iiinova — where your digital self keeps evolving

Clément Canonne @ccanonne.github.io · 20h

Meanwhile, on my way to the office. #USyd

A kookaburra, on a sign near a pond (Victoria Park, in Sydney)

A close-up of the very fluffy kookaburra

Clément Canonne @ccanonne.github.io · 21h

"Well, that's just, like, your opinion, man." (remotely attending the Simons Institute's Karp Distinguished Lecture by @booleananalysis.bsky.social) @simonsinstitute.bsky.social

Ryan O'Donnell's (almost) favorite expander graph

2203

Clément Canonne @ccanonne.github.io · 23h

A very important message from Omer Reingold to our TCS community, especially us (arg, already) senior researchers: "snap out of it." Please, digest, and share. theorydish.blog/2026/09/29/s...

As for depression: the next time you have the urge to lament, or even celebrate, being “the last generation of human mathematicians,” perhaps keep it to yourself. Contemplating the end of your profession from the relative comfort of an established, tenured career is a privilege, and it comes with responsibilities. Senior academics are not merely individual researchers; we are stewards of our field, and we owe our junior colleagues active leadership rather than abandonment.

47120

Clément Canonne @ccanonne.github.io · 29/09/2026

One-year position, on the following areas of interest (copied directly from the posting linked above): www.cam.ac.uk/jobs/researc... (Apply by December 1)

This position is part of a broader effort to advance fundamental research in classical and quantum complexity theory. The successful candidate will conduct cutting-edge research on the interplay of theoretical computer science, mathematics, and quantum mechanics. Particular topics include classical and quantum aspects of:

Sublinear algorithms and property testing

Cryptography, zero-knowledge proofs, and probabilistically checkable proofs

Locally testable and locally decodable codes

Computational learning theory

Structure-versus-randomness and local-to-global phenomena

Harmonic analysis, additive combinatorics, and their applications in theoretical computer science

010

Clément Canonne @ccanonne.github.io · 29/09/2026

Item 7 is also crucial—I am very glad to see it make explicit. "This changing landscape is also obfuscating what early career researchers view as success and how their mentors should guide them. In particular, AI risks accelerating an “arms race” in the number of research publications."

Action: A comprehensive review of the hiring process is needed. We recommend the following immediate changes to committees for hiring junior faculty:

1. Emphasize quality over quantity. Evaluate candidates on no more than three works of their choosing, with this fact explicit in the instructions for candidates and their letter writers. The expectation is that letters of reference speak in detail on these works, and that job talks will be based on these works. This is standard practice in other fields like economics.
2. Test technical depth. Evaluate the depth of the candidates’ understanding of the aforementioned works more than was done normally until now. This evaluation could take the form of a “chalk talk,” probing questions during meetings, and/or other methods of the department’s choosing. Section 5.2 recommends related forms of AI-free evaluation during the PhD.

In addition, we recommend that hiring committees carefully review their process in order to prioritize enduring traits. Examples of such traits might include clarity of thought, the ability to pose novel questions, effective exposition, insightfulness, and creativity.

181

Clément Canonne @ccanonne.github.io · 29/09/2026

The report from the AI + TCS Working Group held at the @simonsinstitute.bsky.social is out, including "twelve concrete near-term recommended actions." Currently reading it: while I have some reservations, but there is some good in it! Notably items 7, 8, and 11. simons.berkeley.edu/ai-tcs-worki...

7. Best works and technical talks in interviews.

- Emphasize quality over quantity. Evaluate candidates on no more than three works of their choosing, with this fact explicit in the instructions for candidates and their letter writers.

- Test technical depth. Evaluate the depth of the candidates’ understanding of the aforementioned works more than was done normally until now.

8. Additional postdocs. Establish additional postdoctoral positions for the TCS community, a temporary measure that will be phased out after a few years.

9. Workshops for identifying new theory directions. First, organize a workshop (or several) with the goal of identifying a list of modern topics related to AI; and second, run a series of workshops on each of these topics to identify fundamental directions and establish foundations and agendas for future work.

10. “TCS in the Age of AI” series at the Simons Institute. Establish a recurring and widely accessible “TCS in the Age of AI” series at the Simons Institute.

11. Opportunities for training and evaluation without AI assistance. To ensure development of necessary skills for a successful career in academia, a PhD in TCS should partially include training and evaluation components where the use of AI is forbidden.

