Generative Agents - Deep Dive and GPT-4 Recreation

Generative Agents - Interactive Simulacra of Human Behavior - Deep Dive and GPT-4 Recreation
In this video, I look at the paper and also look at how some of the ideas can be used with GPT-4.

For more tutorials on using LLMs and building Agents, check out my Patreon:

My Links:

Github:

[filipewnunes]

Looking forward to create a big house with 20 agents, powered by Auto GPT, all top programmers, all with access to virtual machines, IDEs and a bunch of libs and frameworks, with the main quest being to develop a AI Generative Agent.
Save all experience, test the newly created agent, add it to the 20 prior, repeat.

[countingtls]

I've tried to recreate it with various public models (Vicuna, Koala, Dolly, Alpaca, LLaMA), there seems to have a pattern where only LLaMa model (13b above) can constantly generate stable output one entry at a time. All other models, even dolly v2 will simply output a stream of time steps and actions disregarding the "realistic time" (like 7:01 open a restaurant, 7:02 take a break, ... and close it 7:18, and sleep at 7:19). They need very specific prompts to "hold" their generation for one step. Usually without time stamps.

Even LLaMa model's output is a bit repetitive and short, and not as creative as ChatGPT (even GPT-3 models from APIs have some issues). It seems that the model has to be optimized not only for tasks, but also conversations in order to generate more "natural" and plausible actions. Still trying to figure out how to implement "effective summaries" for with these limited LLM. The models that can generate entries don't seem to be able to "rate" events (more "chatty" models seems to be able to somewhat generate rating, but not very consistent).

[sitedev]

I can see someone making this into the new version of a Tamagotchi where if you don't deliver resources to your 'colony' bad things happen or if you provide an abundance of resources even worse things happen.

[abhijitkadalli]

This is so cool discovering your channel was the best thing that happened to me ♥️

[DivineLlamaLove]

This was such an awesome paper. I read it yesterday and thought Westworld! 😂

Thank you again, Sam.

[_________]

Very interesting. I’ll want to play these out using LangChain by giving the the Generative AI deeply detailed back stories and behavioral rules to check not only garnet to agent interactions but a UI to allow for users to connect to the “game” as well.

[patrickhenz]

A fascinating idea, and another example how fiction predicted or inspired potential reality. On of the first books related to the Metaverse had been "Simulacron-3" by Daniel Galaouye. Here the Metaverse was created for market research purpose, to test the reaction to particular products or ideas.

[PenicheJose]

great video, where I can see the example game is not in the paper.

[ripreal]

It's literally insane how fast AI is moving, the implications of this are so massive lol

[impolitevegan]

I can imagine future games being much more addictive than they are now. Imagine better VR, better graphics and realistic and interesting characters. Imagine, say, Witcher but the NPCs are much more intelligent and much more interesting.

[aaronward]

Amazing paper, hats off for the researchers. Thanks for covering this! It's exciting to see how this will be used to replace the standard Oblivion-like dialogue we have in games today. Using LLMs in conjunction with speech synthesis models will create a very immersive experience.

Also, anyone aware if there is a git repo for this? or is it closed source?

[TheInfiniteVoid]

Been saying it for years, eventually we will have "artificial reality tv" reality tv shows that is simulated, staring people who don't know they are a simulation also doesn't know its on a tv show.
like the Truman show, except Truman is a generative agent, the best part, the show doesn't even have to be good(right now, in the future it will have to be good, but rn it can thrive on the novelty)
cant wait.

[wukao]

Hi Sam, I am also interested in reproducing this to play around with it but found out that the gaming side of things are the hardest, would you able to setup a video to go over the Phaser side of implementation?

[andy]

Hey Sam, any plans for followup videos on this topic?. Regards, Andy

[shouldb.studying]

I can't wait for the bibites guy to get onto this

[KolTregaskes]

BabyAGI and AutoGPT could do with reflection. 😉

Again, an open source GPT model would be great if somone could replicate this for us to try.

[ИльяФрост-тн]

And I imagined this mechanism of reflection in the ordinary world and most likely it will cause the effect of an ominous valley.

[charlotteong]

Heard the code has been released on git! Will you be doing a video on that?

[daafd]

It’s basically like Sword Art Online Alicization

[ArmoredarmadilloX]

Might be weird to ask but... Do you think this would make for an interesting "game"? Just running/interacting with a lil village of AI agents?