mjn

I've also started doing this, and it's surprisingly enjoyable to both do and even to read. The end result is often more readable to me than using a 3rd-party JS visualization library, because I only need to know standard HTML/CSS concepts to understand what's going on. And a side benefit is smaller pages with less bitrot due to being able to skip the dependencies.

Besides growth, Nvidia also has much higher profit margins. In Q1 2024, Apple's margins were 42% gross, 26% net, versus Nvidia's 78% gross, 57% net. May or may not be sustainable, but for now Nvidia's basically printing money.

In practice, wrong findings that aren't due to misconduct and aren't very recent are usually not retracted though. It's just considered part of the history of science that some old papers have proofs or results now known to be false. It is pretty common in mathematics, for example, for people to discover (and publish) errors they found in old proofs, without the journal going back and retracting the old proof. A famous example is Hilbert's (incorrect) sketch of a proof for the continuum hypothesis [1].

[1] https://mathoverflow.net/questions/272028/hilberts-alleged-p...

I think the flags have more been because of the poor sourcing of the actual death. It seems various people and groups heard that he died and are posting their obituaries and remembrances, but there's no real factual confirmation or details (e.g. one thing that was briefly on the front page here [1] was a link to an unsourced, now-deleted tweet from Momentum UK).

[1] https://news.ycombinator.com/item?id=40721271

Yes, this article doesn't appear to have any independent sourcing, it just cites the Jacobin one that you link: "de acordo com a revista norte-americana Jacobin" ("according to the North American magazine Jacobin").

"Unfit for deployment" or "not intended for deployment" is semi-standard wording for research models that are just raw language models with none of the safety/bias/offensiveness filtering that is usually desired for product applications. For example, if you deploy it as a customer-service chatbot, it might tell your customers to kill themselves, or call them racial slurs.

It doesn't mean that there's anything technically wrong with the language model per se as a model of language, just that there has been no effort made to ensure it's fit to be deployed as-is for any given generative-AI use case, and the model authors would prefer you didn't do that.

I submitted the Reuters article as a summary, but here are some primary sources as well:

FinCEN's press release: https://www.fincen.gov/news/news-releases/fincen-proposes-ne...

The actual proposed rule: https://www.federalregister.gov/documents/2023/10/23/2023-23...

> DeepL and Google Translate definitely also make things up

I think what they make up is different, but this is a good point. They have a particularly odd tendency to either do something like autocorrect where it wasn't appropriate (translate a different word that is similar in spelling to the requested word), or to make up false friends, doing something like transliterate + then autocorrect in the target language.

One example, which I blogged about 5 years ago but is still mistranslated, is the word "ribbit" (what a frog does): https://www.kmjn.org/notes/google_translates_ribbit.html

In 2018, if you translated it to Greek with Google Translate, it gave you κουνέλι (kouneli), which is Greek for rabbit. A word that is one letter away from ribbit but not close to a similar meaning. When I tried it just now, it translates it to ραβδί (rabdi), which means stick and is completely unrelated to the correct answer, but I guess starts with similar letters as ribbit?

Same situation in Greece. Formal documents like identity cards, university diplomas, and even mailboxes, generally put the surname first, while in running text, like a news article, the given name is first. Some formal documents also list the patronymic, in the order: Surname Given Patronymic.

In some situations it is less clear which convention applies. For example most of my older relatives use surname-first for their Facebook account name, which is less common with younger people. I have also seen both orders on business cards. I think most Greeks would not expect foreigners to be familiar with or follow these conventions though.

Retirement savings are in fact excluded from parents' assets for the purposes of FAFSA, at least if they are in some kind of account legally recognized as a retirement account (401k, 403b, IRA, traditional pension, annuity, etc.). Details of what is included and excluded from parental assets: https://studentaid.gov/2324/help/parent-investments

There are a few others as well. As of the 2023-2024 cycle, there are 13 ways to be considered independent (any one of them suffices), according to the official list [1]. To summarize them here:

1. Born before Jan 1, 2000; 2. Married; 3. Will be attending grad school [2]; 4. Active-duty U.S. military; 5. U.S. military veteran; 6. Have dependent children; 7. Have live-in dependents other than children; 8. Both parents deceased and/or were a foster child or a ward of the court after age 13; 9. Current or former emancipated minor; 10. Legal guardian is someone other than parent or stepparent; 11,12,13. Homeless or at-risk-of-homelessness within the past year (three different routes of determination).

[1] https://studentaid.gov/2324/help/need-parent-info

[2] Graduate students in the U.S. generally don't get the kind of need-based financial aid that undergrads do, though, so this one is often not very helpful.

That matches my experience. Of places I've spent significant time in the past year, the one with the most reasonably priced food for eating out is NYC, of all places. It has very expensive places too, of course, but a lot more sub-$10 options than anywhere else I've been recently. There are even a few 99-cent pizza slice places still in business (not as many as 10 years ago, but they exist, and $2-3/slice places are abundant).

