harshreality

It doesn't magically mean anything, but you're comparing a case where there's a plausible argument for fair use (no commercial impact from the training set sharing, not done for profit, transformative), against... what specific instance are you talking about with NV or Samsung? Theft of trade secrets? Samsung leaking their own IP to chatgpt, through stupidity, isn't illegal.

This doesn't even affect the major AI companies that are doing training, does it? Don't you think Google, OpenAI, Anthropic all have their own datasets by now? If they wanted to use bulk content from pirate ebook and academic paper libraries, they would've already mirrored most of those sites a long time ago.

I understand your moral position. But you must realize that's not what copyright does. That's wishful thinking, the sort often manifest in modern written works on the copyright page where the publisher helpfully writes, "The moral right of the author has been asserted". Begging the question: What moral right, other than the legal copyright that was already asserted separately?

Note the irony, too: It's the publisher writing about morality on behalf of the author. It's the publisher that would give consent for any AI-training use in your concept of the proper legal and moral order of things. The author probably doesn't know much about copyright law, but wants the work monetized as much as possible, because they like having a house and eating food, and sure, creating more content sometimes. To that end, they've assigned copyright to their publisher. Or independent creators would have to create a union and assign licensing power to the union leadership. But again, only licensing to big tech, because nobody else could afford it.

Also see the ETA: paragraph in GP, regarding the difficulty of judging whether simple works are copyrightable, or whether complex works have actually been reproduced in a copyright-violating way.

This feels like an argument over angels on the head of a pin. Look at the markets. They're all-in on AI. Big tech (and big startup) AI models will move forward, no matter what licensing agreements between publishers and AI behemoths, or copyright exceptions, end up being necessary. The only question is whether others who don't have that power or money should be able to try to train their own models (admittedly probably much lower-parameter, because they don't have datacenters full of H100s) on datasets that are at least in the same ballpark.

If you're complaining about the training, you're confusing the entity doing the sharing in this case, with the entities doing the training.

ETA: If you're complaining about use of the model, this becomes a dispute over how close to an original work an output can get before it becomes a copyright violation. That seems unresolvable. Look at court cases where musicians sue over short sequences of notes. Nobody knows where the threshold should be for that, for poetry, for flash fiction, for academic papers, for novels, for paintings, or for voice samples. Courts just go with whatever they think feels right in a particular case. That is an untenable legal situation in a world with AI models like these.

Also, to reiterate, Google has its own fulltext, license-free corpus of books, which it undoubtedly has used to train models. What's your view on that? Should Google be the only entity allowed to train on large books datasets because it happened to get one through a loophole of collaborating with libraries to OCR their collections?

What do you think the market is for people who want to read books or articles condensed into lines of text suitable for model training?

I think it's 0. Maybe a handful if pirate libraries of readable ebooks and papers didn't already exist, but even a handful of copyright violators wouldn't be a serious commercial threat to the publishers.

So far, governments have been unable to do anything about pirate libraries. Complaining about AI datasets, poorly formatted for human consumption, seems misplaced.

Criticizing this as "stealing" or "piracy" is a vote for a future where only big tech, or the major publishers themselves, can train models on good datasets. Nobody else will have the money or market power to license that much material.

Scene releases are in human-consumable formats.

Books as lines of text are not human-consumable (by reasonable humans).

There's a strong case for sharing this being fair use, even if distributing hundreds of thousands of ebooks would ordinarily violate copyright.

I don't know what OpenAI's training corpus is or how they got it, but in case anyone has forgotten, Google has its own license-free collection of scanned books. It had already run that through OCR (and probably periodically re-OCRs) to generate its google books indexes, so there is no copyright argument against Google's book-trained AI models (assuming they train on books).

An argument that Books4 is piracy is an argument for an oligopoly on good AI models.

Julia, R, Mathematica, Matlab, Fortran, Smalltalk, Lua...

Julia being a mostly functional non-systems-programming language, where memory addressing is usually in the background, I think you'd find you rarely use numeric-literal indexing and you were getting upset over nothing.

The typical looping over i from 1 to length(a) is not idiomatic in Julia, where looping over eachindex() or pairs() abstracts away any concern over index type (OffsetArray can have an arbitrary first index). There's firstindex() and lastindex(); bracket indexing syntax has begin and end special keywords. Multidimensional arrays have axes().

"But what about modular arithmetic to calculate 1-based indexes?"

Julia has mod1(), fld1(), etc., which also provide a hint to anyone reading the code that it might be computing things to be used as index-like objects (rather than pure arithmetic or offset computation).

