wxnx

The notion of a standard language is a modern idea (not in the sense of modern vs. postmodern, but historically).

The idea that the only valid basis for distinguishing between dialects is geography strikes me as odd, but suppose we accept it as a premise. Why is it less valid to allow correctness to be defined, say, with respect to local borders as opposed to the borders between countries?

There's confounding here that you're ignoring. For reference, I'm a machine learning research scientist who started in bioinformatics, initially lured by the possibility of a machine learning solution to the protein folding problem.

Google's research arm has made leaps and bounds in a particular field (deep learning) and then managed to apply it successfully to a very, very hard problem (protein folding). That other companies failed to adapt Google's successes in deep learning faster than Google is not surprising at all to me.

One might argue that the impact of academic-big pharma collaboration (in the form of funding for research projects related to CASP) is what enabled a company like Google, with no independent desire to collect the massive amounts of wetlab data required to evaluate or develop a tool like AlphaFold, to even participate.

More importantly, AlphaFold hasn't really solved the protein folding problem well enough for drug development. So, the entire debate might be moot.

Interestingly, Poe is not alone as a famous writer of fiction who became fascinated by a particular branch of zoology. About a century later, Vladimir Nabokov was obsessed with butterflies, with one of his hypotheses (then thought of as largely insignificant and incorrect) later being borne out by genetic studies -- ironic, given the man's distaste for the idea of genetics.

Nabokov himself even went so far as to say:

> It is not improbable that had there been no revolution in Russia, I would have devoted myself entirely to lepidopterology and never written any novels at all.

I think the usual argument is more along the lines of: Google's greed prevents them from putting the proper controls in place to prevent these sorts of abuses of the system, because those controls would cost too much. Not so much that any particular specific instance of abuse is particularly profitable.

I'm not entirely sure what you mean by human biodiversity.

As for the others on your list, non-alt-right people (cf. non-alt-right sites) might be more inclined to discuss (cf. host discussions about) these "facts" if their primary use wasn't to be woven into otherwise fictional narratives claiming to be coherent justifications of racism and homophobia.

Quite right, poor choice of phrase on my part.

My point was that there are people who are e.g. legitimately surprised when their high-speed data connection drops and it prevents them from using Google Maps on their smartphone. To them, it's functionally a GPS, not Internet-based at all. Similarly for certain messaging apps, which integrate with the phone's OS so tightly that it's not apparent that they're using Internet-based communication vs. SMS.

My larger point was that to then expect all respondents to a phone survey (on a measurement significantly linked to technical literacy, no less) to make similar distinctions unprompted is perhaps a bit of reach.

> My first reaction was “7% sounds low”, but again, that’s probably biased by my own experiences.

That was my reaction as well, particularly based on how the question was asked - it was very specifically about the Internet, not mentioning specific services that someone with low technological literacy might not associate with Internet usage (e.g. Google Maps, or Facebook Messenger) [1].

[1] https://www.pewresearch.org/wp-content/uploads/2021/03/Non-i...

A significant portion of the US is illiterate [1]. One might assume that being illiterate affects internet usage. Not quite as romantic as neo-Luddism, or even making a conscious choice not to expose oneself to potentially problematic patterns of communication.

[1] https://nces.ed.gov/datapoints/2019179.asp

> I have nothing against supporting GPUs (although I think their use is overrated and most people would do fine with CPUs), but Python really needs a general purpose, high performance autodiff.

As someone who works with machine learning models day-to-day (yes, some deep NNs, but also other stuff) - GPUs really seem unbeatable to me for anything gradient-optimization-of-matrices (i.e. like 80% of what I do) related. Even inference in a relatively simple image classification net takes an order of magnitude longer on CPU than GPU on the smallest dataset I'm working with.

Was this a comment about specific models that have a reputation as being more difficult to optimize on the GPU (like tree-based models - although Microsoft is working in this space)? Or am I genuinely missing some optimization techniques that might let me make more use of our CPU compute?

> By that I mean not that expressing the meaning of a word is forbidden, not the idea, but the utterance itself. If, like me, you tend to assume that cultures become more progressive and tolerant over time, this is an interesting very strong counter-example. Having a word that you that literally can't say even in quotation or reference implies something almost akin to magical thinking. As if the two syllables themselves, even when stripped of meaning, still have some kind of power.

I don't know that the people who leveled accusations against this guy are completely equating "usage" and "meta-usage".

I do know that there are people who view "meta-usage", particularly in casual conversation, as problematic. The n-word is viewed by them to have connotations of severe power imbalance and violence, particularly in the mouth of a white person. The idea that they should have to hear the word in casual conversation, without warning, whether it's intended to wound or not, is contentious. Intent matters, but so do the words themselves. To put it in your terms, some words cannot just be "stripped of meaning" to some people.

I personally empathize with people who feel this way, despite being a white guy for whom no such word with such visceral effect seems to exist. Even if discussing racism in a conversation, I probably just wouldn't whip out a historical quote from someone that happened to have the n-word in it. And I don't think that's as ridiculous as you're making it out to be.

