Genie 3 is out the bottle. New AI development.

Everyone´s talking about gaming, but this could seriously become important for movies, like the use of Unreal Engine right now, or in medicine, desaster management (imagine a VR situation where you have to rescue someone from an earthquake and your instructor can generate new obstacles along the way for you to react to).
 
When it's all said and done? It already is. It is practically a utility at this point.
Nah. Matter of fact, I change my mind. It's going to be nearly impossible to knock off the top 3 inventions of all time.

1. The harnessing of electricity
2. Medicines
3. The wheel
4. Writing
5. The Printing press

Ai is not a top 3 yet and I don't see anything ever knocking those off.
 
Last edited:
It's still generating video ... The energy and computing resources needed to create a "dynamic world" for just 1 person is enormous. Now extrapolate that to thousands, tens of thousand or even millions of people ... That's just ridiculous.

What game developers want is AI that generates assets: 3D worlds, fully features 3D models, textures, etc etc etc.
Generate everything once so the whole world can play a game without needing as much energy as the sun.

Give it 2-5 years.

People don't realize how much everyone is going to change by 2030.

Create your own games on demand, etc.
 
Buckle up, a few more iterations and you have a universe wide infinite sandbox with any fantasy or time periods you can think of to play in
 
Give it 2-5 years.

People don't realize how much everyone is going to change by 2030.

Create your own games on demand, etc.
Maybe. I think the problem is that we cannot scale energy fast enough to support this, unless we start building coal plants again.
 
As inevitable as flying cars.

Seriously, this is some bullshit. If by "games" you mean absolute soulless trash, as everything created by AI so far, yes. If you mean a genuine videogame like Elden Ring or Expedition 33, the answer is no.

If you think Ubisoft or EA produce slop, just you wait.
In 20 years every game made will use AI heavily and the games will be much better for it.

AI is a tool. We do not need videogames to be completely automated but also we do not need to waste time doing work the computer can do for us.

Instead of thinking of AI as a human replacement, think of it as a new type of computer input. We have keyboards. We have mice. We have mics. Now we also have AI.

We can input concepts into the PC. So an example would be macros. Currently I can create a macro in excel and run it in a workbook. I cannot run in a million workbooks without writing code to do that which I don't have the ability to do.

However I can ask AI to do it and run that macro across a million workbooks with just one entry into the computer. No coding. Can you see how this would be useful for developing videogames? Because the PC can now understand more than just numbers and letters, it can understand concepts and act on them, we have a new input into the computer that can save a lot of time. It is a tool, not a person. It removes busy work. It allows us to extrapolate ideas across a whole build in seconds instead of weeks/months. Even if it does the job poorly it allows us to sample the idea in real time in the game to decide if we want to pursue it and began the process of work. It improves the game and is one of the most valuable tools we have already.
 
In 20 years every game made will use AI heavily and the games will be much better for it.

AI is a tool. We do not need videogames to be completely automated but also we do not need to waste time doing work the computer can do for us.

Instead of thinking of AI as a human replacement, think of it as a new type of computer input. We have keyboards. We have mice. We have mics. Now we also have AI.

We can input concepts into the PC. So an example would be macros. Currently I can create a macro in excel and run it in a workbook. I cannot run in a million workbooks without writing code to do that which I don't have the ability to do.

However I can ask AI to do it and run that macro across a million workbooks with just one entry into the computer. No coding. Can you see how this would be useful for developing videogames? Because the PC can now understand more than just numbers and letters, it can understand concepts and act on them, we have a new input into the computer that can save a lot of time. It is a tool, not a person. It removes busy work. It allows us to extrapolate ideas across a whole build in seconds instead of weeks/months. Even if it does the job poorly it allows us to sample the idea in real time in the game to decide if we want to pursue it and began the process of work. It improves the game and is one of the most valuable tools we have already.
I think this gets it mostly right.
However, AI will be and remain limited by the user.

AI generated fake-games/video like in this thread? Nope. Not real. Not simulated. Not shareable. Not directable enough.
AI-assisted traditional workflows (programming, animating etc.?) Absolutely. We see some of this already nowadays, although pretty limited.
Cascadeur for example uses AI for its new Inbetweening feature, that generates motion between poses. Still somewhat limited, but great for what it is right now.
AI assistants coming to UEFN, helping you generate Verse code to make games in Fortnite.
ChatGPT and variants can write some small amounts of proper UE-code (emphasis on the small).