The strict way to handle indexing, I suppose, would be to have distinct 1-indexed (e.g. Ind1) and 0-indexed (e.g. Ind0) Int-based types, and refuse to accept regular Ints as indexes. That would eliminate ambiguity, at the cost of Int-to-Ind[01] conversions all over the place (if one insists on using lots of numeric literals as indexes). And how would one abstract away modular arithmetic differences between the Ind0 and Ind1 types? By redefining regular mod and div for Ind1? That would be hella confusing, so that abstraction probably isn't viable.

Reusability, good maneuverability, range, loiter time.

Something to close the gap to dispense cheaper munitions than long-range guided missiles. And to do so with a dramatically faster OODA loop.

Flying sausages with rocket motors and stubby wings don't do very well by most metrics.

Highly maneuverable reusable aircraft (even drones) are expensive, and a scarcer resource than missiles.

In modern great power military conflict, if you get to the point where you can fly F-16s or a hypermaneuverable AI dogfighting drone into visual range against an opponent [with inferior planes] who's out of missiles, haven't you already won?

If the opponent isn't inferior in dogfights, then instead of using ~$1mil munitions as expendables, you're using planes (and possibly pilots) as expendables. That's not a good or economical trade-off when missiles are as good as they are today.

A sea of low-quality content isn't a problem if the training process incorporates a way to estimate quality and eventually train the NN out of low-quality output.

There are many areas where that estimation might be difficult without human intervention: arts of all kinds.

However, in a more fact-based area like programming, a NN training routine could be modified so that what a LLM thinks it's learning can be tested. That could apply to CS and math, at least.

Eventually, NNs integrated with robotics could learn physical sciences without human intervention or pre-selected training material, and would still converge to superhuman knowledge even starting from random ideas. Giving a NN access to robotics without a human in the loop is probably too dangerous, though... it might decide to unsafely test its understanding of nuclear reactions, or the r-number of ebola...

In theory, maybe. In practice, no.

In theory, a publisher could sell you an ebook, declare that you own it and can transfer it (like a digital right of first sale) as long as you don't keep a copy.

In practice, it would be a hassle for libraries to verify that permission, so even if you had such a book, they might say thanks but no thanks. Essentially all ebooks are sold under license that each book is for your use only, and no transfers are allowed. Because ebooks are licensed, and not sold complete with a right of first sale, retailers can set whatever conditions they want depending on who they sell to, and the market can't arbitrage. That's why libraries have to pay more for an ebook they're allowed to lend 10 times than you have to pay for the same ebook on Amazon.

Cost per flight hour is still much higher than its slightly less capable competitors.

I don't know what the argument is for that. Maybe that sim training will be a larger percentage of training in the future, and therefore operating costs don't matter so much? Take whatever jet has the best capabilities, period?

I think what that misses is that maintenance of real aircraft will atrophy without constant pilot feedback. Of course mechanics can follow the maintenance guidelines, but so much of that, historically, is guided, modified, and improved by experience from wear and tear from actual use.

Aren't they comparing high-supersonic to supersonic to subsonic?

I thought the point was that aerodynamics change from one domain to the next as shockwaves cause flow separation or eddies on or behind surfaces.

If you're playing the game as stated, and if the host opens a door revealing a goat, you must switch for the best odds.

The offer to switch doors is part of the problem statement. You can't justify your decision based on the possibility in the real game that you might not get the offer to switch (which would mean that when you do it might be an intentional red herring).

Even if there's a 50% chance there are all goats, or a 75% chance, you still switch. That reduces the overall probability of a win, but it doesn't affect the relative probability from switching. The only reason you might not want to switch is if the first choice had been forced, such that there was an above-random chance that your first pick was the car.

This is not surveillance. This is content shaping and censorship.

Pop-Tarts™ are junk food, though.

For truly realistic CSAM, there would probably need to be some quasi-illegal images in the training set.

But "CSAM" also includes "character looks under 18" even if they look fully developed, which an AI model could do without training on actual CSAM.

Whether that makes the model illegal is a separate question. Is a LLM illegal if a clever prompt can get it to output a copyrighted poem? Is a human illegal if they can draw "CSAM"?

They (other than the cops investigating actual crimes against actual children) don't care about the children. They care about control. AI "CSAM" in particular (combined with deepfakes of adults made without the model's permission) is an emotional justification for more control in the form of draconian internet censorship and other preemptive measures.

Decades of megacorporation propaganda have paid off, and a fair portion of gen-z/y/x is hypervigilant about use without permission, even if the use is transformative and doesn't displace the original work. Their misguided concerns will play into the hands of the politicians seeking more control.

