My personal side-blog.

Perhaps some sort of ringoid like structure, seeing it as a category and formally adding inverses to certain maps? Being very careful not to invert too many things. It certainly can't be modeled as a one object category where you just invert some maps. But if you construct it delicately? Maybe a new object for every limit ordinal or something?

Well, I guess the structure of the ordinals in terms of limits and such is some extra structure. I’m not sure how much of that can be equipped onto the subfield you’re talking about.. probably you can get all of it if you do it right. I guess you can just equip it with a copy of the countable ordinals (which has that limit structure (/topology) on it) and the (order preserving) embedding of the countable ordinals into the field you’re speaking of.

Personally, my answer is that a number is any element of a (insert fairly broad class of kinds of structures here) that I feel like calling a number.

Perhaps some sort of ringoid like structure, seeing it as a category and formally adding inverses to certain maps? Being very careful not to invert too many things. It certainly can't be modeled as a one object category where you just invert some maps. But if you construct it delicately? Maybe a new object for every limit ordinal or something?

Small water drops on an iPhone screen break up the seemingly white light into sparkles of red, green, and blue. It reminds me of Newton using a prism to refract white sunlight into a red-green-blue spectrum, 350 years earlier. But we are really living in the anthrophocene now: when Newton did his experiment he uncovered the fundamental laws of nature, but when we scratch the surface, the foundation is something that was designed by an engineer.

I imagine in a century or two literally everything will be like this and we will spend our lives in some elaborate virtual reality setup, or as uploads. But for now you can get a preview by working as a software developer, where everything you deal with is a matter of contingent convention.

And when they are natural in some sense, it’s because of the underlying mathematics. Anyways that’s why I quit software and so math now.

More natural than a Turing machine at least, and I don’t think the concept of a Turing machine is all that arbitrary (though not as natural as an admissible numbering).

The lambda calculus also seems relatively natural?

Now, if you mean things like “learning to use some particular software library/programming language”, sure, these have a lot of arbitrary things to them, but this seems like, dealing with the details of how some concept is defined in one work vs the (almost certainly equivalent, but proving it would take some work) way the same* concept is defined in another context.

One idea is, for each source, we choose a (random?) bitstring (long enough that each source gets a different one), and then like, for each parameter of the model, we only update it while training if (some particular bit of the bitstring is (0 or 1)), where different parameters would vary as to which bit they cared about and whether they were frozen if 1 or frozen if 0. I imagine this would both cost too much during training, and I’m not convinced that it would even really allow one to determine the source of the information afterwards. Because, like, just because certain nodes get more activation than normal during the response to some question, does that tell us which parameters played a role in encoding the information retrieved? Would we be able to estimate the bit strings from this? Seems hardly clear. Especially because, the information could have been reinforced through many different sources, and taking a weighted sum of their bit strings seems like it would just give us useless mud.

Another idea, which I also think would be much too expensive, would be, instead of training the model just to output logits, maybe we could train it to output like, a tensor product of the logits with a vector encoding the source? Of course, that would increase the dimensionality way too much. Also, not sure why I’m thinking the tensor product. That probably doesn’t make a lot of sense..

The proof of your being principled isn't in what you say about your enemies.

Some ideas of “what would be a good choice” are either too vague to apply everywhere, or just aren’t actually a good idea to apply everywhere, but where the cases in which they are/aren’t a good thing to apply would be too hard to explicitly specify. But, these things can still be useful in guiding action, even though they don’t serve quite the same purposes as principles as you describe here.

I guess “heuristics” would be the term for these? And, when speaking about morality (or ethics), “moral (or ethical) heuristics” might be a decent term for these?

I suppose your original point still applies to these moral heuristics to some extent though. It may be less critical, because they are only heuristic, but, to crumple completely at any pressure does seem like it would negate the value of such heuristics as it would for the principles.

Still, I feel as though a formulation of your point when extended to apply to heuristics rather than to strict principles, would benefit from some kind of caveat, but I’m not sure how to formulate it. Maybe I’m wrong and it doesn’t need such a caveat?

Hm. I guess moral heuristics that aren’t strict principles are perhaps most valuable when considering one’s own (potential) actions, or the actions of someone asking for advice. It seems like there’s maybe not so much of a point in pointing out that an adversary’s behavior goes against a moral heuristic that isn’t exactly a clear/strict rule?