Nobody is suggesting that contextualized, or (ideally, to some) forewarned, meta-usage will never have its place. But that's not really what happened in this story, is it?

> If Ehrlichman did say that (as you pointed out, even that is disputed) it's quite absurd. The prohibition of drugs started decades before the Vietnam War or the Civil Rights movement, and it was an international effort, in response to real dangers. It's not tenable to argue that it was all about control, nor that Nixon's motives (if those were his motives) were even relevant to the passage of the CSA with only 6 opposing votes in Congress.

Nominally, perhaps. But nobody can argue against the fact that widespread enforcement efforts stepped up drastically in the late 60s and 70s, with drugs being a particular focus after Nixon declared war on them explicitly. Just look at any chart documenting the incarceration rates [1], not to mention the literal creation of agency with (in modern times) a multi-billion dollar budget dedicated exclusively to drug enforcement.

[1] https://www.nap.edu/read/18613/chapter/4#35 figures 2-1 and 2-6

Well, you've successfully constructed a hell of a strawman! My comment was to express my support for the following statement:

> I'd say more generally, it's about control and having something police / government can hold over almost anyone, or any demographic, if convenient for whatever reason.

I didn't say anything about Democrats or Republicans. I happened to use a quote that touched on Nixon in support of my claim.

I would also note, for people not well-versed in American history, that up until around the 1960s, the concept of a Southern Democrat still existed - people mostly identified with opposing desegregation in the American south. Some of these politicians, including Strom Thurmond, who voted against civil rights and for the Controlled Substances Act, later became Republican figureheads. Party ideals shift over time.

With that said, I don't think the current state of drug laws in the US are an issue of party. It's a lack of evidence-based policy, and it's extremely problematic.

One might not blame politicians for looking at the effects of narcotics and their relation to organized crime in the mid-20th century in America and say: "we need to legislate this, because the problem is ballooning and we don't have any other tools at our disposal." Addiction science didn't exist yet.

One might even forgive the Controlled Substances Act as an extension of this reaction, were it not for the fact that through the subsequent creation of the DEA and the scheduling system, the effects of drugs and addiction became significantly more difficult to research in the US. This happened under Nixon, and I think that's why he receives a significant portion of the blame for the current state of affairs (rightly or wrongly).

The DEA has also simply failed in its remit. They've shown themselves to be completely incapable of assessing the medical utility and controlling the abuse potential of drugs (marijuana on the one hand, prescription opioids on the other). This, combined with the fact that the CSA and the DEA's existence prevent us as a country from evaluating other approaches to the drug problem, make for a lose-lose situation.

This, in my opinion, is closest to the truth - not to discount the issues of race that have historically propagated downward through the system and continue to affect people today.

From John Ehrlichman, a Nixon associate:

> You want to know what this was really all about?” he asked with the bluntness of a man who, after public disgrace and a stretch in federal prison, had little left to protect. “The Nixon campaign in 1968, and the Nixon White House after that, had two enemies: the antiwar left and black people. You understand what I’m saying? We knew we couldn’t make it illegal to be either against the war or black, but by getting the public to associate the hippies with marijuana and blacks with heroin, and then criminalizing both heavily, we could disrupt those communities. We could arrest their leaders, raid their homes, break up their meetings, and vilify them night after night on the evening news. Did we know we were lying about the drugs? Of course we did.

Note: the veracity of this particular quote has been contested, but the story it purports to tell is compelling in my opinion regardless.

Well, it's a fundamental technique in statistical genetics, so the paper doesn't explain it in depth. MR uses the technique of instrumental variables. [1] looks like a good explanation.

The basic idea is to identify a variable Z (genotype, in the case of MR) that causally impacts your outcome Y (COVID, in this case) only through your exposure X (vitamin D levels). If you can find and measure such a variable Z (which meets a few other conditions [2]), then you can robustly identify whether X has a causal impact on Y only from observational data.

[1] https://www.youtube.com/watch?v=NLgB2WGGKUw

[2] https://en.wikipedia.org/wiki/Instrumental_variables_estimat...

> Yeah, this would make sense if you ignored all dietary and solar variations in Vitamin D productions

No, actually, MR specifically accounts for these variations.

However, one idea (one that would need a lot of evidence in my opinion, but to give you a clearer picture) that is not captured by MR is horizontal pleiotropy. So, if the genetic variants used in this study e.g. actually cause someone to seek out dietary modifications/solar augmentation of vitamin D, that would not be accounted for here.

I don't think the biology in this case allows for that to any significant degree, but I'm not an expert in vitamin D metabolism/action or COVID.

> I don't think it's an honest assessment

The title here is maybe a bit simplified, but the paper is quite honest about its weaknesses if you read it in its entirety.

Good question!

