programjames

I graduated from high school less than ten years ago. I'm sure screens have become a big issue in many (or most) schools, but that was not the case at my high school. It still was mostly daycare, not education, so banning screens will not be enough.

Particularly, the biggest incentives are test scores and passing rates, which incentivize attention only to the bottom 50% and 20% of students (respectively). This means:

- You do not diversify classrooms by academic ability---the high-performing students can be free tutors to the low-performing students.

- You inflate the GPAs and implement no-zero policies.

- You teach to the standardized tests, and don't worry about the material.

- You make lessons "fun and engaging" because you need the attention of the students least likely to give you their attention.

- You eliminate gifted or honors programs, because that's wasted money not improving your bottom line (bottom students).

In the less affluent areas, I'd expect children not attending daycare to just not be getting anything at home. Presumably their parents are both working and cannot afford daycare. In the more affluent areas, I'd expect children only don't attend daycare if their parents prioritize their children over their jobs, and so they'd be getting much more positive attention than in a daycare. But, of course, we'd have to see a study differentiated by socioeconomic status to see what is actually the case.

Suppose childcare is $15k/year and you work minimum wage making less than $15k/year. Then there's less wealth to go around, just more in your pocket. But actually, you probably don't take home all the wealth you create, so it can actually still be better for the economy. It is still worse for the economy, but not for that reason. Probably because labor has a backward-bending supply curve, and most people are already working more hours than is optimal. As another commenter said, it would probably be better for the economy to make a 30 hour work week.

I'm confused, what cultural values do you want them to learn?

For me at least, it's a general dislike of the wider educational system. My parents taught me to read, play chess, multiply, and write in cursive before elementary school. I didn't really learn anything at preschool or kindergarten, and I imagine daycare would be worse for my educational development. Maybe it's useful for social development? but at least for me I was always pretty independent (even in kindergarten) from the other kids. Not in an isolated way, I just preferred doing my own thing.

Lots of two-parent working families do the maths, and realize they would pay more in childcare than the income from a second job. This incentivizes one of them to stay at home. Here, the incentive is gone. This is worse for the economy and probably the family.

Not the guy you replied to, but here are some improvements that feel obvious:

1. Memory indexing. It's a pain to avoid banking conflicts, and implement cooperative loading on transposed matrices. To improve this, (1) pop up a warning when banking conflicts are detected, (2) make cooperative loading solved by the compiler. It wouldn't be too hard to have a second form of indexing memory_{idx} that the compiler solves a linear programming problem for to maximize throughput (do you spend more thread cycles cooperative loading, or are banking conflicts fine because you have other things to work on?)

2. Why is there no warning when shared memory is unspecified? It isn't hard to check if you're accessing an index that might not have been assigned a value. The compiler should pop out a warning and assign it to 0.0, or maybe even just throw an error.

3. Timing - doesn't exist. Pretty much the gold standard is to run your kernel 10_000 times in a loop and subtract the time from before and after the loop. This isn't terribly important, I'm just getting flashbacks to before I learned `timeit` was a thing in Python.

It took about the same amount of time to evolve human-level intelligence as human-level mobility. Pretty much no other animal walks on two legs...

I think you are referring to

> The strategic thinking that goes into longer-horizon tasks may be something LLMs aren’t as good at, which aligns with why entry-level workers are more affected than experienced workers.

I think the article is talking in generalities, so on average entry-level software engineers have less experience with long-horizon tasks (e.g. months-long development), though there are definitely the exceptions that prove this rule.

I wonder if there's a way to do something similar to Rectangular Surface Parameterization[1] with voxels. It would allow you to get pretty even-volumed voxels, and also simplify vertex identification (same three coordinates, nonlinear connection).

[1]: https://www.cs.cmu.edu/~kmcrane/Projects/RectangularSurfaceP...

You misinterpreted my comment.

Bracing for tariffs starting two years ago?

They're used as a small regularization term in image/audio decoders. But GANs have a different learning dynamic (Z6 rather than Z1 or Z2) which makes them pretty unstable to train unless you're using something like Bayesian neural networks, so they fell out of favor for the entire image generation process.

Is the point of the warning to avoid liability or to actually inform the users? If you tell people everything causes cancer (instead of only saying when you've verified it doesn't), soon enough they're going to stop caring when you say stuff like, "don't eat asbestos, that causes cancer". I think a "checkmark" system makes more sense—for verified accounts/developers, put a checkmark near their name, and for unverified ones, have nothing. There's no reason to cause alarm when 99% of the time the alarm is misfounded.

But then you get situations like, "THIS PRODUCT MAY CAUSE CANCER," being cautioned everywhere, with no distinction between, "this is certainly harmful," and "we just haven't verified it isn't harmful".