However: AI aka LLMs doesn't 'understand'. It's not just an input <-> output mechanism. It's a black box. You can give it your intent and hope for the best, unlike a mouse or keyboard.
AI gives you approximations to what you want. Sometimes those reach 100% accuracy and therefore correctness, but whenever the use case gets a bit more advanced, you will forever be limited by your own capabilities.

If you know a little bit of Javascript, and you generate code for a Snake-type of game, you are able to understand the generated Javascript code, and make tweaks as needed, and even expand it a little by using AI and your skills.
You will get somewhere as a beginner. But that's all it can do for you. Any kind of topic that has nuance and technical details, such as game dev, AI will get wrong, 100% of the time.
And on a topic where correctness is required, you can not rely on AI.

Code that doesn't compile won't mean you have 99% of a game, it means you have 0% of a game until you fix the issue. And if AI generated the wrong code, and you aren't smarter than the AI, you won't be able to build a game, at all.
The way I see it, the barrier to entry in some fields is going to get lowered, but the ceiling will remain as high as ever, and AI is not going to help with getting to that ceiling because you are offloading your mental process and understanding and achievement to a separate entity.

So at most, AI is going to help automate tasks that a professional already knows how to do. It will help with iteration times and R&D, but will be limited by what its user can do.
 
Quantized model quality isn't quite the same though. I have used a variety of Qwen and DeepSeek versions and they are ok, but nowhere close to full on cutting edge modern models, IMO.

And things get a lot worse once you go beyond light coding assistance and text generation.

I am not any sort of an expert here, but the further we go, the more "power hungry" the more advanced models seem to get, especially if you throw reasoning in top and then agents into the mix.
Quantization is extremely good these days, though, thanks to very clever techniques used during the compression .

For a quick demonstration -- because visual is easier to spot the general alignment of performance -- this is the same prompt on FLUX (transformer-based image model), where the left was full precision and right was 8-bit quantized version. Despite massive reduction in model size, the final image has only very minor differences that are not even errors so much as just tiny differences in random sampling result.

Numovs.jpg

(and there are also "distilled" models, a somewhat different set of techniques from quantization, but similarly making great strides and managing to compress levels of understanding that you could never train on a small scale, but we're finding that you can "train big then shrink into smaller student model" with little loss)

Similarly, something like Qwen3-Coder is extremely intelligent and useful under quantization. And the trend is that better and better models can run on consumer hardware, particularly with chips like these Silicon series laptops from Apple with integrated memory that increases each gen.

The concept of "GPT4 at home" is already realizable, and was a pipe dream a few years ago.
 
Last edited:
There is no guarantee AI will continue to improve exponentially to do some of the things people are saying.
Just because it can currently generate a somewhat consistent video feed that allows you to loosely navigate doesn't mean it will be generating GTA10 one day.
Like in 1969, after the moon landing, lots of people thought we would be landing on other planets in the not to far off future, considering the rapid progress we were making. Obviously that didn't happen though.
 
There is no guarantee AI will continue to improve exponentially to do some of the things people are saying.
Just because it can currently generate a somewhat consistent video feed that allows you to loosely navigate doesn't mean it will be generating GTA10 one day.
Like in 1969, after the moon landing, lots of people thought we would be landing on other planets in the not to far off future, considering the rapid progress we were making. Obviously that didn't happen though.
You're comparing software progression to a much more complicated physically limiting problem (moon landing, space travel).

Most professional developers today are actively using AI to generate production level code. This article is from 2023: https://github.blog/news-insights/research/survey-reveals-ais-impact-on-the-developer-experience/
 
In 20 years every game made will use AI heavily and the games will be much better for it.

AI is a tool. We do not need videogames to be completely automated but also we do not need to waste time doing work the computer can do for us.

Instead of thinking of AI as a human replacement, think of it as a new type of computer input. We have keyboards. We have mice. We have mics. Now we also have AI.

We can input concepts into the PC. So an example would be macros. Currently I can create a macro in excel and run it in a workbook. I cannot run in a million workbooks without writing code to do that which I don't have the ability to do.

However I can ask AI to do it and run that macro across a million workbooks with just one entry into the computer. No coding. Can you see how this would be useful for developing videogames? Because the PC can now understand more than just numbers and letters, it can understand concepts and act on them, we have a new input into the computer that can save a lot of time. It is a tool, not a person. It removes busy work. It allows us to extrapolate ideas across a whole build in seconds instead of weeks/months. Even if it does the job poorly it allows us to sample the idea in real time in the game to decide if we want to pursue it and began the process of work. It improves the game and is one of the most valuable tools we have already.

20 years?

No, more like 5-10 years at most. The cost savings are to significant.
 
