tototavros

[tototavros]

"try being a doctor for a bit to see if you enjoy it, before you invest all that money going medical school; yeah, you might get sued, but defending yourself from medical malpractice suits is part of real doctoring too"

[tototavros]

forbidden neil the rat photos

[tototavros]

BABES WAKE UP TODAY’S A NATIONAL HOLIDAY 🌈🐀

🎶NEIL IS BANGING OUT THE TUNES🎶

My (old) contribution to the beautiful day that is april 13th

Keep at it with the tunes.

[tototavros]

What if I decided that Luke-Acts was a work of history but the other gospels weren't. Just an annoying thing I could do

[tototavros]

if youre a 1200s merchant you need to be facilitating cultural exchange. If you achieve something else good. If you achieve nothing else thats okay. but you have to be facilitating trade-network mediated cultural exchange.

[tototavros]

I think what's bad about tariffs is they're like the opposite of antitrust law. They're pro-trust.

Federal Trade Commission – Market Division or Customer Allocation:

Plain agreements among competitors to divide sales territories or assign customers are almost always illegal.

If you want a competitive market for something hard to manufacture, let's say a car, it has to be a global market. That's the only way. Before we had Japanese cars threatening the US automakers, they were an oligopoly, the Big Three, and an oligopoly isn't much better than a monopoly, they don't really have much incentive to undercut each other's prices. One of them could at most capture the whole market, but they'd only sell three times as many cars. None of them are in a situation like Tesla was in the 2010s, when by investing in cheaper models and more efficient manufacturing they 10x'd their sales in <10 years. So why make the investment? So yeah I think competition is really important in keeping down prices and putting up barriers to this competition is just donating your income to corporate profits. The people who disagree with me about this on tumblr tend to say stuff like there are no competitive markets anyway, either because they naturally consolidate or because firms don't really cut prices to grab market share, but, I mean, look around you. There's lots of car companies, they do cut prices to grab market share.

However, tariffs aren't really that bad, because of foreign direct investment. In response to the threat of Japanese cars, Reagan did force a reduction in imports. But the Japanese car companies opened factories in the US South. In this way the US government addressed the threat to employment in the US car industry without seriously sacrificing consumer choice.

Foreign direct investment by Chinese companies is harder, I think mainly due to Chinese regulation, but it's happening. Consider solar panels for example, where China has outcompeted the rest of the world.

Reuters:

Construction of U.S. solar-manufacturing plants by Chinese companies is surging, putting China in position to dominate the nascent industry, as other American factories struggle to compete despite federal subsidies.

Chinese companies will have at least 20 gigawatts' worth of annual solar panel production capacity on U.S. soil within the next year, enough to serve about half the U.S. market, according to a Reuters analysis of corporate statements, government documents, and interviews with eight companies and researchers.

The group includes seven companies backed by Chinese firms including Jinko Solar, Trina Solar, JA Solar, Longi, Hounen, Runergy, and Boviet, according to the analysis.

I haven't heard of any of them actually making silicon wafers or even solar cells here, just assembling panels and modules using parts imported from China. But it's a start and it could go farther.

There has been some opposition to this trend, and I guess really I'm more worried about a government crackdown on Chinese investments than I am about tariffs on their own. I don't want us to be stuck with just First Solar charging whatever they want.

Our first clear government attack on Chinese investment in the US is the forced "sale" (nominally compensated confiscation) of TikTok.

It is important to protect property rights, not just of Americans, but of foreign owners of American capital.

[tototavros]

Another thing I presumably had the opportunity to learn in my college thermodynamics course, but instead I'm realizing it now over ten years later.

A small increment of heat is transferred; the quantity is T dS.