That's a good point. I wonder how many of those '70s theorists of technology would have predicted that, 50 years later, UPS driver and plumber would be (seemingly) harder to automate than knowledge work. Not really to blame them; it's only obvious in retrospect. But I agree we may need to rethink some of the established analysis that was based on different assumptions.

Some of the alliances are more surprising to me than others. I'm not surprised that religious social conservatives eventually discovered that they have things in common with other religious social conservatives, even from other traditions. If your main issue is that you like the traditional patriarchal family and don't like homosexuality, it's not a huge stretch to find others who agree.

I'm more surprised by the movement in this direction of people like Tommy Robinson, former leader of the English Defence League (discussed in the article intro). I thought of this kind of far-right as more of a "clash of civilizations" type of cultural and/or racialist right-wing, defending "the West" against foreigners and immigration. They did not historically have much of a religious angle to them, nor particularly care about social conservatism (other than opposing immigration). It feels like a genuine change that people in that camp now find common cause with Islamism against a shared enemy of Western liberals.

It's a bit more than six courses a year at community colleges and teaching-oriented 4-year colleges (though you're still right that adjuncts are much cheaper). In California, most full-time faculty at community colleges have a 15 credit per semester teaching obligation, which works out to ten 3-credit courses a year. Some do get credit in lieu of a course for taking on other responsibilities (dept chair, director of a program, etc.), but that's the baseline workload. Even the Cal State system, which is made up of 4-year schools, has a baseline teaching load of eight classes a year ("4/4").

This appears to be the new policy: https://twitter.com/TwitterSafety/status/1603165959669354496

Chase is also more restricted in how the investment banking side of the business can use funds from the retail banking side, precisely because banks in the 1920s did play fast and loose with investing customer deposits, which caused a bunch of depositors to lose savings in the 1929 crash and associated bank runs.

The amount of interaction between the two sides that's allowed has increased and decreased over the years, but it hasn't been unrestricted since the Glass-Steagall Act of 1933: https://en.wikipedia.org/wiki/Separation_of_investment_and_r...

This is the reason UBI was historically popular among (some) libertarians, even though most current proponents in the U.S. probably come from some variety of left-liberalism or market socialism. Milton Friedman supported a version of UBI as a replacement for the welfare state, for example. His view was that if you think it's the role of the state to provide some minimum floor for standard of living, fine. But instead of trying to administer a dozen different programs with complex eligibility requirements and a whole bureaucracy behind them (food stamps, public housing, housing vouchers, Medicaid, etc.), it would be more efficient, less economically distorting, and less invasive into the lives of the recipients, to just pick whatever level of spending you're willing to put towards this goal, and give it out as unrestricted cash.

In the general case you're right, it's equivalent to the halting problem. The outline of the proof by reduction: set up two communicating processes in a way that will deadlock iff a particular loop in one process fails to terminate. So if you had a deadlock detector for arbitrary communicating processes, you could use turn it into a termination detector for arbitrary loops.

> If you're using a dynamic language, then _by definition_ the language will not enforce your static hints at runtime.

I don't think this follows by definition. It's common not to check the static hints at runtime, but a dynamic language could choose to treat them as contracts or assertions. Over in Lisp-land, SBCL will enforce static type declarations at runtime [1], contrary to most other Lisp compilers, at least under the default compilation settings (you can force it to skip the checks by compiling with a low "safety" level). If the compiler can prove that a static type declaration always holds, it will omit the runtime check; otherwise it will compile it into a runtime assertion.

[1] http://www.sbcl.org/manual/#Declarations-as-Assertions

I read that as being about what language industry uses to write ML applications, not about technical feasibility of integrating machine learning methods into a codebase. Put differently: industry most often uses Python (especially in ML), therefore the author wants to target Python in order to maximize uptake outside of academia. They even admit that doing it in Python is technically harder than doing it in Julia ("Unfortunately, this type [Type III] is also the hardest to get right"), but consider it worth the trouble for the broader accessibility.

(That's more or less the direction I've been going with research code lately too, so I can sympathize, although I'm not entirely happy with the situation and definitely also sympathize with the Julia folks being unhappy about it.)

There's an earlier blog post by the same author where they discuss three possible ways of moving away from the Fortran/C status quo towards higher-level models. They mention Julia as one of the routes, but not the one they decided to take: https://dionhaefner.github.io/2021/04/higher-level-geophysic...