A number of good thoughts here. A nitpick: spreadsheet software very much does support random number generation, and as such I don’t think that’s a good basis to make a distinction. However, effective randomness of “reward” might be. I’ve also thought things that might be called UBL might be good. Perhaps e-ink displays could be helpful for some of this?

Good point, I should have assumed that Excel/etc would have PRNG, although I suppose if a society were implementing something like this, a spreadsheet suite without one would appear pretty quickly.

"Reward" is more subjective, though. CAD software would be another issue. Overall, I think it's undesirable to have a mandatory software classification board, because then that would be used to choke out political dissent. (Dissent against who? Over what? Whoever's in charge at the time, over whatever issue is inconvenient for them.)

The benefit to my mind of e-ink isn’t it being monochrome, but- well, I remember when I was a child kind of resenting the TV because it drew me to look at it even when I disliked the show that was on, such that I would avoid the living room while it was on - I think refreshing backlit screens, perhaps like fire, have an effect of drawing the eyes, which e-ink lacks, even in color.

Hmn. Now that you mention it, I do have a habit of turning off televisions that I'm not watching if I can get away with it, although that's more to do with sound, rather than light.

Interesting.

More conceptual sketching follows.

So, my way of thinking was that, rather than a singular device, what we would have would be machines composed of multiple, dedicated hardware modules.

For example, there might be an hourly Internet-queue module. An email application would write to the queue to gather emails during the next update cycle. These would then be loaded into some kind of buffer.

Alternatively, the machine might have a high-powered GPU used to render video games... with a hardware timer limiting its use to a few hours per day, possibly with rollover for the week, or limiting use based on how often it's called up.

Alternatively, the machine and its OS might use the hardware to help it track the generation of interaction frames and the number of frames used within some time period.

The idea is that software is too flexible. Borrowing from the diet idea of not buying junk food at the store, with a general purpose computer, it would be too easy, in a moment of weakness, to download and install a version of a future open source clone of Minecraft that has not been registered as a game with software authorities.

Hardware modules, however, are more expensive and more rigid. You have to physically go somewhere if you want to get one, or order one and have it delivered some time between 8 hours and 30 days later. As such, we decompose the general purpose personal computer into parts, and then allow the user to assemble a machine with capabilities which fit their personal rhythm.

A user might then have different time allocations set for different capabilities. Maybe they have 2 hours set for real-time 3D rendering, and 8 hours set for a flat reader mode that displays files from long-term storage, or edits text.

My personal feeling, and this is very vibes-based, was that the user should then have the dignity of the device looking like a 1990s personal computer.

An e-Ink device sounds like a mass-market tablet like the Amazon Kindle Scribe, either with no keyboard, or with a tablet case keyboard.

This is going to sound weird, but an e-Ink tablet with pen sounds like the correct device for Frieren, as in the elf. On the other hand, for some reason, the idea of having an Amazon Kindle Scribe as my only computing device sounds humiliating.

I'm not sure why. I think it's something to do with not being able to call up the power of a PC, combined with seeming (in that scenario) like I'm not able to operate one?

Anyhow.

Intuitively, the consideration above calls up two concepts.

First, imagine a 38-inch e-ink display mounted to a wall, with an almond-colored bezel. It is displaying a high-resolution version of The Great Wave Off Kanagawa. It's probably some kind of elf technology.

What does it do? I'm not sure. The visual came into my mind in contrast with the necessarily small size of cathode ray tube monitors.

Sometimes, these sorts of intuitions don't go anywhere. Other times, they have a logic to them that becomes more apparent on further examination.

Maybe the center of the 38" display is placed at eye level on a desk, and the device is accessed using a keyboard and a mouse. If we consider a 39" 16:9 monitor, it's 33.86" wide and 19.29" tall. With US letter page size, that's 3.98 pages wide and 1.75 pages tall - trim off the margins, and that's about 8 simultaneous textbook pages.

This may be what the power user version of the low-speed e-ink device looks like - unlike a textbook, it isn't limited to 2 pages at a time, so it can represent a schematic with an 8-page spread. Alternatively, it could have 3 separate e-books open at once, plus a text-editor for writing. ...and as a computer, one could easily search, copy, and paste text, flip through dozens of books, and so on.

Thinking about it after writing that, and seeing that large e-ink displays have been manufactured, but that the market is apparently small... a GPU module that only does 3D rendering for games for a couple of hours a day based on hardware limitations sounds like a tax dodge, or rather a min-maxing of allowed computational power through time-shifting.

