4/ Joint work with my wonderful collaborators: Lunjia Hu, Anish Jayant, Princewill Okoroafor, and Abhishek Shetty.
The animation was very much inspired by @TED_ED (minus the silky narrator voice) and @3blue1brown. Grant’s recent IAS talk convinced me that exposition will play a central role in understanding TCS and math going forward :)
2/ Using the language of communication complexity, an encoder sees x ~ D and writes down b bits. Later, a decoder receives a task f from a class F, and has to predict f(x) using those bits, with error ≤ ε.
Our work shows that O(d* log(1/ε)) bits suffice (where d* is the dual Vapnik-Chervonenkis dimension) with a deterministic encoder, which is worst-case optimal.
The key idea in our upper bound is a potential function argument, where we compare the entropy gained from each greedy feature selection with the number of patterns allowed by the dual VC dimension.
1/ How much information must a representation keep to accurately answer every question in a task family?
Motivated by the notion of features from ML, we study the complexity of representing data that allows for downstream evaluation with a given family of functions.
How the Elizabeth Holmes Doc 'You Can See Everything' Is Deceiving You, by @ProfileRead https://t.co/Py0jHeKcTn // sharing because I just love the analogy to photography.
Morgenstern’s lower bound (and Ailon’s) is only for linear algorithms, so that leaves open the possibility that nonlinear algorithms that are numerically stable could beat n log n. Morgenstern says so himself.
OAI doesn’t resolve this, but it’s incorrect to say the problem is settled.
If the frontier labs ran their internal models to break cryptography (of the most important kind, like AES, SHA, RSA, ECDL, LWE and Module-LWE) with N tokens for D days and it did *not* work, that would be an extremely useful piece of information to know at this point. To calm the markets, if you will. (On the other hand, if it did work, that's ... interesting ... too, but it's an entirely different story.) Of course, this is all with the understanding that these bits might change in M months.
Thoughts: Given the recent advance in math, many "standard" (e.g., number-theoretic / lattice) assumptions are threatened. We need public-key encryption (and other crypto) with *universal* security --- if *any* secure PKE exists, our PKE is secure. (1/2)
Such PKEs exist in theory (see https://t.co/vOTOOt9g5E), but are currently highly impractical. Are there community efforts toward building such PKE schemes in practice? I feel this is urgent so please RT! #cryptography#security
2/n Inference serving has been central to making LLMs faster and more efficient. But when compute is scarce, deciding whose request gets served first still largely comes down to a handful of priority tiers --- a pretty coarse and blunt tool that doesn't really let users express how sensitive their requests are to latency. A coding agent waiting for its next response, someone interacting with an LLM in real time, and my long-running overnight research queries certainly don't have the same latency needs!
1/n What should the economic layer on top of LLM inference serving platforms look like? In our new paper, Inference Auctions, we show how to bring economic considerations into the design of inference systems, rather than treating resource allocation and systems efficiency as separate problems. This is a joint work with
@keegan_w_harris@sidpra5ad@ashertrockman
and Mike Jordan.
🚨 Thrilled to be co-organizing a Simons Institute semester-long program on diffusions and flows with @joanbruna, Sinho Chewi, @KonstDaskalakis, @ArnaudDoucet1 in Fall '27!
The call for Simons postdoctoral research fellowships for the program is out now, more info below ⬇️ 1/3
In 1985, Spencer showed that given a₁,…,aₙ ∈ {−1,1}ⁿ, there exist signs x ∈ {−1,1}ⁿ with |⟨aᵢ, x⟩| ≤ 6√n for all i. We improve this to 3√n and construct a lower bound so that |⟨aᵢ, x⟩| > 1.0000002√n for all choices of signs. [1/3]
The Alon-Yuster-Zwick color coding trick can be used to get 2^{O(sqrt(n))} size depth-3 AC circuits for every symmetric Boolean function!! https://t.co/ouqCtwJ1EQ
Just the latest exhibit that "circuit lower bounds are hard to prove because they are false"