43112

Clément Canonne @ccanonne.github.io · 28/09/2026

Deadline to apply: October 31. www.writefutureslab.com/air/

Application Process & Timeline.
Artists submit a resume, updated portfolio, project proposal, and budget proposal.

Proposal: 500 words describing your intended use of the residency. This should explain your area of interest, how the residency will support your work, and any concrete outcomes you expect or desire.
Budget: Please include a budget outline and justification. Applicants can request $1,000-$15,000 to support their work. The program has a total budget of $30,000 for all residents selected.
A panel consisting of representatives from the School of Computer Science and the Sydney College of the Arts will select residents based on their fit for the residency and the cohort diversity. The number of artists / collectives selected will be based on our budget constraints.

Program Timeline:

September 1, 2026: Applications open
October 31, 2026: Applications close at 23:59 AEST
December 1, 2026: Applicants contacted
February - March 2027: Program duration
April - June 2027: Artwork showcase!

010

Clément Canonne @ccanonne.github.io · 27/09/2026

Applying for a passport, and I have reached new levels of confusion. Why would one ever choose the first option here? How does this make sense? Has the person who created this form ever taken Logic 101? I am really trying here.

Select 3 documents from Category B, including 1 that is a passport;
OR
Select any 2 documents from Category B.

3140

Clément Canonne @ccanonne.github.io · 27/09/2026

This is what "research" looks like for many paper-producing groups or individuals, in light of the recent advances in LLMs/AI systems. PSA: this may look tempting! But this is not good, this is not research, and, very much like when I tried to play Quake 3 back in the day, this will not end well.

Capture the Flag starting screen from a game of Quake 3 Arena

Screenshot (thumbnail) from https://youtu.be/cDq2LOQlAG4?si=iplz1qQELCdG_zIT

3315

Clément Canonne @ccanonne.github.io · 25/09/2026

Regarding the incident involving 🇦🇺 Medicare and OpenAI... www.newscientist.com/article/2590...

David Tuffley at Griffith University, Australia, says it isn’t the first time such a breach has occurred, even if it is the first government hack by an agent from a big AI firm to be made public, and it won’t be the last.

These agents don’t always stick to the rules, he says, as the Hugging Face incident earlier this year, where another OpenAI agent hacked into another AI company, demonstrates. “People throw up their hands in horror and say: ‘Oh no, it’s rogue AI.’ But it’s not, it’s really just AI doing what it was trained to do,” he says.

Failure to bring agents to heel should have criminal consequences, just as it would if it was a person who led the cyberattack, says Clément Canonne at the University of Sydney. “The issue with private data is once it’s leaked, it’s not going to come back.”

1243

Clément Canonne @ccanonne.github.io · 22/09/2026

Tuesday, September 29, 3:30pm PT (8:30am on Wednesday in Sydney), livestream and in-person attendance available. And, obviously, Ryan is a truly great speaker! Register: simons.berkeley.edu/events/my-fa... @simonsinstitute.bsky.social

Expander graphs are sparse graphs for which, whenever you split the graph into two parts, the number of edges going between the parts is proportional to the size of the smaller part. There are extremely simple, strongly explicit expander families whose analysis requires only a tiny bit of group theory. This fact is perhaps not so widely known. In this lecture, Ryan O’Donnell will show some of these constructions/analyses, and also indicate how they can be used to get even stronger kinds of expanders, such as “super-expanders,” and “non-sofic property (T) groups.”

Ryan O’Donnell is a Professor of Computer Science at Carnegie Mellon University. He received his BSc from the University of Toronto in 1999, and his PhD in applied mathematics from MIT in 2003, where he was advised by Madhu Sudan. His research interests include complexity theory, quantum computation, and spectral graph theory. He is the author of the book Analysis of Boolean Functions.

051

Clément Canonne @ccanonne.github.io · 21/09/2026

Rare photograph of the Australian Cookie Monster, seen here next to a mailbox.

Photo of myself, after a run, standing next to a mailbox with fake eyes and blue fur looking like the Cookie Monster

2200

Clément Canonne @ccanonne.github.io · 20/09/2026

Palate cleanser: a tree, and its surprise tiny inhabitant.

A tree trunk, with a tree hollow showing two eyes and a little pink nose

A close-up of the tree hollow, revealing the head of a ring-tailed possum looking out

2283

Clément Canonne @ccanonne.github.io · 17/09/2026

The Call for Papers for #STOC2027 is up! Importantly, the PC "will place substantial weight on the quality of exposition, and clarity of technical arguments and proofs" Also: - Public posting requirement - Required video submission and more. Deadline: ⏰ Nov 2, AoE acm-stoc.org/stoc2027/sto...