The emergence of the "turbo library" considered above might imply some sort of regulatory environment that creates the product category.

If you did catch that part, calling it a “hellscape” seems very strange to me.

Rather than proposing a single solution, this post covers some of that thinking, to provide you with concepts to chew on.

The Medieval1900/1960s/1990s/2010s/+ tech zone poll was based on a general sense regarding the speed of movement and the speed of information, and noticing that in the year 2000, "the Internet" was still a place you went to (by sitting down at your computer), and not a cyberspace layer that surrounds the planet.

But what, exactly, would "1990s/Y2K computer limits" cash out to? 800x600 pixel resolutions? 800MHz processors?

Just what was this mysterious "it factor" we would be trying to bring back by making computers slower and less advanced?

From my notes:

One way to view virtual reality is that the VR environment has extremely low weight, and is therefore extremely mutable. Any physical good, such as a sports car, can be simulated, and in almost any number, which raises a question: why buy a real sports car when we can simulate them? People are at risk of getting lost in virtual reality, with simulations becoming more satisfying to them than real life, leading them to underinvest in their real life. We can think of technology as altering the ratio of effort to environment change. High effort is required to move dirt with a shovel. Low effort is required to move the same volume of dirt with a bulldozer.

The assumption behind the zones-by-tech-level question is that beyond a certain point, except for medical technology, additional high-tech development is superfluous, because virtual reality means that nearly arbitrary sensory experiences can be generated relative to an agent's sense limits.

Demand results when the expected value of a change in the environment loops back through the agent, resulting in a potential change in behavior.

The point of video games is to produce a high stimulus feedback relative to the amount of effort. (I read that in an article on GamaSutra once years ago, and it really stuck.)

In a virtual reality environment, reward signals can become disconnected from agent well-being.

Movies, television, and books appear to be less addictive than video games and social media. What separates them? Interactivity appears to be the primary difference.

With this in mind, I was then able to work backwards and develop an intermediate regulatory concept: interaction frames.

With a book, the content is static the entire time. With a DVD, you press a button and then the DVD may play to completion. You might skip back to a previous scene, but there is no need for further interaction.

A paged website is static until you update the page. With continuous-scroll social media, new content is always being added, and at any moment you may receive a notification to get into an argument or that someone liked your post. Video games in general tend to have continuous interaction.

This model does not adequately address the situation, and, importantly, cannot distinguish between a video game and a spreadsheet program like Microsoft Excel.

We can adapt a concept from gambling: human beings seek mastery or identification of patterns, and the randomness of gambling prevents mastery of the pattern from being achieved. A spreadsheet is very much not random, while loot drops in many video games are quite random!

We could at least measure commands to produce random numbers, or require registration of pseudorandom number generators.

However, this is still insufficiently general. Simulator games may generate complex behavior from simple rules. Are we stuck regulating "game-like elements" by committee? That doesn't sound right.

Still, people do get bored of single-player games, either by mastering the gameplay, or when they get used to the story elements of the game. A visual novel is more like a spreadsheet or a book than like a slot machine.

Also, not all social media seems to be equally addictive. Why were chatrooms seemingly less dangerous than twitter?

There is an anticipation of information gain (informative tweets). There is an anticipation of exciting experiences (someone shows up to fight!). Also, from the user's perspective, it's somewhat random.

In a chatroom, there are fewer people, and they stay for longer. This means that (1) you have more information about them and their positions, (2) you have more incentive to be cordial, and (3) because there is a limit to how much they can have gotten up to while you weren't looking, there is a limit on the anticipated amount of new social information.

On Twitter, there is an endless stream of new people to argue with, and they can show up at any time. You can post about some chalupas you made yesterday, and some lunatic will show up to fight you. Thus, there is always the background anticipation of an (emotionally stimulating) attack.

In a chatroom, if someone would attack you over chalupas, you already know him as the chalupa guy. In real life, an argument is limited by space and time - the chalupa argument ends by default when the bar closes and the chalupa guy is no longer within earshot, and you cannot reopen it until you see him again.

Tumblr is slower-paced, but things like "likes" are continuous, so there is always the incentive to check in to see how your post is doing.

This allows us to get into a model based on food.

There is a dieting strategy involving not buying junk food at the store, so that it is not at your house when you get a craving for it.

This implies that the craving is a temporary impulse or peak, and that it just has to be outlasted. If the craving were uniform, the strategy wouldn't work, because the dieter would just buy the junk food at the store.

