Azhdarchid

I'm Bruno Dias, a São Paulo-based designer (systems, narrative, mechanics) and writer working in video games. Azhdarchid (named after the family of large pterosaurs) is my personal blog; I also have a more professional landing page at brunodias.dev.

You may know me from such video games as Where the Water Tastes Like Wine, Neo Cab, Pathologic 2 and 3, or Fallen London—you can see my full credits on MobyGames. I also edit the Game Narrative Reader, a link anthology of great essays about video game narrative design on the web.

For my sins, you can find me on Bluesky (@brunodias.dev) and on the Fediverse (@bruno@azhdarchid.online). For my cinema sins, I'm also on Letterboxd (notbrunoagain).

Recent Posts

Jump to full post list

On Waymos and their safety record

On Waymos and their safety record

Laura Michet writes:

Long-term, there is little incentive for self-driving car companies to prioritize the safety of road users like cyclists. They can eat any amount of fine payments. They can bring any amount of legal influence to bear. It is simply not wise to allow this kind of power imbalance between road users! I think the threat of a tech company becoming a road user is, long-term, greater than the benefit of automatic driving.

Laura Michet's Blog, Why I oppose self-driving vehicles

This is a really good point I hadn't really considered myself, although I also generally oppose "autonomous" vehicle deployment by corporations for a few related but distinct reasons.

The first one is that I think safety claims from companies like Waymo need to be treated with extreme skepticism; tech companies lie basically all of the time about basically everything, so they're probably lying about this. Don't tell me they wouldn't do pivot to video with people's lives.

But beyond outright lying, many of the claims are just misleading. Take this Futurism article from last year, very impartially titled "New Data Shows Waymos Are So Safe That It's Almost Comical".

The crux of the article is a comparison between the crash rate of Waymos and the crash rate of "all drivers", which as far as I can tell from Waymo's own descriptions of their methodology, means all drivers in US counties that Waymo operates in.

These comparisons inherently require statistical adjustment to approximate an apples to apples comparison, because Waymos are just not driving the same routes that human drivers are; in particular, they do very little highway driving, which is significantly more dangerous than city driving—per the US DoT, rural driving has about a 50% higher fatality rate per distance traveled.

Waymos are also, you know, Waymos—small minivans and crossovers, all electric, all pretty new, all very inherently safe. We're comparing this to the median American car, which is an SUV with a fetish for rolling over or a pickup truck with a fetish for running over children.

Waymo has published peer-reviewed papers about their safety record—all in the same journal. It's worth questioning whether anyone with a real incentive to interrogate their methodology is in a position to do so. Everyone in power is kind of broadly incentivized to go along with Waymo's claims. Auto makers want a new market for cars; policy makers want a convenient tech solution to road safety problems that doesn't require them doing work or the state spending money. The idea that self-driving is extremely safe is extremely convenient, which is itself a reason to be skeptical.

It's also worth saying that by far the most convenient way for Waymo to exaggerate their safety record is this: Waymos are not always operating autonomously; they are often being guided by remote operators, whom the company will not refer to as "drivers" for liability reasons. Whatever Waymo says about when exactly or how often those remote operators are in control is, of course, something that Waymo gets to disclose or lie about with total impunity; it's not independently verifiable.

The way Waymo talks about road safety is also designed to induce in the reader and overestimation of the safety gain from Waymos. A large proportion of road accidents is caused by either drunk driving or speeding. If you're walking into a rideshare or taxi vehicle with a presumably sober driver who isn't going to drive recklessly, you are already significantly safer than the average person on the road, so the broad comparison Waymo is making is exaggerating what the safety gap is in specific circumstances.

The conversation also inherently marginalizes questions of what things we could do to promote road safety that are not giving over control of the road to one of the world's most evil corporations. How would Waymos compare to human drivers who were being more effectively induced to drive safely through traffic calming and other infrastructure that generally isn't being adopted by cities in the Americas?

However, all of this is just pointing out that Waymo is almost definitely exaggerating their numbers and presenting them in a very rosy light; even so, you might look at the actual numbers and say "well, it's such a huge difference, there's definitely a meaningful benefit there."

