18
votes
They plugged GPT-4 into Minecraft – and unearthed new potential for AI
Link information
This data is scraped automatically and may be incorrect.
- Authors
- Will Knight, Andy Greenberg, WIRED Staff, Shayla Love, Amy Martyn, Chris Stokel-Walker, Lauren Goode, Amanda Hoover, Joel Khalili, David Nield, Omar L. Gallaga
- Published
- Jun 1 2023
- Word count
- 649 words
I like the AI trend that has emerged. I just hope developers of games don't jump on it just to say they have "AI" in their game as part of a trend like the NFT bullshit. Outside of gaming though, AI chat bots like ChatGPT and Microsoft's Copilot stuff have been super impressive.
I think AI (or at least LLMs) have potential to make games' worlds feel a bit richer, like they could be used to generate random crap NPCs say in games like Grand Theft Auto to make the world feel more alive. But they'll never replace actual writing of plots and missions.
Yeah. That's how I see it, too. But I don't think they'll have a huge impact beyond just being a gimmick until the average gaming PC can comfortable run a model locally for those NPC's dialog so people don't need to play those games online.
If you want a game or two games that have actual AI in them you can check out supreme Commander two or planetary annihilation.
They are old RTS games who's developer that created the AI for them was a big fan of neural networks back in 2012/ 2013. He coded literal AI into the game to fight against the player and as a result when you play those games you're actually playing against a real AI, just like the ones they use all over the other places.
Now it's still 2012/2013 technology and it's a real-time strategy game so it's not exactly up to the standards we have today, but it's real and it's there and you can play it today.
Yeah the enemy AI in those games were great. I didn't play a lot of them but I did play some when they came out.
But that's a different kind of "AI" than what's being discussed here. This is machine learning big model stuff. Which is different than creating algorithms that run on the gamer's PC to dictate how NPCs and enemies in games behave under certain circumstances.
Absolutely different, I understand that, but they are both still matrix multiplication piles powered by gradient descent built to do tasks.
Fair. I just think the distinction is important. Especially since the AI we're talking about when referencing things like ChatGPT can't run on a user's PC in most cases and will require an internet connection to work.
Well hey, if you want that, and I'm not so much directing this to you but more broadly because I'm assuming you know this exists already:
I've been having a lot of fun running llms on my local machine and I've been buying way too many graphics cards to do it.
Check out this guy!
https://github.com/KoboldAI/KoboldAI-Client
And this guy!
https://lmsys.org/blog/2023-03-30-vicuna/
I strongly encourage everyone to check them out, because we need AI to not be ending up in the control of centralized companies like OpenAI, and the more people who use something like this the better.
Oh yeah. This stuff is fun. I did Stable Diffusion on my machine a while back. It was fun. My GPU couldn't output more than pretty small images though. Haven't tried any chat AIs though. I love to see this stuff being developed independent of the big companies and able to run without internet though.
If you're willing to deal with jank You can get some Tesla P40 GPUs which are about as powerful as a 1080 for $200, and they have 24 gigs of ram so they can run pretty decent AI models, and you can get nice big images out of it in stable diffusion.
However, you're going to need a 3D printer and you're going to need an adapter to get it to be powered and it's lots of terrible fun and it's only good if you want to deal a lot of trouble.
And if you're willing to deal with slow responses, you can LLMs on your CPU as well with something called kobold c++.
Yeah I think that's more than I want to deal with for now. But I'll definitely check in again when things progress a bit more. And maybe I'll buy a new GPU with more than 8 GB of VRAM in the next year or two.
Today's AI will be like Ultron of old marvel comics. Once you defeat it in one way, you can't defeat it twice with the same technique. It learns and never forgets.
Terrifying, actually.
As far as the idea that AI will replace a lot of office tasks - I think they're right. I expected this to happen first with self-driving cars. I thought we'd see millions laid off because trucking could be fully automated. But now I'd predict that millions will be laid off because menial office tasks will be done by AI for less than the price of a human. It might even get so bad as to cause a major economic or social event.
I don't think GPT-4 will be the model to do it. But imagine a model trained on a million hours of annotated data entry footage with key-logging.
I don't expect it to look like a major economic event (or at least, not a sudden one) because it will happen over many years.
Businesses have been using computers to automate jobs for decades. For example, when the Internet took off, businesses started to build websites that let customers interact directly with the business's computers. It's a big improvement over processing paper forms sent in by mail, or standing in line to do it. But despite the hype of the dot-com era, it took many years to do. Some people didn't start banking online until the pandemic hit.
It's more difficult to change business processes than we imagine when thinking about them in the abstract. There will be delays and failed projects.
I'm thinking we'll see a plug-and-play solution that replaces the PEBCAK with an AI-driven VM.
Sure, but who would dare use it? Maybe scammers would be first.
I think a lot of people would use it if they're already using something like Mechanical Turk.
We could think of it as much cheaper competitor for Mechanical Turk and similar crowdsourcing markets. Looks like Mechanical Turk might have about 100,000 workers worldwide?
Some of that is human-subject research or generating training data for machine learning, so it wouldn't be replaceable. (Unless a worker is using it secretly.)
I'm witnessing AI replacing people in small pieces already at my job. I'm in marketing and we had an agency develop a Microsoft Excel Online spreadsheet and script for generating tracking URLs for our ads. The generator needed updating and instead of going back to the agency to do the work, we just asked ChatGPT to do the code updates, and it worked great. It was a small job, but it clearly took it away from a human in that instance, and I could easily see that happening more and more frequently as companies begin to recognize the possibilities and the tools for enabling it become more widespread.