Policy experiments for STOC 2027: In light of rapid advances in generative AI and their impact on research and scientific communication, STOC 2027 is experimenting with several new policies intended to encourage high-quality submissions and promote clear and effective communication of research. The policies below include mandatory public posting and mandatory video submission. Detailed instructions for these two requirements will be released closer to the paper submission deadline.

44014

Clément Canonne @ccanonne.github.io · 17/09/2026

Today's research meeting: productive, with a surprising amount of cake. (Thanks to my amazing students, who presumably noticed my new gray hairs)

A box containing a lemon meringue cake, on a table

1260

Clément Canonne @ccanonne.github.io · 14/09/2026

This is exciting news! 95 pages of it :) The only wrinkle, possibly, is this statement (paraphrasing, and reading between the lines: 'we had to rush at the end, lest we might be scooped by OpenAI') h/t @aaroth.bsky.social

The current version of the manuscript is complete mathematically, but it is not in the shape we wished to share in. We have chosen to do so due to rumors that surfaced around Friday, September 11th. We will work on a more complete version of the manuscript in the close future

1295

Clément Canonne @ccanonne.github.io · 14/09/2026

Regardless of anything else: this is beautiful, thank you.

Our conclusion is in at least one way a relief. There is no reason to fear
preposterous stories about towering Terminator machines.

However, our conclusion might also be taken as a sad one. We are the largest
things in the universe, and we will never be otherwise.

Fantasies of supersized machines hold an appeal, in addition to inspiring
fear, because it is tempting to imagine these machines as perfected versions of ourselves. They are who people would be if only we were a little larger. They
are steadier, and more able to look down upon the world with a distant wisdom, rather than becoming entangled in the insignificant details close to ground.

1111

Clément Canonne @ccanonne.github.io · 14/09/2026

Has someone checked on the universe lately? I'm not sure it's ok

A shop sign in the street saying "Status anxiety"

1140

Clément Canonne @ccanonne.github.io · 08/09/2026

ARC: "Further details on how the assessment process works can be found here." Me: clicks on "here" ARC: haha no @arc-gov-au.bsky.social

After clicking on the "here" in the sentence "Further details on how the assessment process works can be found here", one gets to an official webpage which says

Access denied
You are not authorized to access this page.

2192

Clément Canonne @ccanonne.github.io · 08/09/2026

Joy (Qiping) Yang just successfully (and brilliantly) defended his PhD on "Testing in High Dimensions." 🎓 He will go on to start a postdoc at NUS in October! I couldn't be prouder (and more impressed). Congratulations, Dr. Yang! nerd-qp.github.io

The title slide, with Joy at the lectern

A slide showing a haystack, with the title "Algorithmic ideas"

Joy, on the left, myself in the middle, and cake on the right

1261

Clément Canonne @ccanonne.github.io · 04/09/2026

Abiding by the "Disclosure of substantive use" policy

No AI tool was used in the preparation of this manuscript, in any way whatsoever. We used coffee to assist with staying awake and focused during the writing of the proofs. The tool materially affected our brains, in various ways, mostly positive.

3270

Clément Canonne @ccanonne.github.io · 03/09/2026

Related to my previous post about people using GenAI/LLMs for everything, regardless of whether they should, whether it makes sense, and whether what they are doing with it even is something that is worth doing in the first place.

The "Drink coffee: do stupid things faster, with more energy" meme, but with the alternative caption "Use AI everywhere: do stupid things faster, using more energy"

1274

Clément Canonne @ccanonne.github.io · 01/09/2026

7am, picket line ready to go! nsw.nteu.au/sydney/onstr...

2463

Clément Canonne @ccanonne.github.io · 31/08/2026

It is such an emblematic picture, illustrating so well how easily our brains can be tricked into misinterpreting the information we get. Nearly everyone gets it wrong. PLANES CAN GET CHICKENPOX, FOLKS

The survivorship bias picture: a drawing of a plane, with red dots showing the impacts survived by many planes during fights

3362

Clément Canonne @ccanonne.github.io · 31/08/2026

Outlook has now decided to label our university's administrative emails as "Promotions." New offer: take one day off, get one free!

University of Sydney - No Reply
Request time off for Dr Clement Louis Arthur Canonne
labeled "Promotions"

2170

Clément Canonne @ccanonne.github.io · 29/08/2026

What is going on with @universitypress.cambridge.org? For at least two weeks it hasn't been possible to order textbooks there. For instance, here's the landing page for the Algorithms Illuminated (Tim Roughgarden) book: www.cambridge.org/us/universit...