(This suggests that the baseline craving for drugs is higher, because drug addicts are willing to take much more extreme actions to feed their addictions, over a longer time period, and that the symptoms for withdrawal (a more constant negative stimulus) are worse.)

So is control in the hands of the agent, or is it in the environment? Is there choice, or not? The dieting strategy appears to split the difference - removing junk food from the cupboard is a kind of prosthetic self-control and willpower shifting. Willpower exists and can be exercised, but is limited, and reserves vary over time. Altering the environment at a point of high willpower can reduce willpower requirements in later contexts, until they are within the window of reliable feasibility.

This suggests a strategy of altering the digital environment to enable users to act on meta-preferences for prosthetic self-control.

For social media services, this suggests regulations imposing new usage modes. For example, it might be required to provide access to third-party user interfaces, which might do things like hide the number of likes. Alternatively, a user might receive all the tweets from a specific set of accounts as a daily summary. Since social media companies require revenue, this access might be a paid service based on average foregone ad revenue. It's a matter that would require a good deal more consideration.

For other items, as part of a broader social movement, we might imagine users being able to buy dedicated hardware. Attempting to control interactivity via software requires a great deal more regulation and is easier to bypass. By contrast, we might imagine a hardware module that writes to a virtual canvas at some rate. The user could then scroll the canvas without receiving updates until the next refresh.

The user could buy an appliance device with built-in limits, similar to not bringing junk food home from the store.

Regarding welfare...

If social media makes people insane, then refraining from social media is pro-social, but suffers from a coordination problem.

However, that's more speculative. More broadly, this is about the liberal concern of consent under capitalism, and what it means to consent to technology. There are two considerations, in tension.

If you have to either use a technology or be homeless, then can you really have been said to have "consented" to the technology?

On the other hand, why should everyone else be expected to subsidize some guy using a horse and buggy?

Maximization of pattern efficiency is likely to be hostile to the continued existence of the human species, not that differently from how hard drugs distort and kill people. On the other hand, insufficient pattern efficiency means reduced production, which may mean making hard moral choices and having wasteful suffering that could have been avoided.

(Regulated capitalism is actually pretty good about consent relative to production levels, compared to say, feudalism or command economies.)

Suppose a standard S for a type of device, with the following constraints on how they send/receive data on the internet:

For information that is publicly readable by everyone, an S device will by default only display it to the user if it is at least 24 hours old (Note: doesn’t imply a delay between requesting the page and receiving it

Information published from an S device to be globally readable, is made available when a certain (configurable) time of day passes (but an S device doesn’t let you use the time being configurable to make it sooner)

A user of an S device can add other S devices (or perhaps user IDs) to a whitelist, but where adding to this whitelist has a 24 hour delay, where messages from the whitelisted user/device can be displayed immediately, provided that the whitelisting is reciprocated, but there is a limit (of say, 1000? Maybe 10,000?) devices/addresses/ids that each device/id can whitelist. (This allows for chat rooms and video calls, etc., but one has to pre-emptively both decide to white-list each-other )

There are some things that I’d like for such a setup to support but which these restrictions don’t quite support. For example, it doesn’t support online matchmaking in video games, or joining an irc channel about a technical topic in order to get help on it on the same day that one learns of the technical topic.

But, such a setup should allow traditional websites without issue (though RSS feeds would be a day behind. Eh.) including like, homestarrunner (“It’s dot com”), and even like, sharing a cool video one found online in a chat room one is a part of.

It seems somewhat compatible with web forums, but not necessarily entirely?

Ah, I guess it depends how you apply the requirements. If you say the forum is on an S device or has an id, then as long as the number of active users of the forum is below the limit, then it works without issue (well, the public view of the forum would be a day behind what the actual users of the forum would see, but that’s no big deal). But if you consider the messages to be from each of the other users directly, that’s a bit of a mess. I think the former way of applying the rule probably makes more sense. Though, in that case, there could end up being like, a hierarchy of ids which could be used to circumvent the limitation on the number of real-time connections, and there would need to be some way of either disallowing that, or at least giving a user ample warning before letting them whitelist (a part of) such a thing..

oh, and, bonus : if chat rooms or forums under this paradigm were limited in the number of active real-time users they could have, they probably couldn’t really use on-the-fly targeted ads, because for users not in the real-time whitelist, they could only receive versions that have existed for at least 24 hours, and the real-time active users probably wouldn’t number enough to do real time targeted ads for? So probably the ads would have to either be semi-static or it would have to be financed in a way other than ads?