So consider a heat transfer from a hot system to a cold system. This is a little cutesy but I'll call the temperature of the hot system T and the temperature of the cold system t, so you can remember that T is bigger than t. The corresponding entropies are S and s. The heat lost by one is the heat gained by another, so:

- T dS = t ds

T/t = ds / (-dS)

So the entropy gained by the cold system is more than the entropy lost by the hot system:

ds / (-dS) > 1

ds > - dS

So the total change in entropy is greater than zero:

ds + dS > 0

Remember, ds is positive because the cold system received heat so its entropy went up, according to the formula t ds for the increment of heat. dS is negative, because the hot system gave heat so its entropy went down. The change in total entropy is positive, because the positive number here is larger than the negative one.

So the second law of thermodynamics, which tells us that entropy changes are nonnegative, also tells us that heat has to flow from hot to cold rather than the other way around.

That heat flows from hot to cold sounds a lot more obvious than that entropy never decreases, since it's hard to tell what entropy "really is". In fact, temperature is only a little more intuitively accessible. For this derivation we only needed to know one intuitive fact about temperature, which is that hotter things have higher temperatures. But if we wanted to do anything with that ratio of temperatures besides note that it is greater than one, we'd have to confront the fact that this ratio is a bit mysterious, since the temperature scales you learn before you study thermodynamics have arbitrary zero points. If I understand the history right, absolute temperatures are "for" thermodynamics, like Thompson introduced them so that they can be measured as work in some sense which I don't really get. I'll get it after I learn to derive Carnot efficiency I guess. I'll derive that from the laws of thermodynamics, but that's backwards from the historical order; I think the history is more like, temperature is defined as it is so that Carnot efficiency will work out that way.

So this is the character of thermodynamics, and what makes it more difficult than an intro physics class. You start with some mysterious-looking laws relating quantities with nonobvious meanings, and have to fight your way from there back to the physical facts that the theory was based on.

Quantum mechanics has this characteristic as well, but it's a much longer and harder fight.

[tototavros]

So, about this new "AI 2027" report...

I have not read the whole thing in detail, but my immediate reaction is kind of like what I said about "Bio Anchors" a while back.

Like Bio Anchors – and like a lot of OpenPhil reports for that matter – the AI 2027 report is mainly a very complex estimation exercise.

It takes a certain way of modeling things as a given, and then does a huge amount of legwork to fill in the many numeric constants in an elaborate model of that kind, with questions like "is this actually a reasonable model?" and "what are the load-bearing assumptions here?" covered as a sort of afterthought.

For instance, the report predicts a type of automated R&D feedback loop often referred to a "software intelligence explosion" or a "software-only singularity." There has been a lot of debate over the plausibility of this idea – see Eth and Davidson here for the "plausible" case, and Erdil and Barnett here for the "implausible" case, which in turn got a response from Davidson here. That's just a sampling of very recent entries in this debate, there's plenty more where that came from.

Notably, I don't think "AI 2027" is attempting to participate in this debate. It contains a brief "Addressing Common Objections" section at the end of the relevant appendix, but it's very clear (among other things, simply from the relative quantity of text spent on one thing versus another) that the "AI 2027" authors are not really trying to change the minds of "software intelligence explosion" skeptics. That's not the point of their work – the point is making all these detailed estimates about what such a thing would involve, if indeed it happens.

And the same holds for the rest of their (many) modeling assumptions. They're not trying to convince you about the model, they're just estimating its parameters.

But, as with Bio Anchors, the load-bearing modeling assumptions get you most of the way to the conclusion. So, despite the name, "AI 2027" isn't really trying to convince you that super-powerful AI is coming within the decade.

If you don't already expect that, you're not going to get much value out of these fiddly estimation details, because (under your view) there are still-unresolved questions – like "is a software intelligence explosion plausible?" – whose answers have dramatically more leverage over your expectations than facts like "one of the parameters in one of the sub-sub-compartments of their model is lognormally distributed with 80% CI 0.3 to 7.5."

---

Maybe this is obvious, I dunno? I've just seen some reactions where people express confusion because the whole picture seems unconvincing and under-motivated to them, and I guess I'm trying to explain what I think is going on.