End all scarcities.

Let AI generate in seconds what would take us years and hundreds of people to make.
People who are against that don't understand how liberating it will truly be.

Agreed. Now excuse me while I generate bespoke celebrity pornography and animate it in ways I don't even want to understand.
 
Last edited:
So at most, AI is going to help automate tasks that a professional already knows how to do. It will help with iteration times and R&D, but will be limited by what its user can do.
Fully agreed, this wont ever replace game engines - but we will see AI getting more integrated in the form of neural shaders / models generation.
 
Last edited:
I'm always cautious about these wild claims. Real world implementation is always the hard part.

However, I hope soon we come to the point where one person can create something like expedition 33 (with outsourcing) soon though

What game developers want is AI that generates assets: 3D worlds, fully features 3D models, textures, etc etc etc.
How would that be different from just putting in UE5 assets? Isn't that even easier already?
 
I'm always cautious about these wild claims. Real world implementation is always the hard part.

However, I hope soon we come to the point where one person can create something like expedition 33 (with outsourcing) soon though

How would that be different from just putting in UE5 assets? Isn't that even easier already?

The idea is that AI will someday be able to help with every step of the game design process: textures, 3D models, complete assets. Just use a prompt or use a 2D image and an AI assisted 3D modeler will turn all of that into a complete scene. That will be a game changer and companies are already working on that.
 
Maybe. I think the problem is that we cannot scale energy fast enough to support this, unless we start building coal plants again.
Big corporations are already planning/building private nuclear reactors/solar farms/wind farms (all mixed together) to power new AI datacenters (faking sustainability).

If governments don't invest in this, private companies will do. It's basically inevitable.
 
Last edited:
Holodeck comes one step closer

Yeah but on the holodeck you could have a simulation as pornographic as you wanted, most AI now won't let you have any fun due to prudish devs. I mean yeah you can get AI software that generates porn, but it usually seems to be from dodgy sites.
 
As inevitable as flying cars.

Seriously, this is some bullshit. If by "games" you mean absolute soulless trash, as everything created by AI so far, yes. If you mean a genuine videogame like Elden Ring or Expedition 33, the answer is no.

If you think Ubisoft or EA produce slop, just you wait.

Flying cars was never going to work. The two are nowhere near comparable. AI game creation doesn't require new laws, sky highways, or battery breakthroughs to work.

In fact, we're very close to AI generated games now. Tools like GPT-4, Gemini, CoPilot, and Mistral can generate code, dialogue, game mechanics, and world-building from text prompts. The path to creating full games with text prompts isn't that far, especially with the speed AI science is progressing.
 
Flying cars was never going to work. The two are nowhere near comparable. AI game creation doesn't require new laws, sky highways, or battery breakthroughs to work.

In fact, we're very close to AI generated games now. Tools like GPT-4, Gemini, CoPilot, and Mistral can generate code, dialogue, game mechanics, and world-building from text prompts. The path to creating full games with text prompts isn't that far, especially with the speed AI science is progressing.


Yeah I get sick of the never ending false equivalences. There are tons of things that worked and didn't, so the idea that every potential big thing hinges on the things that didn't work is silly.
 
Yeah I get sick of the never ending false equivalences. There are tons of things that worked and didn't, so the idea that every potential big thing hinges on the things that didn't work is silly.

This.

AI is advancing at such a break neck pace. I can see people being able to create any game they can think of with just text prompts within 5 years. Not just games either, but AR/VR simulated worlds as well.

I can see this being fantastic for education. What to see what Rome was like in 2nd century? No problem. Type this request into a prompt, stick on your VR headset and you'll be able to take a walk around a simulated ancient Rome that looks real.

This is tech that we can see developing in real time.

Flying cars will never become mainstream because there are too many problems that are frankly either impossible or far too costly to solve.

Could you imagine the chaos if millions of people were zipping around in flying cars? No lanes, no stop signs, no traffic lights, just people flying about everywhere. It would be carnage. It's sci fi fantasy that was never going to work in the real world.
 
This.

AI is advancing at such a break neck pace. I can see people being able to create any game they can think of with just text prompts within 5 years. Not just games either, but AR/VR simulated worlds as well.

I can see this being fantastic for education. What to see what Rome was like in 2nd century? No problem. Type this request into a prompt, stick on your VR headset and you'll be able to take a walk around a simulated ancient Rome that looks real.

This is tech that we can see developing in real time.

Flying cars will never become mainstream because there are too many problems that are frankly either impossible or far too costly to solve.

