Are there examples of AI generated content where people like it for something beyond the fact that its AI generated or its a "shocks content"?
Generative AI is cool but the novelty of started to wear off IMHO and when the AI content is looked without the "wow AI made this OMG" glasses, the outputs are blunt and uninteresting.
I've been thinking about this too. I don't know of any such examples.
I guess I've not really seen anything "novel" come out of it. So it's cool that "AI" is remixing the training data , but it's fairly boring 5 minutes after that.
Great question. We've seen "AI Seinfeld" and nonsensical machine-generated kids' videos on YouTube featured on HN, and the comments are usually in th3 vein of "just think what will be possible in x years".
What, indeed? More grist for the entertainment mill that no human will ever voluntarily sit through?
6K view and It’s clearly marketed as AI generated. Is there any clue that people are watching this because of its qualities as a film and not as another demo about “OMG AI will destroy Hollywood next year this time around”?
I don’t believe that the hardship of making content is about it being pretty or resemble something.
> People are making AI-generated films that rival anything you'll find on television.
No, they aren't.
> Here's just one example.
There are some decent frames in that, but as a film it is atrocious. Heck, even the individual animated clips that are stitched together aren’t particularly good. It does not “rival anything you'll find on television".
So is it safe to say that any concern about what AI should and shouldn't do has no gone out the window at Google and they will release (or show, I know this is just a paper) anything they can at the risk of being left behind by OpenAI?
Like seriously, why was this worked on?
I don't really buy that the goal is to train other AI based on this generated AI. Is that really valuable? That seems... like the end result will be pretty bad.
Eventually we will all trying to sell gear on the street, with a couple of select druids being allowed to work for the big brother corps, with a couple of leetcode trials during the planets alignment phase.
More seriously, I think the reality is that software dev as an industry will go through a shrinking phase that will be materially significant. Right now you probably only feel it if you are an outsource consultancy in Latin America or something. But eventually as an industry, we'll go through a drawdown everywhere.
That makes sense, and it will be necessary. But I'm equally sure that people who used to be software devs will find something else to do. Some will get enterprising and start using the AIs to generate their own products. Some will go into management. Some will turn to other industries. AIs generating things will make the pie bigger, and there will be more ways to cut something out for yourself.
It's incredible the number of side-projects that Google can afford with its advertising revenue, I bet this will never generate them a single cent, the folks creating it will probably make big bucks and move on their careers to create a company that will do this better and make others pay for it.
It's a company really needing activist investors that have the billions and billions to kick the CEO and executive team out. Also split its business. Imagine how many business lines from Google would be more efficient if they would operate on its own.
I hope they do this, I'd totally invest in Google. Before they do that, I wouldn't, as it would be considered charity for those making $1M comp making useless moonshot projects.
Not to put too fine a point on it, but the vast majority of game makers are putting out Really crappy games nobody wants to play. Only a relative few make the hits we hear about. There is still some money in that long tail of crap, and the AI companies will get better and better at gathering those pennies to themselves over time. I even think they'll go upmarket soon. That is, I think they'll get much better at generating even the hit blockbuster type games.
Exactly. Not even talking about app store which has a wide selection of games, if you just look at xbox store, which is presumably of higher quality (there is a higher bar for publishing in Xbox store -- you can't just create a new account and start publishing), it has a ton of games that have bad reviews and don't make sense to me - including from publishers like EA. Well, some people are playing those games, and EA is still around and doing well. Sounds like the person doesn't really play games much or understand how this works.
It can be entertaining to go on the store on the Nintendo Switch. Before owning one I assumed that it would be high-quality curated games that meet some minimum standard for Nintendo to allow it on their ecosystem.
Nope! It’s a sewer of $0.99 games that obviously prey on people thinking they’re getting a popular game like “Grand Theft Auto V” and not “Grand Auto Theft: V Edition” which turns out to be a top-down mobile game using the same fonts and imagery.
>It's incredible the number of side-projects that Google can afford with its advertising revenue
Part of the reason they are affordable is because they are built on the existing shared infrastructure. It would be much more costly for an independent company to spin up all of the necessary infrastructure, frameworks, practices, etc.
This is basically the line of argument used to push for layoffs in 2022-2023[1].
From an investment standpoint, Google is a curious play, as it shares a massive ad market duopoly with Meta, yet invests billions in dubious product lines with uncertain payoffs or futures as leads move out to move up. The graveyard of cancelled Google products has been documented to the point of parody.
Personally I think the whole executive team is out to lunch, too comfortable with their printing press to do much if anything about Google’s structural dysfunctions.
I wonder if there's more variety in game. From all the examples in the blog and release, it looks like the model is segmenting static background, non interactive background, platforms and the player. And just builds the game to allow you to control the player and interact with the platforms.