And I'm also worried – as always with this stuff – that there are some people who will look at all those pages and pages of fancy numbers, and think "wow! this sounds crazy but I can't argue with Serious Expert Research™," and end up getting convinced even though the document isn't really trying to convince them in the first place.

---

Now, if you do buy all the assumptions of the model, then yes, I guess this seems like a valuable exercise. If you are literally Daniel Kokotajlo, and hence believe in all the kind of stuff that Daniel Kokotajlo believes, then it makes sense to do all this legwork to "draw in the fine details" of that high-level view. And yeah, if you think the End Times are probably coming in a few years (but you might be able to do something about that at the margins), then you probably do want to get very precise about exactly how much time you have left, and when it will become too late for this or that avenue for change.

(Note that while I don't agree with him about this stuff, I do respect Kokotajlo a lot! I mean, you gotta hand it to him... not only did he predict what we now call the "Gen AI boom" with eerie accuracy way back in 2021, he was also a whistleblower who refused to sign OpenAI's absurd you-can't-talk-about-the-fact-that-you-can't-talk-about-it non-disparagement agreement, thereby bringing it into public view at last.)

But, in short, this report doesn't really touch on the reasons I disagree with short timelines. It doesn't really engage with my main objections, nor is it trying to do so. If you don't already expect "AI" in "2027" then "AI 2027" is not going to change your view.

[tototavros]

lots of people want understanding what's going on with consciousness to be like science and im just really not sure this is the kind of question science is going to be very helpful with. like. these theories of consciousness dont make testable predictions (that honestly depend on the theories in a systematic way, arent just guesses), of course they dont, because what were interested in here is exactly that thing which is not externally observable! any discussion of the nature of consciousness is going to be all philosophy, and when neuroscience tries to intrude it's usually kind of bullshit (i *hate* that study people love to quote about how you press the button before you form the intention to press the button, very slightly. theyre just ASKING people when they "formed the intention". you believe that shit?!). at best it can rule out certain theories, which DO make explicit material claims. it can tell us what consciousness *isnt*, but not what it *is*, yknow? there's a "smooth core" that doesnt poke out

[tototavros]

"I only get all Jacobinard in the presence of unjustified monopolies, I swear, I'm really a Proudhonist at heart"

[tototavros]

The Yemen bombing group chat messages accidentally shared with The Atlantic are now public. Here’s three high-ranking government officials celebrating the destruction of a what sounds like an apartment building, part of a series of airstrikes which reportedly killed 53

[tototavros]

I was going through some archived scans of 90's otaku magazines, as is my sacred duty, and I stumbled on this ad for a Sega Saturn game I did not know:

The pitch of Roommate (as seen here) is that of a "real time" romance simulation:

What makes it real-time is that the game progresses in sync with the Saturn's internal clock. In that way [main girl] Ryoko is just like a real girl; she has her own daily habits and lives her life accordingly. So if you start the game in the afternoon, you might not be able to meet her because she's at school [...] The purpose is to enjoy living together with Ryoko in real time and communicating with her.

And this is exactly the kind of way-too-convoluted gimmick that sacrifices gameplay functionality on the altar of conceptual novelty based on random technology add-ons present in new-gen consoles of the era that I just love. Obviously the concept of starting a game and having the main girl not be present so you cannot play is completely asinine - but think of the realism!

Between that and the discount-Sadamoto 90's character designs, I wanted to see it for myself; so I spent way, way too long setting up a Sega Saturn emulator. In my experience early CD-ROM-based consoles often require much more bespoke set-ups to get working, in this case custom BIOS files in the emulator firmware directories, and JPN-language ones at that for this game. But I got it to work and oh yeah, this is some early "digital" console era crust:

Playing this game is just painful. The clock of course means that you essentially can't play it at all - looking at YouTube comments on the very few Let's Plays and such that exist, people are reminiscing about how they could never find Ryoko because their schedules didn't align. One person even comments:

This game is for NEETs and shut-ins

Which is a valid demo I guess! But it doesn't really stop there - your house is a "fully realized" 3D environment of bare walls which you navigate with clunky controls. Let me log in and take some screenshots...

