papeda

I agree that plenty of people have been too cavalier about slapping together some models, predicting something, and calling it a day. On the other hand, it's not like fair sentencing or fair loan recommendation is a solved problem for humans either. There is evidence that, when carefully designed, algorithms can produce more equitable outcomes than humans, for example when deciding who and how to release on bail [1].

So I think we don't want to hand over sentencing decisions to a neural network that nobody understands, but I also think careful use of (simpler) machine learning can still improve a lot of our decision-making. The question is how much these decisions improve over what we have now. There's not exactly a surfeit of wise, highly trained ethicists who are happy to make consequential decisions all day. Many human decision-makers are quite flawed too, they just get to hide behind the opacity of being a human rather than an algorithm.

To that end, a whole "fair machine learning" field has sprung up over the past few years to study this. There are like a dozen papers in the area at NeurIPS and ICML every year. There's some progress.

[1] https://www.nber.org/papers/w23180

This complaint pops up over and over again on HN. A lot of people like writing that "paints a picture", or like fancy GUIs, or want a website with a bunch of bells and whistles, or small talk without much real information. HN, in general, skews against all of these things (and so do I) but it's not a representative sample of what people in general want.

Slightly tangential question, but I have seen at least one proposal [1] for "malpractice insurance" for cops. The claim is that this would shift expenses for bad policing settlements from taxpayers to bad actors -- a bad cop attracts suits, creates settlements, pays higher premiums, and perhaps eventually stops being a cop. It has some kind of economic logic, but to the best of my knowledge this has never been implemented anywhere. Has it?

[1] https://www.cato.org/publications/commentary/make-cops-carry...

> More generally, I find that some AI researchers and practitioners with strong theoretical backgrounds tend to dismiss this kind of paper as "merely" engineering. I think this tendency is misguided. We must build giant machines and gather experimental evidence from them -- akin to physicists who build giant high-energy particle colliders to gather experimental evidence from them.

The analogy is a bit off to me. As far as I can tell, there was significant impetus from within particle physics to commit a huge amount of resources and political effort toward verifying theories with experiment. I don't see anything similar in deep learning, because in this case the "theory" is mostly that "bigger is probably better". I think that idea is pretty uncontroversial for stuff like this. And if the work reduces to marshaling enough resources, what exactly is it?

We should give OpenAI some credit for doing the damn thing, but as is the result kind of seems like an answer to a question that people weren't really asking.

Yup! Deliberately fostering phytoplankton growth using "iron fertilization" even has its own Wikipedia article [1]. The idea is that iron, not nutrients, is a limiting factor to phytoplankton growth in some areas, so we could boost phytoplankton just by throwing (a suitable form of) iron in the ocean.

IIRC, it's a controversial topic because the lack of a unified regulatory body for oceans makes experimentation difficult, and there are poorly-understood side effects with various algae. Also, wholesale manipulation of an ecosystem for scientific purposes is usually frowned upon (pollution, though, seems to get along just fine...).

[1] https://en.wikipedia.org/wiki/Iron_fertilization

I click on your link and the video looks suspicious, but I also have no clue who made it or who all those people are. Rex Chapman has a checkmark next to his name, but so do plenty of vacuous trolls. Inserting violent actors seems like a good way to discredit a protest, but "makes sense" isn't evidence.

So I typically end up just waiting a week or two for ostensibly fact-checked coverage and being out of the loop because there's so much uncertainty around developing events. But it feels like choosing feckless over reckless, so it's not a totally satisfying route.

> All carbohydrates become sugar, sugar elevates insulin, constantly elevated insulin leads to diabetes.

But my understanding is that some carbohydrates get turned into sugar more slowly, and this "flattens the curve" of blood sugar relative to consuming sugar.

You may be interested in a long series of YouTube videos put out by Chicago Reader called Key Ingredient. In each video a different pro cook from Chicago makes a dish with a different key ingredient chosen for them by the previous episode's cook.

Each video is short, maybe 5min, but is pretty dense with info about coming up with a dish. And you get to see little snippets of pro cooks putting stuff together in quiet hours outside of service, how they move and what they do without time pressures, experimenting with execution. Even better: most of the cooks are not yet famous.

Michael Carlsen's episode is one of my favorites: https://youtu.be/3hTW9jfoCBU

I saw similar advice from a very different source a couple of months ago. It was a book about exercise with the following advice for stretching: don't go until hurts, but go until it's slightly uncomfortable, then try to relax there, then repeat. Maybe a little trite, but I feel that advice is pretty widely applicable.

> This omnipresent cult of the body is extraordinary. It is the only object on which everyone is made to concentrate, not as a source of pleasure, but as an object of frantic concern, in the obsessive fear of failure or substandard performance, a sign and an anticipation of death, that death to which no one can any longer give a meaning, but which everyone knows has at all times to be prevented.

It's a huge rhetorical jump to go from "people engaging in focused activity" to "ritualistic pre-enactment of inevitable death", and even after re-reading this quote several times, I'm confused by how that jump is supposed to make sense?