Here's the problem with that. One reason Waymos are very safe because they drive extremely conservatively. For this reason, they are also slow and will sometimes get "stuck" when they can't determine a safe way forward. This approach to driving is a good idea for Waymo right now, when they're trying to gain regulatory acceptance for their vehicles in more and more cities.

But let's say Waymo captures significant market share. Let's say a large city, or several large cities, actually do become meaningfully dependent on Waymo for transportation. Being slow makes Waymos uncompetitive against other modes, and it makes the service costlier to operate.

On a self-driving vehicle operated by a corporation, road safety is a dial. It's something Alphabet just gets to turn up or down depending on their goals. If Waymo ever becomes established as a significant transportation provider, very soon after there'll be a meeting in a dark room in Mountain View where they'll decide, for all of those cities they're operating in, what an acceptable level of road fatalities is. What an acceptable level of cyclist-car collisions is. That number probably looks pretty similar to the number that policymakers throughout the US, Canada, and Latin America tend to think is acceptable right now, given the degree to which cars are tacitly allowed to kill people. They'll turn that dial until it hits that "acceptable" equilibrium, because safety costs them money.

There are no technical solutions to political problems. Waymo cannot solve road safety, categorically; it can only shift the locus of policymaking about road safety, away from the democratic control of state institutions and into an unaccountable, dishonest, and fundamentally evil corporation. In the process, the seduction of easy technical solutions can only harm progress towards real policy change.

To think Waymo is a solution to road safety is a trap. You are talking about buying safety from an entity that has zero commitment to human life.

Lyre Developments for November

Things have gotten a little strange with this project.

It's been about a month since my previous post about lyre, and the language has changed very substantially (to the point that the current main branch is no longer compatible with the version I used to write First, Second, and Third Persons.

Probably the most striking change is that I "caved" and made the language an island grammar; that is, the source file is interpreted as player-facing text by default and "code" is demarcated within it, as is the case in ChoiceScript and Ink.

=storylet paris_1886
    # Paris, 1886

    You still recall that heady summer of '86, which stretched on for far too
    long…

    ~ recalling_parisian_memories++

But the interesting thing about this is that this is not the product of just a syntactical change where I realized it's more ergonomic to write text without quotes, it's the product of a deep semantic change that has almost entirely replaced the guts of the language.

Originally, lyre conceptualized storylets and branches as having a "body" and a "title" of text, and each was a single expression that would evaluate to the text that's presented to the player.

Like any narrative scripting language, lyre needs some way to evaluate arbitrary expressions; predicates like language proficiency > 10 or times_spoken_to_vicar < 7, but also calculations like time_elapsed = time_left - time entered. Yarn uses a stack-based virtual machine; for lyre, I wrote a very basic interpreter that's conceptually similar to a Lisp interpreter, which has the advantage of being simple and easier to inspect while the language is still actively changing, and required less binding architectural decisions early on.

In this early form, the expression evaluator sat in its own corner and evaluated small short expressions, but most of the core logic was left to the narrative engine—which is just straightforward code that runs a storylet system, figuring out which branches to display and how game state changes. The expression engine was relegated to calculating values that would plug into the decisions the narrative engine made.

However: the expression engine has, as the language evolved, gradually eaten the narrative engine, to the point that what remains of the latter is now really kind of a very simple orchestrating layer whose job is just delegating the right things to the right expression engine components.

Which is to say, storylets used to be made of distinct kinds of statements, some of which could contain expressions, but which were distinct among themselves. The text body of a storylet was one specified property of that storylet, a single string value (or a string with variable text, which is in fact just an expression).

Now, most of the different kinds of statements that make up storylets (text and effects) have been consolidated into the expression language. Storylets don't have "a text body" as a distinct component, and execution is more similar to something like Yarn; each line of text is its own statement. This allows interleaving text and effects in ways that work how you expect:

=storylet recruitment
    ~ new_sailors = d(6)
    Your crew of { sailors } sailors expands with the addition of
    { new_sailors } new able seamen,
    ~ sailors = sailors + new_sailors
    for a total of { sailors }.