Jesus Christ it's 10 pm and you are cooking dinner?! The one time I don't want this ghost popping out of the cracks in the floorboards, I swear...

Okay, got rid of her (She broke a plate -_- you moved in yesterday, girl):

You walk, in real time (step by step) through this pixel museum just...hoping that one of the rooms will contain Ryoko and proc a dialogue event based on the time of day. There is a little more to it than that but that is essentially the gameplay. This would, very obviously, be simply better as a straightforward visual novel.

But you see how that just isn't as cool in 1997, right? This is the era where the fidelity of graphics and the technology for simulation is progressing at a rapid clip, and everyone wants to see the boundaries pushed. Roommate isn't the first "real time simulation" game, but it is the most pure, the one fully committed to the bit. Your house is completely mapped out, the girl has her routine, you walk step by painful step through the rooms because this is "real", you are living it. They even use a live photo for the outside of the house to sell the aesthetic (and also save money):

Ryoko is waiting in the kitchen of that house when you come home from work, putting on an apron, ready to cook dinner. For you.

Assuming you get home at whatever fucking 30 minute window the game decided to gatekeep its gameplay behind! But of course I exaggerate - it wasn't that bad (there are little mechanics you can use to set some schedule times in the game for example), player tolerance for bullshit was way higher then, and you were expected to buy strategy guides for these things. So even though it was panned by critics on release...it was a sleeper hit with a devoted fanbase.

Which means it got a ton of sequels and ports! We don't have to go through them all, though I will share my favorite factoid about the first sequel - "ROOMMATE ~Ryoko in Summer Vacation~" from the wiki:

The character designs are significantly different from the previous game (especially Ryoko's brown hair and large breasts).

Priorities, baby. But some of the ports are interesting because of the changing tech. A version was ported to the PlayStation, which does not have the internal clock a Sega Saturn had. But coincidentally it did have the PocketStation, a handheld GameBoy/Tamagotchi hybrid expansion tool that did have an internal clock and could sync with the game. It also let you track Ryoko's schedule and play mini-games, with some very adorable animations as you can see in this ad for the product that featured Roommate:

This device absolutely reminds me of the Disc Fax system discussed in my Miho Nakayama essay - a very niche product biting off more than it can chew making games overly complex to play but allowing things that would otherwise be impossible (and this one was a good deal more successful at least). Here it allowed Roommate's central gimmick to function - and is super cute, honestly I would buy a standalone tamagotchi version of this game.

The PS1 also couldn't quite handle how the game was built for the Sega Saturn graphics-wise, and as such a bunch of the 3D elements were sanded off into 2D simulacrums - most notably the house:

Which, despite this being a technological downgrade, is way better! It looks adorable, you can actually see what is going on and where Ryoko is, and you can navigate it way more cleanly. God, did...hold on let me tab back to the game...yeah, is there no clock in the original game on screen. That is insane. Anyway the PS1 version had a lot of these cute little graphical additions, even right on the title screen:

It is definitely the better looking version, which is a classic tale - in 1997 the "bleeding edge" of 3D graphics were impressive to players, even through their roughness. Now they just aren't, and so the retro charm of designs that are optimized what the mediums of the time could reliably handle have a lot more appeal.

There was also a PC port in 1998, which did exactly what I suggested and added an "adventure" mode where you could ignore the clock system. They definitely learned over time what worked and what didn't; but the appeal of the gimmick is what first sold it to players in the end.

All of this is to say, don't play Roommate, and if you do just emulate the PS1 game instead of torturing yourself with the Sega Saturn version. Oh...you weren't gonna play a Japanese-only abandonware 90's not-even-eroge dating sim to begin with? Ah, well, yeah, I guess that makes sense.