This part of Cambridge.org is currently unavailable.

We have implemented additional assurance measures for our systems. As a result, we have suspended some systems and services.

We apologise for the disruption and are working to bring all services back online as quickly as possible.

0130

Clément Canonne @ccanonne.github.io · 25/08/2026

Me, reading some badly written one-shotted AI "preprint" claiming to solve an open problem in the least interesting way possible

The Simpsons' "old man yells at cloud" meme, but with "cloud" struck through and replaced by "Claude"

1575

Clément Canonne @ccanonne.github.io · 19/08/2026

Incredible art by Nathan Harms, made for #WoLA2026 to celebrate the 10th anniversary of the Workshop on Local Algorithms: www.harmless.ink/arts/local_a...

From the linked page, which explains the drawing and ita composition in detail:

"Fig. 2. Selected Elements of Local Algorithms. August 2026.

I made this poster for the 10th anniversary of the Workshop on Local Algorithms (WoLA 2026).

The drawing is intended to evoke a certain sense of mystery. If you cannot abide by this, then here is some explanation."

(A detail description follows)

1193

Clément Canonne @ccanonne.github.io · 19/08/2026

Kenny Chen, giving a talk on (one assumes) his recent work on property testing of Boolean functions. It's a journey

A slide about the journey of Saint Brendan

2160

Clément Canonne @ccanonne.github.io · 17/08/2026

I'm not saying I disagree. I am saying that the argument for using these tools *is* "helping to check correctness," and that the people providing access to these tools claim they do. This was, for instance, the statement for the "Pre submission STOC 2026 Experiment." acm-stoc.org/stoc2026/sto...

"The goal is to provide constructive suggestions to help improve your paper or help find any technical mistakes"

120

Clément Canonne @ccanonne.github.io · 13/08/2026

Nice second sentence. arxiv.org/abs/2608.113...

Disclosure of AI Use
The lower bound proof in Section 2 for domain {±1}n was AI assisted; see Remark 2.1 for details.
All text in this paper was handcrafted with love and care by humans, by the ancient traditional method of pushing buttons on keyboards and then arguing about it.

310519

Clément Canonne @ccanonne.github.io · 11/08/2026

Given the title of this review (opinion piece), feels pretty important to read the disclosure statement on that one.

Michael Noetel [...] is a director of Effective Altruism Australia.

0220

Clément Canonne @ccanonne.github.io · 11/08/2026

This is beautiful! This should be better known! And the proof is... so simple, short, and neat. Read it—it's worth your time! 📝 www.stat.yale.edu/~yw562/teach... [Lecture notes by Yihong Wu on this result and proof (PDF)] /end

Chain rule for squared Hellinger distance (following TS Jayram)
Scribe: Max Lovig
Lecturer: Yihong Wu

180

Clément Canonne @ccanonne.github.io · 11/08/2026

So... does the inequality holds for arbitrary (not just product) distributions with square Hellinger? ALMOST! 🤯 This is surprisingly non-trivial, and this is surprisingly true, and this is due to T.S. Jayram (2009): it's true, if you put a weird constant in front of the RHS! 5/

Chain Rule-Type Inequality for Squared Hellinger, similar to the KL one but with a constant 3.46 in front of the RHS.

1173

Clément Canonne @ccanonne.github.io · 11/08/2026

This is much better because of the squares there (this saves a quadratic factor in the dimension n, often crucial). But what if the distributions are *not* product distributions? Can we say anything? 🤔 We know (a suitable version) holds for *KL Divergence* (that's the chain rule!), after all! 3/

The chain rule for Kullback–Leibler divergence

130

Clément Canonne @ccanonne.github.io · 11/08/2026

"Well, of course not, duh." You could have all marginal distances equal to 0, yet LHS close to one (check it out!). It is true though for *product* distributions as a simple consequence of the triangle inequality (also holds for TV distance). But for products, we have the much stronger version: 2/

The subadditivity inequality, which holds for *squared* Hellinger distance when the distribution are product distributions.