Lyre was originally conceptualized more as a data format than a programming language, closer to YAML than anything. All the storylets in Voyageur were in fact just written as YAML files.

But, as it evolves to develop new and more generalizable functionality, it starts to resemble a programming language more, and specifically the arc has been that it starts to resemble Haskell more and more, which is very funny.

This isn't because I love Haskell—though I have a sort of guarded fondness for it—or because I think it's the best possible design for a programming language, let alone a specialist tool like lyre. This is because it turns out the problem lyre is designed to solve is actually very similar to the space Haskell is designed to explore.

Lyre's central idea is the notion of narrative explicitness; state only ever changes through quality changes, values that are like messages emitted out of the narrative engine, meant to be made visible to the player or used as the origin point of events for things in a game's UI and world.

This is sort of the aesthetic and structural heart of StoryNexus and games in that mold. Lyre is designed around it. And if you have a certain kind of programming background, you may already see where this is going. "No silent or implicit changes to story state" starts to look a lot like functional purity. "The result of a state change is itself a value" starts to look a lot like a monad.

I'm not going to define what a "monad" is here because it wouldn't help you in the slightest if you don't already know. In this context, really, what a monad is is just a way of chaining together different computations while keeping track of their relationships; you can think of the "monad" here as being kind of like a combine harvester, going over all the individual expressions that make up a storylet, sorting text from quality changes from branches, and storing them in the hopper to be read as the combined result of the storylet.

Lyre is designed to have a stateless runtime, meaning that the runtime doesn't store any internal state. This is useful in a narrative engine because it makes it easy to implement things like full undo or redo, or traversal between narrative states, or the ability to look ahead at the result of taking a branch without changing any internal state. It also kind of naturally flows from the idea of the language not having any implicit state changes anyway, and makes the runtime both smaller and less likely to contain subtle bugs.

I also wanted lyre to be viably something you could use to build a server-based live game, and statelessness is really desirable there, because for a game like that to scale it has to work in a way where you can't guarantee that a given player's requests will all be answered by the same server process; there can only be one source of truth, and that has to be a database that lives apart from the game logic.

So the way the lyre API works is that every time you call it (to advance the game state), you have to supply the game state along with your call; what you get back is a representation of how the state changes, but the original game state object you passed isn't mutated. The difficulty here though is that people very strongly expect lyre to behave as if it keeps track of internal state.

That is, you expect to be able to write this:

=storylet gain_wealth
    ~ wealth_roll = d(20)
    ~ gems = gems + wealth_roll
    ~ cash = cash + 100 * wealth_roll

...and have it work as you expect.

And indeed it does; the way lyre interprets a storylet is by evaluating each statement in the context of the previous ones, essentially, and this is done through an implementation of something analogous to the State monad in Haskell.

Conceptually, a storylet in lyre is a function that maps from game states to other game states, and not just a series of instructions for doing something. But ergonomically, you can write it as a series of instructions just like you would in Ink or Yarn.

This is all to illustrate how path-dependent things can be in realizing a tool like this. I started pulling on the thread of "what if we allowed unquoted text" and where that ended was totally rethinking the language conceptually as a prerequisite for allowing unquoted text in a sane way.

Previously, interleaving effects and text didn't make sense, because evaluating the body text and evaluating changes were two distinct steps orchestrated by the narrative engine, and one happened strictly before the other. Now it's all one process.

There are still some open questions about lyre's syntax and semantics, and the language is still evolving. One questionable area right now that lyre has a very regimented structure and specifically limits how entities can nest in the source code. Storylets contain branches; branches can contain results, which are really just anonymous storylets meant as the "leaves" of the story tree. This very much maps onto how StoryNexus works, and it is intended to make it so the "grain" of the language is towards a flatter structure that's more at home in a storylet narrative.

However, as the language becomes more expression-based, the idea of branches and results also being syntactically construed as expressions rears its head... But if that were to happen, you'd need some kind of delimitation for them, whereas lyre currently literally has none—it looks a bit like an indentation language, but it actually isn't at all. Blocks simply last until the parser runs into the start of the next block at the same or higher level.

