This has been bothering me too. Its such a nothing statement with no suggestion of how this would be accomplished behind it. Like sure, technology is likely to continue advancing, but it may not even resemble these current implementations by the time we get there.
The man's a consummate bullshit artist, but I can't deny he used to be one of the best. Populous, Theme Park, Dungeon Keeper, Fable...
The last 10-15 years have been a bit of a write-off, but before that, even when he was shitting out games, they were still better than most people could even dream of making.
When an AI is able to take and write every single line of code, generate all the art, and debug it just by having someone as code dumb as a CEO type in a prompt, I'll believe his words and eat my shirt.
The thing is even if AI could do all that (which is doubtful in my life time), you would still need someone to prompt it with something interesting. And CEO types have never had an interesting idea in their lives
A really good place would be background banter. Greatly reducing the amount of extra dialogues the devs will have to think of.
Give the AI a proper scenario, with some Game lore based context, applicable to each background character.
Make them talk to each other for around 5-10 rounds of conversation.
Read them, just to make sure nothing seems out of place.
Bundle them with TTS for each character sound type.
Sure, you'll have to make a TTS package for each voice, but at the same time, that can be licensed directly by the VA to the game studio, on a per-title basis and they too, can then get more $$$ for less work.
They won't because of hallucinations. They could work in mature games though where its expected that whatever the AI says is not going to break your brain.
But yeah a kid walks up to toad in the next Mario game and toad tells Mario to go slap peaches ass, that game would get pulled really quick.
There is a possibility something like this will be possible in the future, but it's not going to be an achievement of AI, it's largely going to be the achievement of regular developers creating a general-purpose game engine that can be used to put together a game block by block, which can be utilized by both human game designers and AI. (Likely to better effect by the former.) I can imagine Entity Component Systems will play a big part of that.
One of the biggest blockers for AI making games is going to be testing it to select for better performance. With text it's relatively easy to see if some text an AI produced is plausible. Images are also plentiful, but that's a lot more subjective. With both of these it would also not take a massive amount of time to add a human element. It's quick to check if a paragraph or image looks like it is a good response to the input promt. A game, however? How long do you need to play it to see if it's fun? At best, perhaps, you can write an AI to control a bot character to see if it's technically playable.
I don't want to even think about the electricity that wlll be wasted training such models.
ECS has really nothing to do with this. ECS is just a specific way to store the internal state of a program, fundamentally no different from other data structures.
Also, a good game is far more than just text and images and current "AI" can't even generate those individually. A game needs significant thought put into things like game loops, story arcs, balancing,... that are non-obvious when existing games would just be training data. Not to mention that using an existing game as training data is both non-trivial and also we just don't have the vast amounts of them that current systems seem to need to produce anything even half-way decent.
ECS already makes it a hundred times easier for me to conceptualize game mechanics, modify and extend them. Giving AI the ability the ability to create data separate from systems that use them will make it much easier for it to build a game. I don't believe for a second it will be able to write functioning object-oriented game code for example. It will likely be best if it avoided coding via a text-based language altogether, and use visual scripting or another system based on chaining logic blocks together. But that still counts as the "system" part of ECS.
25 years might actually be a potentially realistic timeframe for that. But I'd still bet on a little longer myself. I hope it happens within my lifetime.