I'd guess there are plenty of people in high-rent areas who'd need larger and more expensive living quarters to work from home long-term.

I agree with the main idea. Two nitpicks:

1) the random invitation to join your mailing list mid-way through the article is jarring

2) it's funny to get to the end of the article and suggest reading Seneca when, as far as I can tell, "oh you should read the stoics" pops up on HN all the time

I agree that the multi-year publication timelines for journals are in general silly. I agree that keeping up with a field by reading journals doesn't work. If by editorial review you mean "the editor alone decides", I don't think that's a good replacement.

I think the computer science publication model of a few conferences per year with ~3 month submit-to-decision timelines, based on 3-5 reviews per paper, is pretty good. But it's probably not appropriate for fields with deep contributions that take substantial time to verify, like math.

> However i would beg for highly-technical reviews, providing harsh theoretical and technical feedback, which i guess would come from more experienced or expert reviewers than postgrads or myself.

Grad students are plenty capable of writing these reviews. They might even be responsible for most of them. But the problem is still that there's very little incentive to write these kinds of reviews. As a reviewer, I don't even know how the author will accept it -- will they engage with the criticism and re-work their paper? Or will they ignore it, mildly re-write the introduction, and then re-submit to the next conference? I have seen enough examples of the latter to blunt my enthusiasm as a reviewer for providing detailed constructive feedback. Unfortunately, I think the most useful feedback may come after acceptance, when you give the talk/present the poster and people actually pay attention. It's much easier to interact in good faith when the person is right in front of you.

I sympathize with relative "outsiders" to a field who can only get expert attention through reviews. The above issue makes things difficult. Attending workshops can be useful for this.

> The ~30 papers were an outcome of collaborative R&D projects with universities and companies, most of those just having my name in the authors and requiring me to quickly read and comment them, instead of spending effort to produce the results presented.

Ah, this is not the norm in theory. Our author order is alphabetical, so the working assumption is that every co-author has seriously contributed to at least one (and hopefully all) of 1) original idea for the paper 2) technical work (proving a result, implementing an algorithm) 3) actually writing the paper. There are still people who publish 20 papers a year, but this is rare.

My understanding of the payment strategy is that it would be targeted at people who already qualify as experts but are on the fence. For example, I know postdocs and young faculty who would like to review more as a kind of community service but can't quite justify it in the face of other demands on their time. The money might be enough to tip the scales.

I don't think it will work on the other more senior pool of expert reviewers. On the other hand, I think most of the really bad reviews come from this senior pool, so maybe that's fine.

> We should be trying to maximize the growth of knowledge, not minimizing the number of wrong papers, which is what max(s/n) leads to.

IMO, filtering out noise is a big part of maximizing the growth of knowledge. I cannot read every paper in my area that goes up on Arxiv, let alone every paper in related areas. That's why it's useful, albeit imperfect, to have gatekeeping peer-review organizations to force experts to look at papers and pick out the ones they think are most useful.

There is a lot of noise on Arxiv.

I'm surprised that many people automatically assume "you should cite this paper" means "you should cite my paper". I tell reviewers to cite papers all the time. They're almost never my own papers. They're just relevant papers.

I think there is some aspect here of just finding reasons to dislike the reviewer (though I've certainly had reviewers where those reasons were pretty easy to find).

Peer review is flawed, but I don't think it's broken. For context, I'm a late-year PhD student in machine learning theory, and I submit almost entirely to conferences, which is the norm in my field.

To me, the big problem is that there is little incentive for experts to review papers. I write a few papers every year, and I review 5x that number. Slightly more senior people -- postdocs and junior faculty -- may review more like 30-40 papers. Most papers I review are not good papers. This is especially true at the big machine learning conferences like NeurIPS, ICML, and ICLR, where a disconcertingly large fraction (1/3?) of submissions are outright bad. Most of them are at best "meh". So there's not much benefit to the reviewer beyond some notion of community service.

I think my ~colleagues (PhD students, postdocs, young faculty) have similar opinions. We've certainly discussed these issues in the past. One positive step NeurIPS has taken is giving "best reviewer" awards that come from with a free registration. That's a nice benefit, even if it won't on its own incentivize good reviews. Some people have suggested outright paying people for reviews at some quasi-competitive rate, maybe $100 per review (note that this is not great -- a good review takes at least several hours if the paper is not terrible). NeurIPS also added a requirement this year that all submitting authors must review, which is an interesting experiment.

My impractical wish is that we also have some kind of "worst reviewer" sticker for all the reviewers who just paraphrase the abstract, add some milquetoast feelings on novelty, and weak reject. Less facetiously, some kind of StackOverflow reputation system might be useful. As is there's so little external signal about being a good or bad reviewer that the incentives are pretty wrong. Some kind of points system might help.

But my overall point might be: reviewers by and large aren't malicious, jealous people who are only rejecting your work because they want to protect their own turf. They are more likely good faith actors who are trying to balance time spent on a pretty thankless but societally useful job (reviewing) with time spent on much more personally enjoyable and rewarding activities (research, family, hobbies, etc). There aren't really obvious solutions to tilt the balance toward better reviews that can still handle the volume of papers.