Indentation languages have a bunch of annoying downsides, and I also find them aesthetically objectionable, so I don't want to go in this direction; especially because it would be specifically to support use cases that are likely to lead to gory and hard to understand use of indentation! I am not sure, yet, of how to address this, or if it is even possible or desirable to address. It may be that it's just not practical to fold the syntax that creates branches/results into the expression syntax.

To concretize this, results look like this currently:

=storylet root
    * A branch looks like this now—like in Ink.
        An arrow points to another storylet by name.
        -> elsewhere
        =>  And a "fat arrow" is a result, ie a storylet written inline here.
            A result can have multiple statements, lines of text, etc.
    * The parser knows that this new branch must mean the end of the result.
        Indentation does not matter.

And it'd be nice if you could do this:

=storylet root
    * Branch
        ~ if (some_condition) => This result is part of an expression!
            But it still can have multiple lines, as it would need to to be
            useful.

But the problem is that once results have wormed their way into the expression language, that implies you can write...

=storylet root
    * Branch
        ~ if (some_condition) => Result
        ~ if (another_condition) => This expression statement (everything from
        the ~ on) is actually *nested* inside the other result. The parser
        has no way of placing this on the "outer" level.

Lyre currently doesn't have a type checker, but it will eventually, and as things stand it would have to choke on this because you'd have an expression statement within a storylet that has the type Route; a route is the type of statements like -> elsewhere, a description of which storylet to go to next; they only have meaning as part of a branch.

Thus breaking lyre's nice and neat "blocks go on forever until they must stop" syntax. One way of addressing this is to give lyre semantic indentation... but again, I don't want to do that because it is annoying across multiple vectors, even if the construct is less Python-like "brackets are indents" and more Haskell-like "indentation is expression continuation".

Possibly the easiest version of this kind of syntax to build is the "two-dimensional," truly Haskellish one where everything more indented than the arrow itself becomes part of the result block that encompasses the arrow, but that is so cursed.

An alternative would of course be just adding explicit delimiters, like {}, but of that's less than ideal for something highly ergonomic like a narrative scripting language. It may just be the most practical solution though.


Ultimately building lyre has been a very gratifying exercise that has me doing a lot of deep consideration of the aesthetics and design of this kind of tool, and I look forward to continuing to refine it towards being more useful.

I am finding increasingly that there's this law of equivalent exchange to designing something like this where often building regularities and meeting likely user expectations involves building more complex or more sophisticated inner systems that are then veiled from the user—again, the fact that lyre increasingly resembles a functionally pure ML is actively something the language is designed to conceal from the user.

A ★★½ review of Buffet Infinity

A ★★½ review of Buffet Infinity
The poster for Buffet Infinity (2025

Reviewed by Bruno Dias

2025 · ★★½

For what this is (a horror-comedy told largely through hokey early-2000s local tv advertising, about a buffet restaurant that seems to be the locus of some kind of eldritch horror) it's a surprisingly staid, down-the-middle movie. It really doesn't escape how much it's just a bunch of recycled ideas put into a box and shaken around until they blur into each other. You can kind of start spotting all the influences here a little too neatly: Junji Ito, found footage horror, Be Kind Rewind, the backrooms, millennial nostalgia, jokes about scientology, Garth Marenghi's Darkplace, getting high and watching tv with too many ads, SCTV.

It feels like it kind of gets lost in its feature length, doesn't really know how to convincingly escalate, doesn't really know how to end. I kept waiting for it to find a voice that was something of its own rather than just vaguely mashing up and imitating other, more interesting things. And it eventually did and it wasn't good. So.

Big Balk

On Big Walk, IGF, and the Present Unpleasantness.

So I'm assuming that you, dear reader, are aware of the Present Unpleasantness. But if not, a brief recap:

On Monday (November 9th), a group of IGF jurists published an open letter denouncing the harms of AI, pointing out that using LLMs for coding assistance is no different from other uses of the technology, and vowing not to nominate games made with AI for IGF awards.

At this point in time it was public knowledge that a great internecine fight had happened within IGF judging over the topic of AI, centering on a highly popular game having an AI disclosure. Shortly after those jurists released their statement, the information leaked that the game in question was House House's Big Walk. Multiple IGF judges confirmed this, quickly followed by House House themselves adding an entry to their FAQ that more or less repeats what was said to IGF judges about the LLM usage in the game.