I believe this does work for German (or at least I can't think of an example where it doesn't). But a case I can think of where it doesn't work is with standard modern Greek. In all-caps words, accents are omitted, while at least Python's implementation of casefold produces non-equal strings for the all-caps and lowercase versions:

    >>> str.casefold("παράδειγμα")
    'παράδειγμα'

    >>> str.casefold("ΠΑΡΑΔΕΙΓΜΑ")
    'παραδειγμα'

It's at least consistent with this, though:

    >>> str.upper("παράδειγμα")
    'ΠΑΡΆΔΕΙΓΜΑ'

But I'd consider that incorrect, or at least nonstandard. The Greek alphabet does have accented versions of capital letters, but they can only be used as the first letter of a word in mixed case (e.g. if a sentence starts with έλα, you write it Έλα), never in the middle of a capitalized word. However maybe this slides too far to the "language" rather than "encoding" side of the space that Unicode considers outside of its purview.

Even when they don't have such a ledger, it's not clear that blockchains do much to solve that problem either. The hard part is not the ledger, but the agreement on who owns what, with detailed surveys that are agreed to demarcate "ownership" and therefore enforced by law.

For example: Greece is still in the process of trying to centrally account for ownership of land [1]. There are certain regions where that was accomplished many years ago (e.g. some of the former Italian possessions in the Ionian, where the Italian-era surveys were accepted by the modern Greek state). But in many regions, property boundaries are still informally demarcated and based on oral tradition, which not everyone agrees on. The difficult part is that aspect of it: agreeing on who owns what, and precisely where the boundaries are. Once everyone agrees on the results of surveys and demarcations, actually tracking the result in a ledger is comparatively easy, as evidenced by the fact that nobody has trouble retrieving 19th-century Italian property ledgers.

[1] https://int.ert.gr/greek-land-cadastre-will-be-complete-by-f...

It really varies. Normally the author themselves has to come up with the money somehow though. At least where I've been, you can't just bill it to the university. If you have research grants, you can often pay out of those (this is common in the natural sciences). If you have some kind of general research budget you could pay out of that (e.g. I get $2k/yr unrestricted research money from my university... but I'm definitely not going to spend it on an APC!). Some universities do have an internal OA fund you can apply to, but those are often limited. For example, MIT will cover up to $1k per article for papers written by its faculty/staff/students: https://libraries.mit.edu/scholarly/publishing/oa-publishing...

I'm normally skeptical of journals funded by article processing charges (APCs), but this policy is pretty good:

> no author of an accepted paper will be denied publication due to lack of funds

Most APC-charging journals (like the Nature family) have a weaker waiver policy, where they only promise waivers to authors from developing countries, and say they may consider others "on a case-by-case basis" but with no guarantees.

You can break out of the network effects with some effort. One of the top journals in machine learning nowadays is a community-run open-access journal that does not charge authors, JMLR (https://www.jmlr.org/). There's a bit about its model here: https://blogs.harvard.edu/pamphlet/2012/03/06/an-efficient-j... (previously discussed: https://news.ycombinator.com/item?id=15280736)

On that practical level, public libraries in some of the places with the most need for public computer/internet services also, unfortunately, end up rationing them pretty severely. For example, the Los Angeles library's computer use policies are [1]: 15 minute sessions for walk-up computers (no card needed); 1 hour sessions if reserved in advance and you have a library card; 2 hours/day total limit for any combination.

I know this in part because I'm a middle-class CS prof whose stuff is all "on the internet". At one point I had the great idea that I don't need to carry around a laptop, since my stuff is all on the internet anyway. I'll just use public libraries, or other university libraries, problem solved. Sometimes it works, but there are obstacles strewn all about. On a side note, it's also made me appreciate physical books in libraries more. If you really did want to learn something and had $0 to spend, and you can find a physical book on the shelf of a public library, you can read it for as long as the library is open, without a card or 15-minute time limit.

[1] https://www.lapl.org/about-lapl/internet-use-policy

(Edit: I realize the original post was going for something more philosophical, and I see they commented in a sibling comment that they're sorry if they came across as tone deaf. I don't mean to pile on! But wanted to point out that just the bare practical aspect can surprisingly difficult even if you know what you're doing.)

To be honest this is probably in the top 3 reasons why I haven't investigated leaving academia for industry with any seriousness. A private office with a door and windows is very nice to have, and I don't think I'd want to downgrade. (Would I consider if the industry job were 100% WFH? Maybe... only recently started hearing about that as a feasible option.)

It used to be pretty common for engineers to get at least close to "private office". My dad for decades had a nice office with a window and view on the 30-something floor of the Amoco Building (now the Aon Center), which he shared with one other person. Amoco later moved them to cubicles in a suburban office park.

The Xinhua article on it (which I think is fairly authoritative) is clearer that it applies only to online games: http://www.news.cn/english/2021-08/30/c_1310157673.htm

Specifically,

> Online game providers can only offer one-hour services to minors from 8 p.m. to 9 p.m. on Fridays, Saturdays and Sundays, as well as on official holidays, according to the document made public on Monday.

> [...]

> The official said that parents and minors can decide on by themselves how long the children will play other games that are conducive to minors' growth, except online games.