And it will do very little to prevent the spread of AI "CSAM"[1]. What are they going to do, regulate GPUs or any AI training software?

[1] Or any of the other concerns of today, like voice cloning or mimicry of personal likeness, like what just happened with Taylor Swift. The cat's out of the bag for images; OSS voice cloning is roughly at elevenlabs' quality, and it'll be surprising if there aren't, very soon, easy-to-use interfaces for voice cloning and tts.

All the drugs you can consume?

That's not going to be compatible with a livable pop-up city.

Perhaps if the list of available drugs were carefully selected, it could work, but not if you include alcohol, meth, and opiate/opioid derivatives, and worse. Reducing the list of available drugs would reduce the desirability of such a city for those who'd be better off in it, leading to continued hard-drug-fueled crime in society at large.

Sure. That doesn't mean an author would want to maintain MathML inside an ebook.

a) What epub editor software allow direct editing of MathML, and is it a good idea to depend on such software being available in order to edit the book without tinkering directly with the MathML by hand?

b) Does MathML support everything mathjax does?

> We should also remember how a foolish and willful ignorance of the superpower of rewards caused Soviet communists to get their final result, as described by one employee: “They pretend to pay us, and we pretend to work.” Perhaps the most important rule in management is “Get the incentives right.”

-- Charlie Munger, Poor Charlie's Almanack ch. 11, The Psychology of Human Misjudgment (principle #1)

I think precisely dictating layout is the wrong objective, although some areas (legal documents, academic papers) are still obsessed with that. Arxiv recently started offering a subset of papers in html, which is a short step from epub.

If quick high-resolution referencing (page x, yth paragraph, zth line) is necessary, I think the way to handle it is to reference a phrase on that line ("the cat ran"...), which the reader can search for. If the search interface is lacking, that's an epub reader failure, not a format failure. Or, if the search option is considered insufficient (it does require typing with a [virtual] keyboard), paragraphs can be numbered—as many ancient works or works in translation already are, because such works have many editions and can't be layout-exact copies of each other.

If paragraph referencing is necessary, visibly styling the paragraphs with numbers helps dramatically. There's no reason it has to be exclusive to high-profile ancient or classical works, like Plato [1].

Classic poetry and plays, where referencing needs to be most exact and fast, already tend to give up any hope of everyone using the same edition, and simply avoid flowed text and then number the lines.

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

Any decent publishing or html editing tools fully support utf-8 by now. It's not the tools.

Publisher and editor laziness may be a reason to be cautious about epubs currently for niche or esoteric works, but that's not the same thing.

> I bought a book set in the late Middle Ages which managed to transcribe all “þ” as “p”. Until publishers care...

The book market these days makes it challenging to do high-quality editing up front for republishing niche books in a new format. Publishers try to cut corners, outsourcing epub conversions to people who don't care and don't know what they're doing, or they OCR it, have an in-house editor (who also doesn't have a personal affinity to the subject) give it a once-over (maybe), and release it.

Think about what javascript exclusion means, and all the things a good universal ebook format needs to support. Nicely rendered math? Currently the best option to do that is embedded mathjax (maybe you could pre-build mathml and ship that, but I'm not sure that covers all cases). Graphs or charts? There are nice js libraries for that, while doing it manually means exporting images or svgs. Even static svgs are annoying and brittle to font-size changes without javascript to adjust the svg size appropriately.

Don't confuse what's necessary for standard fiction books with what the format should support.

JS and interactivity are fine, in technical books, reports, or niche fiction.

What I absolutely agree on is that epubs don't need is networking. Resources on the internet get stale after years or decades anyway, so inclusion of any network assets into an epub guarantees that the work will degrade over the years. References can be web links, but nothing from the internet should be embedded.

What problem is not solved by various DNS and local-addressing options, but is solvable by some other aliasing scheme that isn't DNS under another name, and can you outline how a better aliasing scheme would work?

Background and current options, as I'm aware of them:

"IPv6 addresses are too difficult" objectors don't seem to want a solution to the ease-of-use problem that leverages aliases, because DNS is that. They want the ease of typing in IPv4 addresses without any overhead of setting up alias mappings or trusting local hosts' ideas of their own names. That's not possible, in general, with an addressing scheme that's 4x as big. The standard representation has tried to improve things with hex instead of decimal and collapsing the longest run of 0-fields. It could have used a higher base than 16 [1], but that would slow visual spot-recognition of an address. It could have dispensed with all the ':' chars, but that would prevent collapsing addresses and make chunking more difficult.