Two things about this are frustrating to me—I mean, everything about this is frustrating to me, but two things I want to point out are frustrating. One is that discussion of Big Walk and its merits will now overshadow a broader discussion. Another is that a lot of the arguments that took place in the 600-post IGF judge thread about Big Walk are now playing out over social media, recapitulating many of the worst misinformed and disingenuous arguments.

In a lot of ways House House couldn't have created a nastier situation for IGF judges if they were deliberately trying; they used LLMs and disclosed their use, but they used them relatively little, and they phrased their disclosure in a way that made it easy for people to minimize the use even more so. They didn't use any generative AI that would have required a Steam disclosure, so their IGF disclosure was an unpleasant surprise for judges, many of whom had already played the game before the judging period even started.

Big Walk went into the IGF as a favorite for the grand prize and design awards; it is an extremely successful video game from a studio with a significant pedigree, while still being made by a small enough core team to justify an "indie" label.

Thus, the biggest game of the year, which a lot of people had very positive feelings about, was now also at the center of the debate around AI use; and the game had used AI in a way that enabled every dishonest argument about the issue, viz.:

  • The argument that they used LLMs very little, so little it shouldn't count;
  • The argument that their LLM usage was basically indistinguishable from reading a Google summary;
  • The argument that it shouldn't matter because the game is very good;
  • The argument that House House are good guys and we like them, therefore we should overlook this;

Some kind of reckoning about AI was inevitable this year. IGF being beholden to Informa, a company that has never seen a stupid tech fad it didn't want to jump aboard of, makes it politically difficult for the organization to resist the encroachment of slop. On the other hand, many judges and jurists feel very strongly about the issue. It's still unfortunate that it played out in the most annoying way imaginable.

Big Walk is not just popular, it's a game that people clearly have a lot of very strong feelings about owing to their very particular and intense experiences with it. I feel like this has led people to go to some extreme, almost delusional lengths to try to construct a moral universe in which House House didn't do anything they'd consider wrong.

The studio described to IGF judges a scenario where one or more persons at the studio opted to engage an LLM chatbot for coding assistance multiple times, to solve several problems over the course of production. In lot of social media discussion this has morphed into "House House had a showstopping bug that was going to ruin the game and they used Claude as a glorified search engine to solve it", which is just an alarming signal of how parasociality makes people jump to defend someone who isn't even trying that hard to defend themselves.

I don't think House House, themselves, are as invested in the idea that they are blameless as a lot of their social media shooters seem to be; to the extent that the discourse about this may now be actively harming their own ability to amend their behavior going forward.

I don't think House House were being deliberately dishonest in characterizing events—I think if anything that characterizing them at all shows a willingness to have a discussion—but I think a lot of people without subject matter knowledge are reading into their recounting of events a lot of things that just aren't so. Unity obviously dropped the ball in supporting House House (quelle fucking surprise), but many seem to read into that event the notion that the problem they ran into was utterly intractable.

Everything I know tells me that this wasn't the case, and using an LLM was an expedient, not a necessity. Unity being buggy and underdocumented is nothing new. House House could have worked around the issue (note that the worst case with the given issue was that the game was a bit frame-y for 20 seconds occasionally during play) or engaged the help of an actual human expert, many of whom they would have had access to just in their local scene around Melbourne. Many of these people are out of work right now.

House House, an already very successful studio with not insignificant resources, made a real choice to rely on a morally abhorrent software system rather than engage the help of real people—whether in the form of paid contracting or just friendly help. I want to say this not to try to exaggerate the scope of what they did but to make it absolutely clear that there is a real, material dimension to this, even if some want to characterize this issue as a purely abstract test of ritual purity.