I guess it's impressive if it inferred physics and the idea of these abstract concepts organically, but overall doesn't seem overly impressive to me compared to the things I've seen recently
> Similar to Sora, Genie's creators call it a "world model," but unlike Sora, it's an "actionable-controllable world model.
This is talking about 2D platformer games and saying that the world model is actionable-controllable. I wouldnt be surprised if it's just a 2D platformer level where you can browse it WASD and maybe hit the walls with some pretend character.
People who started playing games during covid will probably think this is groundbreaking.
(we have already had this for a few decades, it's called "procedurally generated levels")
I am starting to resent these things more and more - the old phrase "you were so concentrated on wondering if you could, you didn't think about if you should".
Feels like there are loads of researcher's trying to replace people with AI in every possible avenue. While I am sure it is a technical feat etc, I am not enjoying how this is going.
I hope these researchers are able to sleep at night. They are already causing pain for people's livelihoods and yet they just keep on going trying to find more ways for more people to get sacked.
Technology and science are not and cannot be good or bad. They are amoral. The brains deciding what to do with it carry the moral burden. The answer is never to understand less about reality.
Sure, but not all tools have an equal ethical space in which to be utilised. A tool made specifically for torture (say a device to painfully remove fingernails) will tend to be used for morally bad things (using a reasonable assumption that torture is bad). Could it save the planet by torturing a specific person in a specific situation? Could it help remove that deadly fingernail from a patient? Sure, but its space for evil acts is much broader, and very different to say a ventilator, or an Xbox.
The reality is that the potential of science and technology to achieve things seen as good and bad does differ, and the magnitude of that power.
When technology advances we should advance the way our society is set up (which a lot of people forget is possible). We shouldn't try limit researchers from trying to come up with new stuff because all of science is related in some way, these advancements related to ML might help cure cancer in the future.
Most of the current advancements in AI research evolved out of teaching AI how to play games. Gameplay and AI research are very closely intertwined, this research project only helps further push the capabilities of AI. Another small step on a very long march ahead.
Yeah, but in the future, I think we'll see our current dissatisfaction with content discovery the same way I see my pre-social media dissatisfaction with computer privacy.
I would be curious on how well it builds an immersive environment. Early stages might be useful for LEOs to train in VR in a wide variety of settings. No need to hire or outsource to a third party. Or build/maintain expensive real life replicas.
Consequently, also allows for pieces of shit to train as well. Got to take the bad with the good, right?
This seems to be the problem with generative ai. It’s unable to tackle the hard problems due to a lack of massive data sets and existing solutions?
All it is good at is solving things humans have already solved. When I see this I definitely wonder why we aren’t researching bigger issues that really affect people.
I don’t think the screenshots (that it uses as training data) contain enough useful information on what makes a game enjoyable to play, to make this an immediate threat to game designers.
When AI will be able to generate credible stories with credible characters then put them into credible worlds whose data can be fed to a Open Source game engine, we'll hear a great disturbance in the Force, as if thousands of game designers and publishers suddenly cried in terror and were suddenly silenced.
On a serious note, it seems to me technology isn't that far from that point.
> On a serious note, it seems to me technology isn't that far from that point.
It feels at least a couple of decades away.
I think the most compelling use for LLMs and generative AI right now is parsing massive tomes of knowledge, which can then be summarised at use-time.
Like imagine pilots being able to consult an LLM with all of the aircraft's blueprints, technical reports, operational history, etc. during flight. Likewise for spacecraft when debugging in flight.
But the internals really need to go beyond just token prediction, to being able to use Abstract Syntax Trees for programming suggestions with correct types, or building ontology trees for answering queries factually. The former is a major frustration with Copilot (when it gives relevant suggestions at all).
chatgpt was released 1 year ago, the transformer is 7 years old, I feel quite the opposite, and also I am quite excited and frightened at the same time.
Have you seen Sora lately? Every few months we get a new big AI innovation or major improvement. We are not in an end of Moore's Law phase in AI, we essentially just started, and the growth is currently exponential not linear.
> On a serious note, it seems to me technology isn't that far from that point.
Perhaps, but because it is hit-driven, I can easily believe "thousands" being impacted even when the top-10 appears unchanged by the reaching of this standard.
Credible != compelling. AI generated stuff is garbage, it's so far away from generating something exciting, it's as far as it was 10 years ago, the only difference is that now it is able to generate garbage (instead of doing nothing).
Even a kid can have better ideas than an AI fed with billions $ of nVidia hardware and data coming from unpaid labor.
It's really hard to imagine that an AI-generated storyline would be worse than say most Bethesda questlines of the last 10 years or so. Or say Ubisoft circa AC:Valhalla. There are so many storylines and quest objectives in games these days that are more or less:
Go here and get a pencil. <watch a cutscene>
Ok great. Now I need 3 more pencils. <watch another cutscene>
OK now go here and follow an NPC for ages while they walk really slowly. <watch another cutscene>
OK now shoot the guy <watch a cutscene>
OK we're done, have a cookie. Any time you want to, come back here and I'll send you to go fetch more pencils.