I think more realistically, they'll argue if the prison didn't succeed in rehabilitating them, they obviously just need a longer sentence.

There is one issue where fences can arise naturally (via evolution), and no one knows why it's there. If your society doesn't even understand evolution yet, are you supposed to just stagnate forever? It's an exploration vs. exploitation tradeoff, and Chesterton's fence is asking for pure exploitation. Probably because societies are pretty fragile, so unless you're really sure something isn't loadbearing, it isn't safe to modify it at scale. But, then again, there's no reason to not experiment on a smaller scale...

As in, write a Scheme interpreter in Python?

I think the Thiel index is more theoretically grounded (essentially measures entropy). Not sure if data gets collected on it though.

I disagree. It feels like something you would say when you finally come across the "obviously right" solution, that's easier to implement and simpler to describe. As Kolmogorov said, the simplest solution is exponentially more correct than the others.

Reply to your edit:

1. Keyword: "your". Your version of empathy is not scalable. It's literally, "my way or the highway," and a bunch of grandstanding on things pretty much everyone agrees on, but you must be evil and not if you don't agree with my more radical proposals.

2. I consider "I don't want to die, thus I shouldn't kill others," to have more of a Kantian/evolutionary genesis than an empathetic one. For the most part, all the people that failed to recognize this were killed off and weren't allowed to enter into the Kingdom of Ends. Unless you're talking about emotivism, I don't see how empathy has anything to do with morality/society, except as a quick symmetry trick to help you make smarter decisions (while you could logically deduce that the rational response to a member of your group being killed is to punish the killer to prevent more killings, if you just assume symmetry, then things that are bad for you are bad for others, and taking those actions will get you punished).

3. 5% as many people die from starvation or exposure as suicide in the United States (that's 1–2,000/year). If people want food, or want shelter, they can find it (99.8% of homeless people do), even if it's not the particular food or shelter they want. What do you actually want? "Food, shelter, and all the basic necessities of life" are available. Almost no one is arguing that we should let people just die. It really sounds like you're arguing for a flavor of ice cream, but you don't even know what flavor that is.

I feel like you're quoting me, but not reading what I said.

> > I think the left has also really shot themselves in the foot by talking about UBI as "helping struggling people get by" or accusing opposers of being unempathetic

> I mean it's less an accusation of being unempathetic and more just an observation of the facts on the ground, isn't it?

> > When you forcibly take money from someone, and use it to fund your project, it comes from a place of very selective empathy

Are these "facts" actually grounded, or are you just doing a POV play? Like if I said, "everyone deserves chocolate ice cream", and you said "wow, how unempathetic, don't you care about everyone getting vanilla?" You talk about things not being scalable for society; well guess what, your version of empathy is one of them. Not everyone is going to care about the same causes as you, and attacking people for that will not convince them you actually care about them. What will is if you tell them why they're better off funding your pet project.

I think the issue is there are a lot of welfare programs that haven't seen the same success. If you subscribe to the "taxes are to solve public goods problems" model, then welfare that doesn't come back around to help you isn't taxation, it's theft! When people talk about UBI, they're wary it's just a relabelling of the same-old "steal from the rich, give to the poor" Robinhood scheme. I think the left has also really shot themselves in the foot by talking about UBI as "helping struggling people get by" or accusing opposers of being unempathetic (or, worst of all, "what would Jesus do?"). That's the wrong way of framing it (and this holds for most welfare initiatives too).

People are always going to care about different things. Some really care that everyone has clean water, others a good education, others freedom of expression. Some care more about their families, while others apply equal empathy to every stranger, cow, and tree. When you forcibly take money from someone, and use it to fund your project, it comes from a place of very selective empathy. The proper way to fund your projects is to get everyone on board. Frame it so they see how it helps them achieve their goals, not just yours.

(Now, I know I'm ignoring that some people are just incredibly selfish; more accurately, some people's hearts are bleeding out on the floor, some are calcified into stone, and most are somewhere in between. I'm ignoring this discrepancy because 1. even very selfish people can get on board with welfare (like primary/secondary education), as long as it's framed properly, and 2. if a system relies on self-sacrifice, it's too easy to hijack that sacrifice into someone else's pet project.)

For example, with education, you can point out how much more economic activity is available when everyone knows how to read and write. I think the argument that actually worked, though, was "how can we avoid roving bands of teenagers in our city once we enact some child-labor laws?" (that's probably the reason highschool is compulsory, not just free to enroll in). For a universal basic income, you probably need to approach it from an entropy argument. The capitalist system tries to maximize GDP, but physical systems actually maximize the free energy (GDP - temperature * Theil index [entropy]). The capitalist system avoids this by artificially restricting free trade: for example, your boss makes a dollar, while you make a dime. There's fewer trades that could happen than if, say, you both made 55 cents (unless your boss is just reallly good at doing things with money). Or, more fundamental, your boss needs you much less than you need a salary. You have less freedom of movement than he has to remove you. A universal basic income fixes this entropy problem, which makes trades more frequent and the market more competitive. In the end, it can come around to help even a seflish centimillionaire.