It is, of course, also true that to use LLMs is to willingly participate in a noxious political project and to use the product of exploitative and morally unacceptable processes; and that objection is also real, material, and grounded. I do not trust the ethical instincts of people whose response to those objections is to wave them off and say "go yell at Sam Altman", dismissively pushing the site of struggle far away from anyone they might know or talk to. The reality is that many people in the industry have opened the door for the encroachment of tech corporations that want to do us (the professionals who actually work in this industry) harm, who want to do humanity harm, who want to do the planet harm. That matters. Other people's behavior matters. You are not powerless cogs in a capitalist machine that can't be held to account any more than the batteries in Mark Zuckerberg's Roomba.

I'm not going to try to argue that any and all LLM usage is equally bad, but either an ethical boundary exists or doesn't. And we are living in a moment where the urgency and necessity of this ethical boundary is enormous; vast resources have been marshaled to simply make everyone's lives worse in service of a vision of the future that serves nobody but five dudes who are already wealthy. Is arguing that people are taking it too seriously how you want to spend this time?

There are some ethical faults we generally expect people to not be 100% compliant with, of course; it's very difficult to live in society without wasting some amount of single-use plastic, eg.

But there are some ethical faults we think it's reasonable to expect people to never breach without sanction. I think using a completely optional set of highly subsidized SaaS products that did not exist two years ago is in the latter set. You might as well be talking about the "purity test" that a studio should credit 100% of their contractors and pay 100% of their invoices.

If you disagree, perhaps stop to think about how heavily you've been propagandized and how much you've mentally surrendered to the normalization of these systems. It is very possible to do things without using an LLM, even things that are hard; and if anything, one of the harms of the "AI" political project is the propaganda machine dedicated to making you feel stupid and helpless.

I don't want to go into another litany listing the harms of LLMs here, that's all been very thoroughly discussed at this point. But I feel like it needs to be absolutely clear that there is an ethical problem with using this technology, and this ethical problem is not reducible by scale (no ethical problem is, really). You either recognize this ethical problem or you don't; and if you do, using an LLM "just a little" is still crossing an ethical boundary (even if just a little) and it still merits being treated as the crossing of an ethical boundary.

In regards to IGF, of course, nobody is entitled to win an award or nomination. The vast majority of entrants win nothing, including some who submitted extremely high-quality work. An award like IGF is meant to elevate and promote excellence, and it's obvious to me that therefore we want winners to reflect the practices we want to see in the industry. Which means that using an LLM, even if "just a little", should indeed be disqualifying.

Excellence in this sense shouldn't just be some nebulous anointing of "quality"; there are plenty of other awards for that. Excellence should mean an aspirational sense of what the industry should be, and what the industry should be is not one where we prefer getting help from a corporate software product meant to make everyone else worse off rather than someone in our own community.

If we recognize this ethical boundary, then we recognize that Big Walk may be an enjoyable game or one that has valuable things to say but it's not excellent in the way that IGF demands. It should not win anything.


Another issue that is repeatedly posed is the question of "transparency" and "disclosure", namely that negative reactions to uses of AI will discourage people from disclosing that they used AI.

The problem here is, what exactly do you expect? What do you think should happen? Do you think that, eg, when Paolo Pedercini posted his self-pitying little blog post about using AI (much more egregiously than House House, mind) in Future? No Thanks, that people should have stroked his ego for being brave enough to admit it?

Like, has your ethical compass not advanced since kindergarten? Are you still in the "it's good that you admitted you did it" phase? Is the idea that something can be an ethical violation, and someone can admit to that ethical violation, but no one is allowed to be disappointed by the behavior that was admitted to? "You can't be mad at me because I was honest" is, frankly, just immature.

Those studios have done things that were disappointing. I don't think they need to be harassed or run out of the industry over it but people are allowed to be disappointed, wary, or unhappy about it.

Again, either an ethical boundary exists or it doesn't. If it exists, there is nothing you can do to make voluntary disclosure a desirable thing to do. People who believe they are bound by an ethical rule are bound to behave accordingly, and that means disapproving of violations of that rule! And therefore, people who are unscrupulous and willing to lie will always lie to avoid that disapproval.

This is, frankly, kind of just the nature of having ethics, so I sometimes wonder with no small amount of alarm if this is some persons' first time doing it. In order to actually have an ethical stance it has to bind your behavior, which means that if the ethical boundary exists then you have no choice but to create the situation in which disclosing is undesirable. We cannot escape the gravity of our convictions, other than by becoming totally amoral.