DNS as an aliasing scheme has many administrative forms. It can be configured per-machine with /etc/hosts (copied around manually, or with scripted or configuration management tools); on a trusted local network with mdns and ip autoassignment, if the network is trustworthy enough; or manually with a full-fledged DNS server, either private or public. More complex network environments leverage tooling to make DNS as painless as possible. The mappings still have to be configured and managed somewhere, even if that's separately per host via what hostname they each think they have.

It's also possible to assign stable, short local addresses that can be remembered and typed easily roughly on par with v4 addresses (10.0.x.1 vs fc00::x:1, which becomes shorter than IPv4 for x==0).

[1] See RFC 1924. Who wouldn't love to use addresses in the form of "4)+k&C#VzJ4br>0wv%Yp" ?

Eyesight is just photon-sensing. There are primitive animals with primitive light/dark sensing.

Hearing is just vibration- and motion-sensing. It evolved from the lateral line in fish, which is not limited to "hearing" per se, but detects motion—current and transient pressure waves that might indicate predator or prey swimming nearby.

Smell is just chemical sensing. Catfish have chemoreceptors across their bodies.

A lot of intermediate forms are impossible to present to you as evidence because all we have is the fossil record which generally doesn't include soft tissue, which is where most of the interesting stuff happens.

One mutation that enables the slightest bit better sensing, provides an advantage, and offspring of that species in that area will gradually end up with that adaptation. Then it repeats. The "must have" is not dictated by some agent, but by environmental selective pressure, relative to pre-mutation versions of the same organism. It's induction, but for biology rather than math/cs.

It's quite possible that other types of photon and vibration and chemical sensing could have evolved, and would be superior, but we started down a path that locally descended to where we are now. And the existing tree of life would predate or out-compete pretty much any new multi-cellular life that tried to "experiment" and "evolve" novel sensing organs.

You want to talk about just-so stories? Cephalopod eyes attach to the optic nerve on the opposite (back) side of the retina, so they have no blind spot. Did a designer goof up on vertebrate eyes and not fix it, but instead implement a very complicated neurological fix involving saccades and "blinding" our brains to the instantaneous hole in our vision?

Liability insurance is not medical insurance.

I agree and I don't know why the myth persists that [baby] Rudin is the best way to learn analysis. It's certainly not so for a lot of people who could learn analysis perfectly well from friendlier (but not necessarily any less rigorous) sources. Rudin requires a lot of self-motivation.

In addition to Abbott which you mentioned,

"How to Think About Analysis" (Alcock), good gentle introduction for anyone who tries to dive into an analysis textbook and hits concrete.

"Real Mathematical Analysis" (Pugh), good, rigorous approach that doesn't require quite as much effort to power through as Rudin does, although some of the exercises are tough.

"Analysis I" and II (Tao), is good and starts from scratch building numbers from the Peano axioms, so there are no other unproven assumptions underlying everything later.

"Counterexamples in Analysis" (Gelbalm & Olmstead), very helpful for understanding some pathological cases that break theorems one might intuitively believe are true.

There's no reason today why someone can't grab pdfs of all of the above (even Rudin, although it's never been digitally typeset), from any friendly internet library, and use them all to build a better understanding rather than relying on one alone.

"Blood work comes back fine" is a very low bar, unless you pay for a high-end analysis (things that even many doctors wouldn't think to look at).

You didn't disclose your diet. There's a huge difference between a diet of meat (from animals raised on their natural diets, because artificial diets can change chemical composition of fats and ratios of proteins to be less healthy), berries, and nuts, and a diet of pizza, sweets and (refined-grain-based) baked goods.

It's not a "trick". We all love and hate foods based on more than their taste. The ones that taste the best are not even properly called "food"; they're distant relatives of natural foods, and they're objectively unhealthy. (Unless you're metabolically starving, in which case "bad" calories are the least-bad option.)

Trilium or Obsidian are not comparable to Outline, because they're single-user with no accounts or access control.

A lot of people have trees growing over, or very close to, power lines; during storms, high winds or ice accumulation cause many branches to break. Around here, utility crews periodically prune trees back from the wires, but a significant storm will still see hundreds of thousands without power. After a day or two they get it down to tens of thousands; fixing all of those can take many days. One downed power line can be repaired in an hour, but 1000s take a while. Some of the tail-end of repairs are more complicated, too.

Hurricanes, long-track tornadoes, and earthquakes are even worse.

It's not that technology is lacking, it's that the cost of making infrastructure resilient to very bad weather (worse than Australia sees) is too high. Most people who want to do something about it buy a generator, which solves most of their problem even though in aggregate it's terribly inefficient.