See cims.nyu.edu/~tristanb/st... for what one side is saying, esp. page 2 onwards.
Who's Who of AI
Clément Canonne
Why they matter
Researcher with public evidence across AI research.
- AI signals
- 16 past 30d
- Sources
- 14 distinct domains
- Discussions
- 32 past 30d
- Latest signal
- 6d ago
Senior Lecturer #USydCompSci at the University of Sydney. Postdocs IBM Research and Stanford; PhD at Columbia. Converts ☕ into puns: sometimes theorems. He/him.
What they're sharing
cims.nyu.edu
Fair Moderation, Equitable Access, and AI: arXiv’s Updated Rate Limit Policy blog.arxiv.org
Leiden Declaration on Artificial Intelligence and Mathematics leidendeclaration.ai
Now is the Time to Give LLMS Access to the ACM Digital Library Why ACM Believes the benefits outweigh the risks in opening up the ACM Digital Library to large language models docs.google.com
Distribution-Free Halfspace Testing with Samples arxiv.org
Give Authors Control — A Petition to arXiv math.mit.edu
Truly Subquadratic 3SUM and Truly Subcubic APSP via Triangles in Sparse Lopsided Graphs arxiv.org
math/preprints at main · openai/math github.com
AI + TCS Working Group simons.berkeley.edu
[2605.23225] Entropy Equivalence Testing arxiv.org
Life on the Uncanny Precipice | Naomi Saphra nsaphra.net
Near-Optimal Pure Machine Unlearning for Smooth Strongly Convex Losses arxiv.org
Articles & links
"We want to encourage authors to be mindful with their work, taking care that they are submitting something substantial of scholarly interest that meets arXiv's guidelines and standards." 2 submissions per month, 3 active submissions blog.arxiv.org/2026/10/01/u... (per submitt…
AI Weekly's analysis
→
Read full analysis →
A list of principles put forth by mathematicians, for mathematicians and other researchers, regarding the use of AI in research. "Number #9 will surprise you!" leidendeclaration.ai
AI Weekly's analysis
→
- The Leiden Declaration, released June 2, 2026, warns AI threatens proof integrity, attribution, and peer review in mathematics.
- Over 2,654 signatories including Fields Medal winner Terence Tao have endorsed the community-initiated declaration.
- The International Mathematical Union backs the declaration, which makes separate recommendations to researchers, publishers, policymakers, and AI developers.
Read full analysis →
Take some time (very little) to weigh on this: "Now Is the Time to Give LLMs Access to the ACM Digital Library" docs.google.com/forms/d/e/1F... (not that I want to influence anyone, but my own position is along the lines of "hell no.")
Nice second sentence. arxiv.org/abs/2608.113...
AI Weekly's analysis
→
Read full analysis →
Clément Canonne reposted
If you are an academic who uses arXiv, please consider signing this. **Give Authors Control: Let Us Decide If AI Bots Can Read Our arXiv Papers** math.mit.edu/~elmos/petit...
AI Weekly's analysis
→
- A petition on an MIT math faculty page asks arXiv to let authors toggle, at submission, whether bots can access full text or metadata only.
- The organiser is not named and no signature count is published; signing requires an institutional email, with webmail and disposable addresses rejected.
- The proposed tiers are "Full Access" to the complete text and source files, or "Metadata Only" to title, abstract and citations.
Read full analysis →
View on Bluesky →
Clément Canonne reposted
Unbelievable stuff! arxiv.org/abs/2610.067... APSP and 3SUM refuted!! This is big.
AI Weekly's analysis
→
- Alman and Williams give deterministic O(n^1.9992) time for 3SUM and O(n^2.9995) time for APSP, the first polynomial improvements on the textbook algorithms.
- The 76-page preprint says the result refutes the 3SUM, APSP, Exact Triangle, and Zero-Weight k-Clique hypotheses that anchor fine-grained complexity.
- The engine is a thin matrix multiplication subroutine that computes a sparse N^2/sqrt(D)-entry subset of XY in O(N^2/D^0.063) operations when D is at most N^(1/18).
Read full analysis →
View on Bluesky →
Clément Canonne reposted
OpenAI just blooped hundreds of proofs and counterexamples to open problems: github.com/openai/math/... Boggling.