Many made the argument that if IGF judges refuse to vote for games made with LLMs, then studios will lie about using LLMs to avoid being snubbed by judges. This reasoning is morally backwards. I, as an IGF judge, have literally no choice but to vote against games with AI disclosures because once you tell me that you used an LLM to make the game, that information now binds my behavior. Because the alternative would be dishonesty; I'd have to pretend that a game which fails the criteria that I think are important in judging is somehow still valid on the merits.

Obviously I could do it, it's not like it's literally physically impossible. But it would feel bad, and other people who share in my stance would then be disappointed in me. I feel like I'm having to explain the very basics of ethics, as if to a six year old, here.

Possibly this is because some of you don't really believe in the ethical urgency of the moment and are just being carried along by trying not to offend those of us who do. Possibly this is because some of you don't really believe in the ethical urgency of anything. I don't know. Fix your hearts or die.

Ultimately, it is impossible for me to care about whether people feel "incentivized" to voluntarily disclose their ethical violations because it is impossible for me to affect their incentives, because doing so would require that I throw out the ethical stance in the first place. You're not locked in here with me; I'm locked in here with you, asshole.

As an actual practical matter, when we set forth a rule about behavior that goes on behind closed doors where we can't see it, we always know that cheating is a possibility and will happen. But this doesn't mean those rules are pointless.

They have value in communicating what a space is, what it is for, who is welcome, and what the values the space holds are. To some extent banning AI is just practical; it spares us from having to engage with the AI enthusiasm of AI boosters or having to spend too much time on the obvious low-effort slop—of which there was plenty this IGF, and the people who submitted these games would have been better off just shredding their entry fee instead.

They have value in removing the possibility of plausible deniability. When you ask for a disclosure, you then give someone a choice: You can not engage in the behavior at all; you can do it and disclose it; or you can lie. Lying has substantial social consequences if you get caught, and it has a cost if you don't—most people find it stressful to lie, especially if you have to sustain a significant lie over a long period of time against people you interact with regularly. You have to spend mental effort keeping track of what you've said to people. You may need to involve others in lying for you.

Conventional morality, ethics, and law are all full of rules that are "unenforceable" in that they sanction private behavior which is hard for enforcers to observe, yet those rules continue to exist anyway. They often inherently give an advantage to cheaters by creating arbitrage opportunities around lying; but it's worth noting that if a behavior is undesirable, you will not make it happen less by normalizing it.

It's also true that people who are willing to lie about one thing are generally willing to lie about others, and unscrupulous behavior tends to be generalized. Which is to say: up to a point, additional rules of behavior constrain scrupulous actors who want to follow the rules, but they also increase the lying burden on unscrupulous actors who are likely breaking other rules.

There are already numerous behaviors that game studios regularly engage in that we'd consider disqualifying (from an IGF, a purchase, talking positively about their game...) if we heard about them. Those forms of misbehavior are old—ancient, in some cases—and therefore the rules against them are ingrained.

Against the destructive advance of AI, principled people are laboring to create a standard of behavior, which is an odd situation that tends to bring out a lot of nonsensical concern. But fundamentally, we don't have to be hostages to other people's misbehavior. We don't need to immediately give up just because an opportunity for lying, cheating behavior exists. We do have the option of resisting and building sites of resistance, instead of simply resigning ourselves to believing in nothing. In fact, if we believe in something, we have no option but to demand better.

A ★★★★½ review of The Secret Agent

A ★★★★½ review of The Secret Agent
The poster for The Secret Agent (2025

Reviewed by Bruno Dias

2025 · ★★★★½

I put off seeing this in theaters then when I went out to see it I fucking twisted my ankle on the street instead so I ended up not seeing it, and having become Sad about it I proceeded to put off watching it for a year. That's my excuse.

Too shaggy and off-kilter to call a "perfect" movie, but inescapably great. Just one of those movies that has another incredible thing behind every corner. Kleber if nothing else proves to be one of the greatest to ever do it simply with how great every last performance is in this movie.

All Posts