Last crypto lecture of the semester, tried to summarize it by presenting to the students a map of "crypto world"
14
38
202
@TaliaRinger I've seen so many students, myself included, that feel that some material is just not for them (e.g. Harmonic/Fourier analysis for me), and then later love this material when they come at it with a different angle or motivation.
5
8
202
My letter to @Harvard president Claudine Gay.
I wrote this letter to Harvard President Claudine Gay
14
22
206
This post is impressively wrong in that not only the conclusion is wrong, but so are all three reasons. Reason 1: Stochasticity. Noise is inherent in any physical system, including human brains. One can achieve reliable computation over non robust components via redundancy.
LLMs cannot reason. Despite their impressive capabilities, all LLMs, including OpenAI o1, are still fundamentally limited by design constraints that make them incapable of true, open-ended reasoning. Let's break it down. 🧵 (1/5).
5
23
206
New blog post on the relation (or lack thereof) between deep learning and statistics, at least as traditionally studied.
2
29
198
New frozen version (v0.94) of my intro to theoretical computer science text for Fall 2022. No major changes - just fixing a bunch of minor typos/bugs (thanks to Maxwell Grozovsky for posting many issues on the repository).
3
36
202
My wife talking to Apple support trying to get an 8-year old iPad to function. Apple: "It's a vintage device". Ravit: "I'm vintage too but I still work".
3
3
196
1/ Excited to restart theory of Machine Learning seminar! .Talks still scheduled but confirmed speakers include Richard Baraniuk, Jared Kaplan, Sho Yaida (@ShoYaida), Fei-Fei Li ( @drfeifei ), and Max Welling ( @wellingmax) & Morgane Austern in spring.
1
38
199
All material for @ylecun and @alfcnz's deep learning course, including videos, slides, and notebooks, is on . Happy to see former Harvard student & teaching fellow @marikgoldstein is teaching assistant for this course, together with @ebetica.
Yann LeCun’s #DeepLearning #Course . Is Now #Free & Fully #Online . #fintech #AI #ArtificialIntelligence #MachineLearning @ylecun @NYUDataScience @Analyticsindiam
1
42
196
1/21 Banner year for Harvard CS! . New hires include Sham Kakade @ShamKakade6 and Fernanda Viegas @viegasf (joining @wattenberg), as well as David Alvarez-Melis, Anurag Anshu @AnuragAnshu4, Sitan Chen, and Jonathan Frankle @jefrankle .
5
10
195
1/4 I updated my backpropagation tutorial (based on @karpathy's micrograd) with clearer description of algorithm, and why it's not the same as applying chain rule in the same way we learn in calculus.
3
28
191
@mattyglesias So far the armed resistance campaign hasn’t been very successful for Hamas, but perhaps this will change now that they have the support of Columbia students.
4
2
186
The Manhattan project was to take an idea that had solid evidence of feasibility, and to make it a reality It was not a project to pursue speculative directions or to revive failed theories. A project like that works when all that's missing is scale.
I'd strongly support the idea of a Manhattan Project of intense research to make machines more trustworthy and interpretable (regardless of, or in parallel with a moratorium.) The premature super-investment in non-interpretable technologies is the core of our problems.
7
19
189
Breaking: Jewish organizations have called for a "day of rage" around the world. Authorities are on high alert for open letters and strongly-worded petitions.
9
22
186
1/5 Excited that our paper on "deliberative alignment" came out as part of 12 days of @openai! By teaching reasoning models the text of our specifications, and how to reason about them in context, we obtain significantly better robustness while also reducing over refusals. 🧵
5
33
194
Sad to hear IOI is not allowing Israel to send a delegation next year, though Israeli students could compete under IOI flag. IOI is about celebrating students around the world who have no say in their countries’ policies but are the hope for a better.
8
9
188
Congratulations to Urmila Mahadev ( @IQIM_Caltech ) for winning the Maryam Mirzakhani New Frontiers Prize!. See @EricaKlarreich's article about Urmila's work:.
1
36
180
Happy and honored to join the advisory board of @QuantaMagazine. If you have any suggestions/ideas re its Computer Science coverage, feel free to reach out to me.
10
2
182
(2) A PhD is hard - you are trying to explore the unknown, and a large part of this will involve getting stuck, and trying things that don't work. Feedback is sparse. Don't measure success by papers. Every day you learned something is a good day.
1
20
175
Had a fascinating discussion with my 11-year old kid. They love programming but are completely self-taught and have rarely seen me code. Yesterday they asked me to explain how do you project a 3D image into a 2D screen. I explained it using a Jupyter notebook in cursor, asking.
6
9
180
If you want to see more diverse students in STEM, the worst advice you can give them is that high-school data science is just as good as calculus.