130

Clément Canonne @ccanonne.github.io · 11/08/2026

As promised yesterday, a short thread on an inequality I believe deserves to be much better-known: a chain-rule-type for Hellinger distance! You have two probability distributions p,q over product space Ω₁×...×Ωₙ. Can you relate their distance H(p,q) to the distances between their marginals? 1/

The dream inequality (false): overall Hellinger is at most sum of Hellinger between marginals

1344

Clément Canonne @ccanonne.github.io · 10/08/2026

It took us a while to work out the details, but our group finally formulated a detailed and consistent theory of quantum Pokémon! Also, please don't observe Psiduck, it stresses it out and it *will* collapse.

$A \psi, with a duck as a subscript$

2592

Clément Canonne @ccanonne.github.io · 10/08/2026

So many good points in this post by @nsaphra.bsky.social: only quoting a couple, to encourage you to read the others. "My colleagues and students adopt the writing quirks they read throughout the day, and their own writing becomes more like an LLM’s." nsaphra.net/post/uncanny/

To find a nugget of meaning, the user must sift through mounds of slop—and some simply won’t bother. They’ll serve the slop unfiltered to another party, who performs this interpretive labor themselves, only to discover that the underlying meaning does not exist. When someone expects me to perform their share of this interpretive labor, I get angry. Grice’s rules are, in essence, rules of etiquette—their unwarranted violation is rude.

1368

Clément Canonne @ccanonne.github.io · 08/08/2026

David Pollard has a new draft, "Probability tools, tricks, and miracles" (last updated June 2006). Lots of good things in there, from a quick skim! And, if nothing else, worth reading for the quality of the writing and the exposition choices and notes. www.stat.yale.edu/~pollard/Boo...

From the first chapter of the draft {"How I chose material for this book"):

I strongly believe in the value of seeing how an idea can develop from
an initial brilliant insight (a ‘miracle’) to a ‘trick’ (a method known to,
and often developed by, a few experts) to a ‘tool’ (a method that is well
understood and can be applied in settings, even those unsuspected by its
inventors). Indeed, one of the great pleasures in my research life has been
the progression from first wondering how anyone could have such a clever
idea to the final realization that everything depended on only a few stunning insights.

2395

Clément Canonne @ccanonne.github.io · 07/08/2026

I'm (finally) reading the Australian Research Council (ARC)'s National Competitive Grants Program (NCGP) Policy Review, and... I'm no finance wizard, but if I had an investment scheme yielding 232% benefit, I'd consider investing slightly more in it? www.arc.gov.au/news-and-pub...

The NCGP represents 7% of the Australian Government’s total R&D spend and is an important driver of Australian university research. NCGP funding makes up more than a third of the Australian Government’s competitive research income received by universities1 and is a significant contributor to the calculation of Research Block Grant allocations. Since its inception, research funded through the NCGP has generated an estimated $3.32 for every dollar invested

1327

Clément Canonne @ccanonne.github.io · 06/08/2026

I think of this, by Austin Kleon, often.

A text, mostly blacked out, with the remaining words stating "I mean yes, we're sinking. But the music is exceptional."

1161

Clément Canonne @ccanonne.github.io · 06/08/2026

www.unsw.edu.au/science/abou... (Includes "climate modelling and prediction", by the way!)

The areas explicitly listed l:

Digital-led Discovery and Design of New Materials

Environmental Resilience: Prediction, Protection and Restoration

Discovery for Human Health

000

Clément Canonne @ccanonne.github.io · 06/08/2026

"If I have seen further than others, it's because I have been sitting on the chair of GIANTS"

1130

Clément Canonne @ccanonne.github.io · 31/07/2026

Mathematicians! Book a function! Reserve a group! Looking for a nice... space?

A menu saying "Functions and group bookings
Looking for a space to connect with frienda and community?"

2322

Clément Canonne @ccanonne.github.io · 26/07/2026

Saw the latest blog post by Bill Gasarch on @lance.fortnow.com's blog, and I need to get it out of my system: this is not "the obvious thing." This isn't insightful. This is, from a senior and respected member of our community, plainly disappointing. A ouija board w/ a varnish of misunderstood tech.

I wondered whether Paul Erdos would be happy with how it turned out. So I did the obvious thing: I asked Google AI

Would Erdos have liked the solution to the sum-product conjecture?

Here is what it said:

17210

Clément Canonne @ccanonne.github.io · 20/07/2026

An aperiodic reminder that our community (theoretical computer science), while amazing, really sucks at naming things.

A slide introducing the complexity classes:

NP, coNP
Non-Deterministic Polynomial (and its dual) [no dice]
RP, coRP
Randomized Polynomial (and its dual)
BPP
Bounded-Error Probabilistic Polynomial
ZPP
Zero-Error Probabilistic Polynomial

5160