Man I should translate it shouldn't I? So people can play it...

[tototavros]

rebellion really did say gay love between girls was the center of everything and the point of it all. and also all that other stuff about love too but let's not focus on that bit

[tototavros]

"She wanted to know why I wasn't a rapper!--No! There wasn't a smile on her face, I think she meant it!"

[tototavros]

How do you *accidentally* make a programming language?

Oh, it's easy! You make a randomizer for a game, because you're doing any% development, you set up the seed file format such that each line of the file defines an event listener for a value change of an uberstate (which is an entry of the game's built-in serialization system for arbitrary data that should persiste when saved).

You do this because it's a fast hack that lets you trigger pickup grants on item finds, since each item find always will correspond with an uberstate change. This works great! You smile happily and move on.

There's a small but dedicated subgroup of users who like using your randomizer as a canvas! They make what are called "plandomizer seeds" ("plandos" for short), which are seed files that have been hand-written specifically to give anyone playing them a specific curated set of experiences, instead of something random. These have a long history in your community, in part because you threw them a few bones when developing your last randomizer, and they are eager to see what they can do in this brave new world.

A thing they pick up on quickly is that there are uberstates for lots more things than just item finds! They can make it so that you find double jump when you break a specific wall, or even when you go into an area for the first time and the big splash text plays. Everyone agrees that this is neat.

It is in large part for the plando authors' sake that you allow multiple line entries for the same uberstate that specify different actions - you have the actions run in order. This was a feature that was hacked into the last randomizer you built later, so you're glad to be supporting it at a lower level. They love it! It lets them put multiple items at individual locations. You smile and move on.

Over time, you add more action types besides just item grants! Printing out messages to your players is a great one for plando authors, and is again a feature you had last time. At some point you add a bunch for interacting with player health and energy, because it'd be easy. An action that teleports the player to a specific place. An action that equips a skill to the player's active skill bar. An action that removes a skill or ability.

Then, you get the brilliant idea that it'd be great if actions could modify uberstates directly. Uberstates control lots of things! What if breaking door 1 caused door 2 to break, so you didn't have to open both up at once? What if breaking door 2 caused door 1 to respawn, and vice versa, so you could only go through 1 at a time? Wouldn't that be wonderful? You test this change in some simple cases, and deploy it without expecting people to do too much with it.

Your plando authors quickly realize that when actions modify uberstates, the changes they make can trigger other actions, as long as there are lines in their files that listen for those. This excites them, and seems basically fine to you, though you do as an afterthought add an optional parameter to your uberstate modification action that can be used to suppress the uberstate change detector, since some cases don't actually want that behavior.

(At some point during all of this, the plando authors start hunting through the base game and cataloging unused uberstates, to be used as arbitrary variables for their nefarious purposes. You weren't expecting that! Rather than making them hunt down and use a bunch of random uberstates for data storage, you sigh and add a bunch of explicitly-unused ones for them to play with instead.)

Then, your most arcane plando magician posts a guide on how to use the existing systems to set up control flow. It leverages the fact that setting an uberstate to a value it already has does not trigger the event listener for that uberstate, so execution can branch based on whether or not a state has been set to a specific value or not!

Filled with a confused mixture of pride and fear, you decide that maybe you should provide some kind of native control flow structure that isn't that? And because you're doing a lot of this development underslept and a bit past your personal Balmer peak, the first idea that you have and implement is conditional stops, which are actions that halt processing of a multiple-action-chain if an uberstate is [less than, equal to, greater than] a given value.

The next day, you realize that your seed specification format now can, while executing an action chain, read from memory, write to memory, branch based on what it finds in memory, and loop. It can simulate a turing machine, using the uberstates as tape. You set out to create a format by which your seed generator could talk to your client mod, and have ended up with a turing complete programming language. You laugh, and laugh, and laugh.