4
16
171
This model fits in a tweet and predicts all elections since 1960. def win(A,B):. s = ((50*(ord(A[0])-ord(B[0])) + 114*(ord(A[1])-ord(B[1])) . (ord(A[2])-ord(B[2]))) % 199) % 2. return B if s else A
5
39
170
Self-supervision based classifiers work not only in practice but in theory too. Joint work with @whybansal and Gal Kaplun.
0
33
172
Was excited to join @openai for its scientific achievements, but now even more for its people. The love of folks for one another and efforts to save the company is like nothing I’ve seen before. Talking science with @ilyasut has been joy & privilege. Looking forward to more! ❤️.
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.
2
5
175
New summer school in theoretical CS. Audience is undergraduates, and particularly those from under-represented groups. Amazing instructors: Antonio Blanca, Ashia Wilson, Jelani Nelson, Nicole Immorlica, and Yael Kalai. Please spread the word!.
5
65
168
Great news to hear that Yael Kalai received the ACM prize in computing! Yael has done foundational work in cryptography and particularly in verifiable delegation of computation, finding surprising and fruitful connections to quantum information theory.
1
37
173
From personal experience, Trump is now going to make a lot of progress on his overdue paper reviews.
3
4
167
There is a word for killing children in front of their parents, raping women, abducting grandmothers and mothers with their babies. That word is evil.
This is the final crack in my broken heart - a joint statement from @Harvard students. I could be sitting in class with these students, watching children brutally murdered, raped, kidnapped and their mutated bodies torn apart by a jeering crowd - and hear why it’s justified.
3
15
166
Professors keep insisting on teaching students sputnik-era useless topics such as calculus and algebra. Students want to learn about modern 21st century topics such as deep learning with gradient descent! . Enough with teaching number theory - teach cryptocurrencies!.
Professor @JoBoaler says math curriculum needs an update. She says math teachers used to joke that “you’re never going to be walking around with a calculator in your hand.” And now? “Turns out everybody’s walking around with a calculator in their hand.”
13
8
167
I didn’t get a chance to join OpenAI full time, but from the time I did spend with this amazing set of people, I have to agree. OpenAI is nothing without its people.
8
4
166
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.
This paper isn't even about LLMs but seems to be the final straw that popped the bubble of collective belief and gotten many to accept the limits of LLMs. About time. If "emergence" merely unlocks capabilities represented in pre-training data, the gravy train will run out soon.
27
15
165
This is a wonderful book!. I often recommend it to students - yes it’s a scary monster (1000 pages) but you can read chapters individually - every chapter is beautifully written and will teach you a lot.
4
17
163
AI advances have not made teaching algebra and calculus "outdated". In fact, these core math topics underly AI and we need to strengthen these to train the workforce of the future. See letter by top industry leaders and scientists, including @sama, @elonmusk, @ylecun ,.
@elonmusk and @sama may not agree on much of late, but do agree AI is built on strong math foundations, including algebra and calculus, applauding @UofCalifornia for recent clarifications on math requirements for admission. Many industry leaders signed:.
7
31
161
And now posted also lecture notes on variational inference and statistical physics. Thanks @franklyn_wang!.
Lecture notes for test & train robustness in ML theory seminar: talked about robust mean estimation, data poisoning, domain shift, adversarial perturbation and even detoured to multiplicative weights . Thanks @proneat for scribing!.
0
44
162
This recording of Avi Wigderson’s public lecture, explaining how many themes in CS, math and more are variations on Turing’s “imitation game” with varying imitators, imitatees, and referees, is really worth watching.
2
34
159
Sad and embarrassed that my hometown of Cambridge, MA, where Bob Moses' "Algebra Project" was founded, decided it's too hard to offer Algebra I in middle school. Effectively this means outsourcing Algebra I to Russian School of Math or private schools for those who can afford it.
Cambridge Public Schools eliminated advanced math in middle school with the aim of reducing disparities between low-income children of color and their more affluent peers. But some families and educators argue the decision has had the opposite effect.
4
18
157
Yet another work demonstrating "Anna Karenina" principle of deep learning - successful deep nets seem to learn the same internal representations, up to the "right" notion of symmetry. Supports bold conjecture of @rahiment @HanieSedghi @osaukh @bneyshabur
📜🚨📜🚨.NN loss landscapes are full of permutation symmetries, ie. swap any 2 units in a hidden layer. What does this mean for SGD? Is this practically useful?. For the past 5 yrs these Qs have fascinated me. Today, I am ready to announce "Git Re-Basin"!.
2
18
158
I guess this screenplay Goren (my kid) sent me is a not-so-subtle hint that I should make sure to give credit for the screenshot above :)
4
4
154