Working on Fully Homomorphic Encryption at Google (see https://heir.dev)
Come say hi at j2kun@mathstodon.xyz
- j2kunI had to upgrade to get DDR5, for one.
- I'd be shocked if this site was never manipulated by its owners to influence opinion in their benefit.
- Read the article, which discusses this already, and maybe respond to that.
- It is worth noting that the Netflix Prize winner's solution was never meaningfully used, because Netflix pivoted from ranking content based on what you tell them you like to ranking content based on clicks and minutes watched.
To say that "we have solved ranking" because Netflix decided to measure shallow metrics and addiction is... specious at best. Instead the tech industry (in all media domains, not just streaming video) replaced improving platforms and services in meaningful ways with surveillance and revenue extraction.
- They claim the algorithm "discovered" the new techniques, but the methods described in section 5 do not seem all that novel to me. It smells like it could be "laundering" the literature [1] and reshuffling existing techniques. This is not inherently a bad thing, but I would hope that if it is borrowing existing techniques, the appropriate citation would eventually make it into this paper.
- > like building an AI product made me part of the problem.
I don't see how the author can believe that quitting their job to work on an AI startup is NOT contributing to the problem of "AI products being shoved down everyone's throats."
Except, of course, that their financial bottom line depends on not believing this.
- Scalar Evolution is one way loops can be simplified
- ...did they exercise it?
- I would contest that this is not a "bitter lesson" in the sense that it has not been demonstrated repeatedly over decades as a truism of computer science.
- > prevents progress
"progress" is doing a lot of work here. Progress in what sense, and for whom? The jury is still out on whether LLMs even increase productivity (which is not the same as progress), and I say this as a user of LLMs.
- I guess you have to read a lot into what the OP means by the italicized "the." I read it as a question of the size and scale of working groups vis-a-vis the scale of the impact.
- I don't see how "they improved the models" is related to the bitter lesson. You are still injecting human-level expertise (whether it is by prompts or a structured API) to compensate for the model's failures. A "bitter lesson" would be that the model can do better without any injection, but more compute power, than it could with human interference.
- "the" Bell Labs was effectively gone by the 1970's anyway. So the Bell Labs described in this article was "the" Bell Labs
- The US gov't broke up AT&T and killed Bell Labs for this, so at least they owe us to bust this monopoly
- > I'm pretty excited for this
Aside from historical interest, why are you excited for it?
- If you enjoy that problem you might enjoy:
Cut one corner off a chessboard. Is it possible to tile the remaining board with 3-by-1 dominoes?
(Spoiler/solution: https://www.jeremykun.com/2011/06/26/tiling-a-chessboard/)
- Nobody is using this thread to actually talk about what's in the paper, just as a place to rant about short form videos... One question that comes to mind to me is: r=-.034 a reasonable effect size? Having seen many scatter plots of r values, 0.3 seems basically like random noise. Is this just falling into the same problem as all huge meta-studies, that there's way too much variability to get any kind of clear signal?
And why, for that matter, do we need science to tell us that SFV is bad and addictive? Isn't that patently obvious from our own lived experiences?
- My position is that the negatives outweigh the positives, and I don't appreciate your straw man response. It's clear your question is not genuine and you're here to be contrarian.
- Well said.