yazi

"machine" "learning" "researcher"

[yazi @autocontrastive]

@MLwagie @adi_baradwaj Unrelated to the broader conversation tho, I realized the kid did notably get into UT, the ~6th best CS program in the country? With 1590+val he almost certainly also got into Turing Scholars, which is, in my books, a very elite program.

[yazi @autocontrastive]

@MLwagie @adi_baradwaj To answer your question about why JS is relevant, I think that selective, meritocratic industries are good evals of how good universities are at doing selection. They have strong financial incentive to recruit smarter, overlooked people, if elite university signaling is bad.

[yazi @autocontrastive]

@MLwagie @adi_baradwaj Personally, I don't feel too strongly either which way about DEI of the modern era. I would also be okay with complete random selection if you get above like a 1550, and guaranteed admission for a small set of exceptional people.

[yazi @autocontrastive]

@MLwagie @adi_baradwaj Factually, I claim
1. Top schools successfully capture almost all the best talent
2. People who would have been counterfactually admitted under fully meritocratic admissions likely weren't going to be in the top 5% of their class.

[yazi @autocontrastive]

@MLwagie @adi_baradwaj But indeed, rationalized arguments can still be true. You have not said anything about any of my claims or arguments!

[yazi @autocontrastive]

@MLwagie @adi_baradwaj Hence why I think that mildly non-meritocratic admissions is good, because you're the tradeoff is a noticeably better campus culture for like 15% less SWEs (as that's the general population would benefit most from better SAT sorting imo).

[yazi @autocontrastive]

@MLwagie @adi_baradwaj Top schools are very very good at getting the really good people. Simply look at the undergrad distribution of the incoming JS new grads. Conversations about DEI/better testing are just shuffling around people around the 50th percentile.

[yazi @autocontrastive]

Once again, the goal of top colleges is not to select the top-k smartest people. The point is to make sure you get the far-right tail of very very smart people (top 150 of a class of ~1k-2k). The rest can be "smart enough" people semi-randomly sampled based on e.g. personality.

[Deva Hazarika @devahaz]

Business idea: make a “hard SAT” and market it as an optional indicator students can submit as part of their application resume

[yazi @autocontrastive]

@zagrebbi erm, not sure the JS incoming employee class is substantially different from the distribution in the chart.

[yazi @autocontrastive]

@afrenkai def disagree. calc + linalg is definitely enough outside of generative image modeling (and even many parts of generative modeling can be abstracted). an undergrad could have done BERT/ResNets/GPT/Attention.

[yazi @autocontrastive]

This power law is also in effect within top schools too. The top 5 math majors at Harvard and MIT simply cannot be compared to the top 5 at Cornell.

[yazi @autocontrastive]

People incorrectly assume that the floor of top schools is higher than the ceiling of state schools, which is completely wrong. The value of top schools is in their tails. The best students at HYPSM are light years ahead of the best students at state schools.

[Aidan @aidannonx]

All of my roommates at KU had perfect test scores, 5.0 weighted GPAs, and were triple majoring in physics, astronomy, and math, double majoring in engineering and chemistry, and double majoring in chemistry and biology for med school

They were denied to UT Austin. You have no clue what you’re talking about

I was in the 99th percentile, a nationally competitive debater and business student, and had two years of college done at 17

Ivy League students are taking remedial math

Setting aside the fact that I will be outearning Ivy League students of the same peer group and that I’ve had many Ivy league students work with me or for me

[yazi @autocontrastive]

@HerbDeanGold @cds979 @aidannonx 99% of people in remedial math are athletes, and I think that's okay. Also, most don't have "remedial" math - the lowest is pre-calculus. And intro classes are actually very rigorous relative to HS versions of the classes. For instance: https://t.co/jRRA3U5c9x

[yazi @autocontrastive]

@flawedaxioms you ask, arxiv provides
https://t.co/ySk45Q8PP5

[yazi @autocontrastive]

note: i'm referring to a form of "hands-on" advising, where mentor is intimately familiar with very low-level details of project. a busy faculty can still provide high-level input or direction, which can be very valuable still.

[yazi @autocontrastive]

hot take: in ML, the role of an advisor is not to "advise" a student anymore. most top phd admits come in with >=1 first author paper and don't need advising. highest EV move for both prof and student is to facilitate access to networks (e.g. by taking industry positions).

[Peyman Milanfar @docmilanfar]

While this is an impressive array of talent, over 80% of the researchers Russ brags about were (self-) advised while he was juggling academic duties with major executive roles at Apple or Meta.

He's not alone in this respect. Elite programs attract talent that buoys the lab, giving advisors the bandwidth to juggle significant outside roles. I'd be careful about assigning too much credit to the advisor and not enough to the students' raw talent and self-discipline.

[yazi @autocontrastive]

@PosterInternet imo the only thing keeping american count somewhat afloat in top phd programs is the fact that a large chunk of smart talent cares about soft, non-monetary factors. e.g. things like doing "interesting" work, good intellectual culture, etc.