But also, hm, if the gradient descent step is reversible, then it has to preserve volume (or, I guess just cardinality, as the numbers have finite bit depth), which seems like an obstacle if we want to end up in the region where the loss is low.

I guess if we define a probability distribution in terms of the loss, as like a Boltzmann distribution type thing, then maybe some kind of, “the number of auxiliary bits spent/erased corresponding to the difference between the uniform distribution and the distribution obtained from the loss function”? I think I saw a video claiming something similar about the distribution learned (rather than a distribution over possible parameters), but the argument I saw seemed unclear.

Ah, if we treat loss as like potential energy, then if instead of plain gradient descent, gradient descent with momentum is used, in a way such that energy is conserved, then that seems like it could be nice and reversible. It is in the continuum limit at least (assuming the loss is smooth) and I would imagine it could be made to be reversible when using discrete time-steps and finite bit depth as well. In the continuous limit at least the energy would be conserved, and probably also if things are set up for the process to be reversible in the discrete case? Then, maybe bits could be spent on decreasing the kinetic energy when the potential energy (the loss) is low? Is there some optimal way to do this?

If we want out computation to start in some initial configuration which may-as-well-be random among N possible states as far as it relates to final answers we want, and we want it to end up in one of M states (with M < N), then by the generalized pigeonhole principle, at least 1 of the M desired final states is reached when starting with at least (N/M) of the starting states, and so this requires spending log_2(N/M) bits to encode the information of which starting state was used.

So, for the size M of the set of configurations that are good enough, hm…

First, what is a spectrum? A classic first definition says that a spectrum is an infinite loop space with specified deloopings. Quickly, if you don’t know, the (based) loop space is the space of continuous maps from the based circle to a based space. That is, pick a “0” on the circle and on your space X, 0 maps to 0, everything else is arbitrary up to continuity. So, loops on X that start and end at a fixed point. (Note for the experts: yes this is an omega spectrum not a general spectrum. Don’t care, I work up to weak equivalence for reasons outlined below.)

Call that loop(X), then you can repeat the construction, forming loop^2(X) and so on. Repeat infinitely (take the limit) and you get loop^infinity(X). A spectrum is the data of this whole procedure, starting from X, looping, looping again, etc.

Now, this is fine and dandy. There’s a problem though. We want tensor products! That is, a continuous product x such that Map(X x Y, Z) is equivalent to Map(X, Map(Y,Z)). This is currying to computer scientists, the hom tensor adjunction to mathematicians, and we love it dearly. It doesn’t exist for spectra as defined above.

And the reason it doesn’t exist is fucking stupid. There’s only one candidate definition, and it fails to be commutative (or associative I think) because of a fucking sign issue. Maps between equal dimension sphere have degree, like spooling a thread around a circle, forwards or backwards. And commuting this tensor product flips degree and everything breaks down. This is stupid.

Now, there is a fix to this, and it’s a rather complicated way of saying “fuck sign issues” and it works out, it’s called orthogonal spectra. It’s obnoxious though. Actually it’s so bad that there’s a newer fix, symmetric spectra, which are also obnoxious.

And, conceptually, I’ll argue it’s silly.

What’s the point of spectra? Spectra represent (co)homology theories. I won’t explain what a (co)homology theory is, but examples include singular (co)homology, topological K theory, stable homotopy, and cobordism. By represent I mean, interpreting square brackets as “maps up to homotopy”, a space X, a spectrum E with deloopings E_n, and a cohomology theory H^n, [X, E_n] = H^n(X). Deloopings as defined above, like, Loop(E_n)=E_{n+1}. Every cohomology theory is represented by a spectrum this way, and every spectrum determines a (co)homology theory.

But this only depends on the spectrum up to homotopy, and maps up to homotopy! The whole point of spectra is something that only matters up to homotopy. Which makes perfect sense, because we’re motivated by the homology of chain complexes in our definitions, and that is itself an up to (chain-)homotopy concept! And, up to homotopy, those sign issues with defining tensor products don’t exist. Because sign issues are stupid and shouldn’t exist anyways.

This wouldn’t happen to be anything like that, would it?

I guess not, as I guess what you described is more like, (A \otimes B) \otimes C = - A \otimes (B \otimes C) ?