Could you imagine the chaos if millions of people were zipping around in flying cars? No lanes, no stop signs, no traffic lights, just people flying about everywhere. It would be carnage. It's sci fi fantasy that was never going to work in the real world.

Oh, its going to be bonkers. I would love to try that drone in VR. It just mindfucks me that I can explore such a massive environment that looks photorealistic, but isn't real, and with breakneck speeds. If its already able to do that? The sky is the limit.
 
Oh, its going to be bonkers. I would love to try that drone in VR. It just mindfucks me that I can explore such a massive environment that looks photorealistic, but isn't real, and with breakneck speeds. If its already able to do that? The sky is the limit.

Still early days and I'd be skeptical on how good the experience is now, but soon you'll only be limited to your imagination.

Holodeck from Star Trek is too far fetched, but with a VR headset and haptic feedback it'll be the next best thing.
 
I think this gets it mostly right.
However, AI will be and remain limited by the user.

AI generated fake-games/video like in this thread? Nope. Not real. Not simulated. Not shareable. Not directable enough.
AI-assisted traditional workflows (programming, animating etc.?) Absolutely. We see some of this already nowadays, although pretty limited.
Cascadeur for example uses AI for its new Inbetweening feature, that generates motion between poses. Still somewhat limited, but great for what it is right now.
AI assistants coming to UEFN, helping you generate Verse code to make games in Fortnite.
ChatGPT and variants can write some small amounts of proper UE-code (emphasis on the small).

However: AI aka LLMs doesn't 'understand'. It's not just an input <-> output mechanism. It's a black box. You can give it your intent and hope for the best, unlike a mouse or keyboard.
AI gives you approximations to what you want. Sometimes those reach 100% accuracy and therefore correctness, but whenever the use case gets a bit more advanced, you will forever be limited by your own capabilities.

If you know a little bit of Javascript, and you generate code for a Snake-type of game, you are able to understand the generated Javascript code, and make tweaks as needed, and even expand it a little by using AI and your skills.
You will get somewhere as a beginner. But that's all it can do for you. Any kind of topic that has nuance and technical details, such as game dev, AI will get wrong, 100% of the time.
And on a topic where correctness is required, you can not rely on AI.

Code that doesn't compile won't mean you have 99% of a game, it means you have 0% of a game until you fix the issue. And if AI generated the wrong code, and you aren't smarter than the AI, you won't be able to build a game, at all.
The way I see it, the barrier to entry in some fields is going to get lowered, but the ceiling will remain as high as ever, and AI is not going to help with getting to that ceiling because you are offloading your mental process and understanding and achievement to a separate entity.

So at most, AI is going to help automate tasks that a professional already knows how to do. It will help with iteration times and R&D, but will be limited by what its user can do.

Your actual idea or reasoning against AI is pretty dated at this point. But go on
 
Last edited:
There's no paper published. This is basically just marketing based on "trust me, bro".
 
Last edited:
not exactly; they published their architecture for Genie 1 (https://openreview.net/pdf?id=bJbSbJskOS) and then the update posts (first Genie 2 and now Genie 3) are basically project updates

It's fairly transparent, and Deepmind in general a group that has contributed a ton of open research / knowledge over the years
Ah, I wasn't aware of the updates (weird way of doing that, they didn't even bother to update their arXiv upload of the paper). Do you happen to have links to the posts? I can't see any (I am logged in).

Edit: Oh you mean Google's blog posts. I thought you meant posts on OpenReview. Well... To me this isn't a publication. So they did a paper for Genie 1 and that's about it.
 
Last edited:
Nah. Matter of fact, I change my mind. It's going to be nearly impossible to knock off the top 3 inventions of all time.

1. The harnessing of electricity
2. Medicines
3. The wheel
4. Writing
5. The Printing press

Ai is not a top 3 yet and I don't see anything ever knocking those off.
I agree with all of that except: `5. The Printing press`
 
Last edited:
There's no paper published. This is basically just marketing based on "trust me, bro".

This.

I would be careful with getting excited for this. Image and video generation is absurdly expensive and it's unlikely most of them will have accessible prices for much longer. There's a reason why Veo 3 is only 8 seconds long.

Also, AI models are trained based on a lot of pattern recognizing models. It took a year for Chat GPT to be able to generate full glass of wine images because it required extensive training.

Unless you curated a local LLM with the stuff you wanted, you will have a very hard time controlling AI generating the details or vision you want. This would be more useful to start or finish boring parts of a video game most people don't want to do.

I would suggest taking a look on the architectural limitations of transformers, which all LLMs are based off.
 
Last edited:
Top Bottom