AI Weekly's analysis
→
- OpenAI released 722 manuscripts organized into 372 families on October 6, produced by an unreleased internal frontier model pointed at roughly 4,000 problems.
- Each accepted result averaged about three hours of ChatGPT Pro thinking compute, with only ten manuscripts shipping with abridged reasoning summaries.
- Many papers include Lean formalizations for computer-checked verification, but the repository warns some unformalized results could have issues.
Read full analysis →
View on Bluesky →
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.berkele…
AI Weekly's analysis
→
- The Simons Institute published 'AI and TCS: The Next Six Months' on September 28, 2026, following a two-day working group at Calvin Lab.
- The report lists twelve recommended actions, from a standardized AI methodology section in papers to grant coverage of reasonable LLM expenses.
- Discussions were informed by 134 survey responses; organizers were Tony Feng, Zeph Landau, Amit Sahai, and Nikhil Srivastava.
Read full analysis →
Our paper on Entropy Equivalence Testing was just accepted to #RANDOM2026. Congratulations to Joy (Qiping) Yang and Yash Pote, who led the charge! arxiv.org/abs/2605.23225
AI Weekly's analysis
→
- Entropy equivalence testing needs significantly fewer samples than standard closeness testing for distributions.
- The paper delivers the first non-trivial closeness testing algorithm for low-degree Bayesian networks.
- Matching lower bounds establish near-optimality, revealing how hard the relaxed problem is in principle.
Read full analysis →
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/
Clément Canonne reposted
Near-Optimal Pure Machine Unlearning for Smooth Strongly Convex Losses Matthew Regehr, Gautam Kamath, Andrew Lowy http://arxiv.org/abs/2606.01527
AI Weekly's analysis
→
- The paper proves that "(ε, δ)-unlearning has no statistical advantage over pure ε-unlearning" for smooth strongly convex losses.
- For ε less than or comparable to the model dimension d, retraining from scratch is information theoretically optimal; you cannot beat it.
- For ε well above d with large unlearning requests, their algorithm delivers an exponential accuracy improvement over retraining and DP baselines.
Read full analysis →
View on Bluesky →
Their own posts
Recent commentary
OpenAI: "We’re releasing a broad range of new mathematical results produced by an internal frontier model." (722 manuscripts)
For context, to ground the conversation and explain one of the issues. USD20M, which is a self-reported (likely low-balled) estimate of what OpenAI burned in *one week* in their scooping race, would fund 100 full-time postdoctoral positions for 2 years (including overhead). 100 postdocs. 2 years.
Me, reading some badly written one-shotted AI "preprint" claiming to solve an open problem in the least interesting way possible
Preparing a blog post, and as a side product I wrote some short thoughts on AI in reviewing, in TCS and beyond.
The OpenAI breakthrough on the unit distance problem seems genuinely impressive to my semi-layman eye. (It is!) One possible takeaway, though, is that if you throw an incredible amount of money and resources at focused research, there WILL be impressive progress. I wish we tried that, too.
The "paperclip maximizer" is a silly thought experiment: an entity mistakes an (imperfect) metric for an absolute goal, and proceeds to diverting all resources to producing paperclips. Silly. In other news, Anthropic and OpenAI are burning millions in a race to establish Erdős' 9173439 conjectures.
I for one am elated to see the big AI companies take societal concerns seriously, and commit a fraction of their time and money to assessing the impact of their products. In other news, the Joker promises to dedicate a couple of henchmen to the study of potential side effects of his laughing gas.
Looking for takes by TCS (Theoretical Computer Science) senior academics/researchers about AI and their discipline/research values. E.g., essays, detailed blog posts. (Specifically TCS: there is a lot from Maths, but while adjacent there are enough differences in culture it's worth distinguishing)
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.
We should also keep track of the number of AI-generated results desk-rejected from conferences and journals due to being of subpar quality, to help keep things in perspective. "OpenAI announces 722 desk-rejections from leading scientific venues!" would not have the same ring, I know.
Their network
In Clément Canonne's orbit
Center = Clément Canonne. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you Clément Canonne? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/ccanonne-github-io)