Your points are mostly fair, though I'll push back on one. For context, I'm a late-year PhD student in ML, but I primarily do theory.

> Reviewers might not have relevant expertise to judge whether a paper is good or not...

Finding enough quality reviewers is hard. There are so many papers (heck, you put out 30 papers in 3 years!) and there is not much incentive to review. I review maybe 5x more papers than I write each year. It's doubly hard to get good reviews out of senior people. For this reason, I think grad student reviews are not a bad solution. In my experience, grad students puts in more effort and write decent, careful reviews. The very worst, low-effort reviews seem to come from established people who can't be bothered. A lower reviewer bar at workshops is fairly common and reasonable. Workshops are not the main publication venue (at least in my areas), so they mostly just want to filter out papers that are bad rather than only accept papers that are good.

> Politics...

I can't say too much about this since I primarily work in theory, but my perception from the outside is that the more "lab"-dominated a field is (i.e., accepted papers tend to come from some group with 20 PhD students and 10 postdocs) the more likely there are to be "mafias". Maybe this is a function of the kind of people who want to head up a lab, maybe it's a function of the kind of work that gets produced in factory environments, or maybe it's baseless prejudice from me.

> Israel had the year of Jubilee, where every 50 years debts would be forgiven

How would this work? Wouldn't people just not enter into contracts near that 50 year mark? Is it randomized?

If anything, I think the AMC exams are more vulnerable to prep than the SAT/ACT in the following sense: I think a bright and disciplined kid who does lots of self-study (e.g. using the free Khan Academy SAT materials [1]) has a decent chance to do well on the SAT/ACT, even compared to a bright and disciplined kid whose parents pay for classes. I think the gap is bigger for AMC stuff because, IMO, it's harder to teach yourself math competition skills. In that sense, access to paid instruction becomes a much bigger edge.

Math competition skills also test something very specific and are a pretty useless metric for most people.

[1] https://www.khanacademy.org/sat

Maybe there are two type of soft power. The first type is where a company makes a globally popular product. The second type is where a company makes a globally popular product that is recognizably "of its country". TikTok is the first, McDonalds/Hollywood/NBA is the second?

Serious question: is it a soft power win for China if most non-Chinese users don't think of it as a Chinese product?

My takeaway was more that fresh eyes and expertise in a related field is a powerful problem-solving combination.

She's clearly smart, but I'm reluctant to call anyone a genius if it causes us to view their success as, say, wondrous light from a distant star rather than recognizable human effort.

Her CV [1] lists her job title as Assistant Professor. Also, this quibble seems to drastically miss the point of the article?

[1] https://sites.google.com/view/lpiccirillo/research/cv

> Generally in the US there's no career path to becoming a tenured professor.

It's true that, in general, there are far more graduating PhD students every year than tenure-track faculty positions. But I think "no career path" drastically overstates the difficulty. It's way easier than making it as a pro athlete, for example.

> A lot of assistant professors, etc. sleep in their cars.

Are you thinking of adjunct professors, who are not tenure-track? Assistant professors typically make at least 50k+, and in the more lucrative fields (computer science, business schools, biopharm I think) >100k is common.

> You would have to have the right connections, or publish papers that were undeniably brilliant. (ie. be a celebrity.)

You're correct that an average PhD student will probably not get any tenure-track faculty interviews, let alone offers. This is especially true in more resource-constrained fields like the humanities.

Heh. I also wonder if Jordan's mentality would have let him take full advantage of the 3. This is pretty much just conjecture, but it's hard for me to square staying out and playing the math on 3s with Jordan's aggression and insistence on beating people.

> > The article even mentions the NBA Finals game where he made 6 threes in the first half, just because.

The 6 threes are a little weird because Portland insisted on guarding him with tons of space, essentially daring him to take what they viewed as a low-percentage shot. It's doubly weird to watch them repeatedly concede open 3s now -- it's not even a result of scrambled defensive rotations, they just watch him spot up and shoot.

I think I agree with you though. The fact that he had such a lovely midrange game suggests he'd have picked up the 3 if he tried. Plus that first step and the need to pack the paint to guard against those weird hanging double clutch layups, he would have had space to shoot too.

So I guess I'm making the tiresome argument that "your claim is probably right, but one specific bit of evidence is weak".

I'm not sure random draws are even better from a zoomed-out "total welfare" score. People get lot of value from preparing for the future and believing good things will probably happen, especially when it comes to their kids. The random straws outcome makes that impossible. That's a lot of utility to leave on the table.

The flip side is that parents who could not ordinarily build for their kids' futures might feel better. I wonder if that's enough to compensate.

> The adage "Popular does not equal good" has never been more relevant.

Eh, it depends. It's not true that popular ==> good, but I think there are plenty of areas where it's more true than it used to be that good ==> popular. For example, many more people now have a much better chance to get an education and demonstrate aptitude in some area. And better communication means talent is better matched to opportunity.

That's not to say we're really optimizing human potential. That doesn't even seem close to true. But I think we're doing a lot better than we were even 50 years ago.