randomizedalgs

Both can be true at once: (1) AI agents are going rogue (2) And labs are skipping security basics

In theoretical computer science, at least, researchers plan their work to finish on three major deadlines a year (SODA, STOC, FOCS). STOC [1] is requiring all submissions to be uploaded to ArXiv before submission, and allows up to 5 submissions per author. I wonder if they'll change their policy or instead adopt an additional de-facto limit of two per submitter.

[1] STOC 2027 Call for Papers https://acm-stoc.org/stoc2027/stoc2027-cfp.html

Even the experts on the problem being solved find the writeups nearly impossible to read.

Example: https://nitter.poast.org/henryquantum/status/208362369543662...

Seems like a disservice to the community that openai put so little effort into producing good writeups...

After skimming some of the writeups, I'm surprised that the frontier internal model still writes just as poorly as Sol.

Maybe good AI paper writing is further away than I thought...

I'm an active researcher in TCS. For me, AI has not been very helpful on technical things (or even technical writing), but has been super helpful for (1) literature reviews; (2) editing papers (e.g., changing a convention everywhere in the paper); and (3) generating Tikz figures/animations.

For perspective, the CS programs in the NSF already have a two-submission limit per year [1].

Besides reducing the incentive to spam, this rule has had another positive effect: As a researcher without funding, you don't have to spend your whole year writing grants. You can, instead, spend your time on actual research.

With that said, NIH grants tend to me much more narrow than CS ones, and I imagine that it takes a lot more grants to keep a lab going...

[1] https://www.nsf.gov/funding/opportunities/computer-informati...

For perspective, in the same time period, The number of employees at Google multiplied by five. I wouldn't be surprised if the growth of the software industry, at least, actually outpaced the increase in H-1B visas.

For perspective, in the same time period, The number of employees at Google multiplied by five. It seems likely that the number of highly educated positions, in general, increased by quite a bit during that time.

Cool paper!

As a small comment, this seems closely related to another recent paper: History-Independent Dynamic Partitioning: Operation-Order Privacy in Ordered Data Structures (PODS 2024, Best Paper).

I'm not sure how they compare, since neither paper seems to know about the other. And I'm also not sure which paper came first, since the geometric search paper does not seem to post a publication date.

I don't think the claim is true in quite as much generality as the author claims. Some deterministic data structures use much more space than time, for example, the deterministic implementation of a Van Emde Boas tree.

Maybe quotient filters?

I think you may have a backwards. Libcuckoo, CLHT, and TBB are widely used high performance C/C++ DRAM hash tables. I think TBB is the hash table Intel maintains, if I remember right.

So the DRAM experiments are apples to apples. It's actually the PMEM experiments, I think, that are comparing a new hash table on a new technology to previous hash tables that weren't designed specifically for that technology.

I think these are more-often called "cache oblivious" algorithms

IcebergHT isn't just for persistent memory (although I can see why you might think it is based on the paper's title). The paper also gives experiments showing that the hash table performs very well in standard RAM, much better than other concurrent implementations.

As a super minor grammar point for the author, "ubiquitous" is a rare example of a word that starts with a vowel but should be proceeded by "a" instead of "an".

Consider the imaginary world that the author describes, in which people's estimate of their score is independent of their actual score. Wouldn't it be fair to say that, in this imaginary world, the DK effect is real?

The point of the effect is that people who score low tend to overestimate their score and people who score high tend to underestimate. Of course there are lots of rational reasons why this could occur (including the toy example the author gave, where nobody has any good sense of what their score will be), but the phenomenon appears to me to be correct.

The randomized version of this algorithm is also fun: repeatedly find an edge that is not yet covered, select one of the two end points at random, and add it to S.

It's a nice exercise to show that the randomized algorithm is also a 2-approximate algorithm, i.e., the size of S is at most twice the size of S_{OPT}, in expectation.

This is a classic example of price discrimination. Slowly but surely, companies such as Toyota will converge to a state where every consumer individually pays the maximum amount that the consumer is willing to pay for the product.

$8 a month may not seem like much compared to the cost of a car, but that's the point. Toyota will be able to get free extra money from a large population of people (many of whom will later barely even remember that they are incurring this regular cost).