I only had the blue pencil after the first quest. I had to go back. Not for the cookies really, but rather because I needed the yellow, purple and especially the red pencil. Only then would I be the level 10 mage capable of retrieving the black pencil.
You act like they were making a fool out of me or something. /s
There was a parody video a while back, can't find it now, about RPGs from the NPC's perspective. PC steals all their stuff and also their grandmother, ends up receiving the sword of grandmothers.
I don’t know. There are a lot of popular games that are basically made to a formula. That said, I don’t see this being possible any time soon for non programmers (with the aid of ai). Consider the amount of debugging needed.
If the computer can actually program coherently, without bugs, and understanding the program as a whole and how the pieces fit together, then it can program itself.
I don’t see this being possible any time soon for non programmers
I don't see this being possible at all for non programmers. If AI companies have models that can generate games in 5 years, there's no way they will let other people use them. The companies with game generating models will simply become the developer and publisher.
If they can generate games, especially those deliberately addictive casual games, then they have a money printing machine. They'll just hire psychologists, economists, and testers to see which games are the most addictive and extractive- er- umm- I mean, which games are the most fun and compelling. Then they'll release those games. Further flood the games market with tripe, and probably get pretty rich doing it.
I hear such comparisons quite often on discussions like this, and I always wonder… have you seen a kid's idea of "creative"? Even one about to finish school?
One time my geography homework was a cut-out of a photocopy of an encyclopaedia image of Italy, surrounded by notes copied verbatim from that same page, and it got top marks. My GCSE English "creative writing" exercise was a mixture of the empty village in Riven, Pratchett's Granny Weatherwax, and the "oborot" shape-shifting spell from Sabine Baring-Gould’s Book of Were-Wolves. When we were tasked with coming up with our own versions of the witches' brew from Macbeth — why is Shakespeare sill so popular? Don't we have any new ideas? — all the students came up with much the same "new ingredients" as each other.
I think all of this AI generated content still has a ways to go until it can be stand alone content, but when it is being used as a supplement to human creation, it can be fairly useful to either generate a starting point, or to fill in the blanks.
BUT...I am seeing a lot of AI generated content pop up in graphic design(emails, work decks, etc...) and some of it is terrifyingly bad.
Humans create shit stories too, and at scale! Just look on Steam at the deluge of shit that humans create.
Less is more.
With that said, I agree that it could be awhile before an AI generates something compelling.
Edit: As an aside this touches on a capability that I was wondering about. Suppose you use Sora to generate a video, how do you fix the issues? It would be nice to rather than generate a video, generate an animation project or something similar so the issues can be fixed more easily. Generating video games seems like a step in that direction.
These guys just humiliated themselves over their static image AI and just recently burned every single bridge and shred of faith the gaming industry had for them with Stadia.
This is going nowhere. It's quite literally non-news.
As someone who has worked in video games both at the hobbyist and AAA level, this is a demo meant to fool non-experts. There are a million tutorials to get a really shitty version of mario running in unity or godot or unreal. And they openly say that they used these shitty mario clones to train this model.
This will not scale in the near term because
a) There aren't that many types of games that get cloned over and over again with small variations that they can train their data on. Mario games, card games, toy 2d strategy games, etc. Even something as basic as the rougelike genre has so much variation between games in style, feel, gameplay loop, system design, online play, etc that I don't think you can mix them up in an algorithmic slurry and get anything legible out on the other end.
b) The output of this so far looks ... really bad to be frank. It's unacceptably low res, no anti aliasing, you immediately see that this is just a sprite superimposed over the provided images. It looks so terrible that they are not getting buzz from word of mouth (the tweet thread has very much not gone viral despite being up for 3 days now).
c) Getting a character to move left or right and jump in a 2d scene with some collision detection is the only thing this has done and that is the easiest part of making a game. I have seen average 13 year olds manage it. There is so much technical depth in games and game engines beyond getting a sprite to render and move. They don't even have a walk animation in this demo. Good luck ever getting this to create a double buffer for rendering. The game industry very jealously keeps its code internal and its gonna take more than the payouts reddit and tumblr have been getting for ATVI/EA/Ubisoft to give up their source to Google. If Google had kept Stadia around, that actually could have provided them with so much free training data from the images being passed back and forth, but since Google is a non-functional company we don't have to worry about that :)
d) Bugginess in a game, once over a certain threshold, causes players to drop the game really quickly. I cannot even begin to imagine the amount of bugs or what the process would be to fix them in this mess.
You could create the same thing back in the day with a webpage where users can submit a set of background tiles and then run a flash mario game superimposed over those images.
Generative AI is cool but the novelty of started to wear off IMHO and when the AI content is looked without the "wow AI made this OMG" glasses, the outputs are blunt and uninteresting.