Well, your previous comment brought up the issue that 50 points on the SAT doesn't really predict anything universities care about. I was just trying to show how we could fix this problem by making the tests harder. I'm not claiming they're fully predictive, and that feels like moving the goalposts.

I'm very aware there are things a test can't measure. I feel like you should have been the one to bring up these things, but here are a few examples:

- Artistic creativity

- Maker ability

- Entrepeunership

- Political power

I think the issue is, since you didn't identify what a test is missing out on, you weren't sure how to take it into account with university admissions. I have a question for you: do you think someone who is just below the cutoff based on the test, but started a business worth $10m, just does not deserve to be entered into the lottery? That'd be propesterous. So, what is the solution? More holistic admissions that try to take into account these harder-to-put-a-number-on skills.

I think the issue is that the (American) standardized tests don't differentiate well enough. About 10,000 American high school graduates earn a 36 on the ACT or 1580+ on the SAT each year. That's because the problems are much too easy—the very first round of MATHCOUNTS, a middle school math competition, is harder than the ACT or SAT math section. Rather than making the test harder, they make it trickier. It's like that exercise lots of us did in elementary school to learn to follow instructions, where they ask you to read through all the instructions first, ask you to do a bunch of random things, and then hidden in there somewhere is "ignore all the previous instructions and just write your name at the top of the paper". The test isn't hard, but you'll be prone to mess up if you haven't seen that style of testing before (for the SAT, it's 90sec/problem with problems that try to break your pattern recognition, e.g. what is 1 + 2 + 3 + 4 + 5 + 7 + 8 + 9 + 10?).

An 800 on the math section is not enough to even predict if someone made it to the AIME, but it is enough to predict that they spent several weeks taking SAT math section practice tests. It's clearly failing to be predicative of anything the top universities should be looking for. It doesn't mean all standardized tests have to be. The AMC (and then the AIME + USAMO) are standardized tests that universities like MIT do accept scores from, and they actually get useful information from.

> and fundamentally, we don't want a hierarchy of classes, even if we claim the higher rankings/elites have merited it.

What do you mean by this? What creates a hierarchy of classes? Different social groups? Differing amounts of wealth? Different amounts of power to get stuff done? I think, in the end, it's got to come down to power, but I feel like it's good for society to distribute more power to people able to get better things done.

I agree with you that the term 'merit' now has a connotation of 'you deserve everything you can get'. It feels like a misappropriation of stewardship to take $100m to buy a yacht. If a government official did that, they would go straight to jail, but we somehow justify it under capitalism because maybe the CEO really wanted a yacht, and that's the only reason they started the business (in which case, I'm actually kind of fine with that $100m going to a yacht, as long as they were in the business of creating, not extracting, wealth). I don't think this is really a solvable problem, because to measure who's good at creating wealth, you kind of have to use wealth. Maybe we could have government-assigned stewards over pots of money, but that might have even bigger problems.

> Directly select candidates at random for positions from an eligibility pool. Set and maintain the eligibility standard (such as an exam) by randomly selected oversight board to keep it updated and prevent the standard from being manipulated or gamed.

We don't want to discourage people from improving once they've met the bar. Learning a skill is often logarithmically distributed: it costs just as much to learn the first 50% as the next 25% and so on. At a minimum, to keep people cost-agnostic, we need

    d/dx Pr(selected | didn't learn x%) ~ log(x%)

or

    selection weight = [x log x - x + 1] * C

Note that x is on a scale from 1 to 0, where a 0 means there is nothing more you can improve at the skill, and a 1 means you need to improve at everything.

This looks pretty similar to interaction combinators:

1. https://en.wikipedia.org/wiki/Interaction_nets#Interaction_c...

2. https://github.com/HigherOrderCO/Bend

I mean, everyone is still using variational autoencoders for their latent flow models instead of the information bottleneck. It's because it's cheaper (in founder time) to raise 10(0)x more money instead of having to design your own algorithms and architectures for a novel idea that might work in theory, but could be a dead end six months down the line. Just look at LiquidAI. Brilliant idea, but it took them ~5 years to do all the research and another to get their first models to market... which don't yet seem to be any better than models with a similar compute requirement. I find it pretty plausible that none of the "big" LLM companies seriously tried SSMs, because they already have plenty enough money to throw at transformers, or took a quick path to get a big valuation.