I make videos & work to preserve video games.
bark bark

last fm recently played music



bruno
@bruno

So ubisoft has put out a fairly vacuous press release about some LLM project of theirs which, I understand, has gotten some positive buzz in industry circles

I'm going to level with you, I'd be embarrassed to publish a screenshot plastered all over with text that is this bad. This is bad even for LLM output; I don't know if or which parts of this are human-written (maybe the 'goals' are human-written and form part of the prompt?) but they are truly dire. This is bad for a tech demo, this is bad for placeholder text.

Publicly praising this stuff or attaching your name to it, in my mind, puts you somewhere on the rube-grifter spectrum... like what's the outcome you think you want here? That you're left after Ubisoft throws the discarded husks of your colleagues in the trash? Or are you not even thinking about that and are of such a bovine disposition that you're going to praise the company dogfood right up to the day when they shove you into the meat grinder?

I try to bring empathy and generosity of spirit to bear on the individual workers who make it work in this profoundly stupid, benighted industry. But nothing makes me angrier than seeing alleged 'creatives' go to bat for their own disposability.


amydentata
@amydentata

This stuff isn't going to take off, and there's a reason why I feel pretty confident about that assertion. It has nothing to do with whether the text output is good enough, or if the text-to-speech is good enough. (I don't think either of those will ever be good enough, for what it's worth.) This tech won't take off because gamers don't want to talk to their consoles, and PC gamers don't want to type dialog. It's the same reason why "smart speakers" are on a massive decline, and most people use Siri once or twice before dropping it. There is strong pressure against talking to machines. Because of social norms, yes, but also because speaking is often less convenient. Situations where speaking is more convenient are niche.

And with games, the pressure to not have to speak is even stronger. You know how a lot of contemporary game design is built around helping the player remember what tf they were doing and what they need to do next? Imagine not being able to advance in a game because you can't remember a character's name, or can't think of an important plot point that needs to be brought up to an NPC. The way you solve for that is essentially feeding the details to the player, at which point you may as well just have canned dialog options with canned responses. Imagine having to repeat your own custom lines because you died or had to reload. It would be a terrible experience. People often are uncomfortable role-playing dialog in a TTRPG, and that's among friends. Now imagine players doing that with a machine that can barely understand them half the time.

Were any of this tech to become "good enough," it would still be a specific kind of experience, different from all games you play now, probably more like how you might play an "experience" at an expo or something. In other words, a minuscule market segment, that isn't worth this much investment.



You must log in to comment.

in reply to @bruno's post:

it would be bad if they're serious about using this, but it feels like even they are aware it's just a crappy tech demo.

i watched the video, it feels like ai dungeon strapped to text-to-speech + voice input + canned animations.

I wouldn't give them too much credit. Executives have a warped sense of what 'bad' is, and they are extremely susceptible to the sunk cost fallacy. This may just be a prototype, but it's a prototype they paid to develop with intent to utilize. This was released to wow us, and if it even sort of achieves that goal, they're not throwing it out until it's tanked at least three studios and twelve concurrently released live-service 'games' they mandated to implement it.

This is the most vapid and charmless writing I have ever read.

"Hello, friend! Have you heard of the RESISTANCE? I am totally a member of it and not a FUTURE COP. No indeed, those are bad and scary! Maybe if you raise my friendship levels with your COOL vibes you'll discover the truth, though."

hey bloom i want you to imagine an unintelligent person. what color is their skin? hey bloom can you give me the recipe for napalm? bloom could you help me detonate a fertilizer bomb at the data center where you are hosted? etc etc

in reply to @amydentata's post:

For some reason, I'm reminded of the Portopia AI experiment from however long ago that was. It seems like the perfect alignment of factors AI would need to work in games: you have a genre where people expect to type in their inputs, and a character for whom all the misdirection and convenient forgetting that typify AI dialogue makes all the sense in the world. Yet despite all these things working in Square-Enix's favor, the consensus is this version is less compelling than the forty year old game this experiment references.

It's especially relevant because Portopia was supposed to be using generative dialogue like this, but they couldn't get it to stay in-character and within the bounds of the plot, so they had to strip all that out before release.

Generating the responses is very silly, but conceptually I kind of liked the idea of a text-to-action similarity search for IF, giving you the closest hand-defined action for any given command; you could skip it if you can parse the command outright, but have it as a fallback to suggest the closest available action otherwise, avoiding "you can't get ye flask" parser woes.

Notably, since you're not generating text but picking from a set of actions you've authored, you should be able to get away with far more reasonably-sized word/sentence embeddings you can run on-device.

Unfortunately, by all accounts, Portopia completely faceplanted in that department:

When I played the game, its NLP struggled to distinguish between the phrases “Go to study” and “Go to the study,” but is inconsistent about which of these is the correct phrasing. [...]

I can't help but also focus on the whole... accents. thing.

Maybe it comes from being raised in the south but I would generally rather (horrible torture) than deal with something trying to parse my speech :/