A popular motif (and perhaps the most illustrative example) is to restrict the instrument to be genotype variants within a specific gene, and use its protein product (for example, its serum concentration) as the exposure [1]. In this way, horizontal pleiotropy is significantly mitigated.

This is sort of a natural extension (on a statistical methodological level, really a simplification) of a popular technique where a drug which has a known protein target is used in a "factorial MR" scheme [2]. Unfortunately, I'm less familiar with the literature here because it's very domain-specific.

[1] doi:10.2337/dc18-2444

[2] doi:10.1016/j.jacc.2015.02.020

Edit: removed one of the references because it's a somewhat convoluted example, and on a deeper read, wasn't actually the paper I was thinking of.

Mendelian randomization (MR; for the interested layman - the term for instrumental variables techniques when the instrument is genotype) is a well-founded research paradigm with several known potential flaws which I discussed in more depth in reply to another comment below. But, fundamentally I agree with your assessment that the study does not definitively answer the question. More generally, the articles I referenced in the other comment may be of interest to you as well - specifically around the reliability of MR studies in general.

[1] doi:10.1038/s41588-018-0099-7

[2] doi:10.1093/hmg/ddy163

Horizontal pleiotropy is always a concern when analyzing MR studies, and I agree that this one does disappointingly little to preempt this criticism. I also agree with your point that causal loops aren't a concern here - vitamin D deficiency was measured pre-COVID, so there's no reason to think that COVID influenced that measure.

I generally question Mendelian randomization (MR) results by default as well, but I do think there is good MR research out there. Some positive things to look for (IMO): attempts to quantify (or at least discuss qualitatively) horizontal pleiotropy (e.g. by looking at known human protein-protein interactions), attempts to mitigate the effects of horizontal pleiotropy through methodology (rare, not sure if I've actually seen this in papers with more exciting results, but an active research area), and attempts to quantify causal loops / elucidate the broader causal structure.

There's also a fair bit of general research on the reliability of MR results [1, 2] which may be of interest to readers.

Additionally, there is a false intuition among people who regularly conduct association analyses but not MR studies that horizontal pleiotropy can only serve to increase the causal effect size. However, this is not true (mentioned in [2] but better elucidated mathematically in the instrumental variables literature): negative results (such as the one posted) can be induced by horizontal pleiotropy as well (of course, not saying that it was).

[1] doi:10.1038/s41588-018-0099-7

[2] doi:10.1093/hmg/ddy163

The IRS, in terms of of maximizing their tax collection, would almost certainly be better off using that money on (admittedly expensive) enforcement action against those who use traditional tax avoidance techniques. As you say, money talks.

That may or may not be true, but I feel it's worth mentioning that often those individuals need some level of security clearance. See listings on job boards for janitorial staff/etc. that require security clearance (often at private corporations, like defence contractors, but I see no reason why it wouldn't extend to public employment i.e. at the Capitol).

No, not necessarily. The objective functions used to train neural networks are generally non-convex (the nets themselves being non-convex as well), but are traditionally trained using stochastic gradient descent (and its variants).

In the context of this paper, you can think of the "gradient descent step" as optimizing the parameters of the kernel (i.e. the parameters that generate the kernel space). There are no explicitly optimality guarantees beyond those of standard gradient descent.

The "SVM step" would still find a global optima within the kernel space, but the qualifications of the previous step mean that the kernel space generated might be useless.

Antisense drugs are the same biological material, but have a very different mechanism of effect.

Specifically, mRNA vaccines require an immune response to be effective. The mRNA (vaccine) encodes a viral protein, and the host (human) immune system still has to learn to respond to that protein.

Antisense drugs are much simpler, with no immune response required. The mRNA (the drug) itself simply binds to the target mRNA.

R is nice for stats stuff, even if performance can be iffy for certain tasks. The existing libraries for basic ML tasks are largely fantastic. The syntax rubs many people (including me) the wrong way. But I don't like the paradigm for basic ML tasks in scikit-learn (python) much better.

Anything involving matrix computation is great to do in python because of existing libraries (numpy). I think most people who are actually writing code (i.e. grad students) don't mind the syntax because they learned it in school starting in undergrad.

Octave's main appeal to me (and most people I work with) is being able to run existing MATLAB code. However, making it work for nontrivial MATLAB projects, is, well, nontrivial. It's also very slow.

Julia is the language I'm most excited about. It's fast [1] even though it has a slow startup time, and that's improving quickly. It's got a nice grammar-of-graphics-based plotting library [2] that's quickly catching up to R's ggplot2. I think it has potential to be the language of choice for non-neural-network stats/ML (that's python for the foreseeable future). I like the syntax a lot, too. The data types take some getting used to, not unlike R. I also think if Julia moves away from JIT compilation, it has potential to be the language of choice for deployed NN models (right now, that's C++).

As for plotting, R's ggplot2 beats MATLAB any day of the week for me. To each his own.

[1] https://julialang.org/benchmarks/

[2] https://github.com/GiovineItalia/Gadfly.jl