7 ms·

Voyager: An Open-Ended Embodied Agent with LLMs

96 points by jedixit 3y ago

startupsfail 3y ago
TLDR. An AI system with an IQ of 110-130, with some careful prompting can generate code to play Minecraft through an API.
- andai 3y ago
  >130 IQ >mines straight up and down
  - startupsfail 3y ago
    I’m not sure what is the point that you are making. GPT-4 does tend to pass various IQ tests with the scores in the range of 110 to 130, with outliers between 90 to 150.
    - notamy 3y ago
      It's a joke about playing the game "right." Mining straight up/down is a rather suboptimal strategy, as: - mining straight up means you either seal your path behind you, or are limited how high up you can go - mining straight down likely traps you in a pit - mining straight down far enough can drop you straight into lava, as many Minecraft players learn early on
jamilton 3y ago
I wonder how many prompts this uses in a minute. Interestingly, they mod the server so that the game pauses while waiting for a response from GPT-4. That's a nice way to get around the delays.
tehsauce 3y ago
This is very cool despite the most important caveat: “Note that we do not directly compare with prior methods that take Minecraft screen pixels as input and output low-level controls [54–56]. It would not be an apple-to-apple comparison, because we rely on the high-level Mineflayer [53] API to control the agent. Our work’s focus is on pushing the limits of GPT-4 for lifelong embodied agent learning, rather than solving the 3D perception or sensorimotor control problems. VOYAGER is orthogonal and can be combined with gradient-based approaches like VPT [8] as long as the controller provides a code API.”
bartwr 3y ago
This is kind of amazing given that obviously, GPT-4 never contained such tasks and data. I think it puts an end to the claim that "language models are only stochastic parrots and cannot do any reasoning". No, this is 100% a form of reasoning and furthermore, learning that is more similar to how humans learn (gradient-less). I still don't understand it and it blows my mind - how such properties emerge just from compressing the task of next word prediction. (Yes, I know this is oversimplification, but not a misleading one).
- chinchilla2020 3y ago
  I read through the code and tried it out for 15 mins. It's a hard-coded program that can do a text search for it's own hard-coded, human-implemented functions. Apparently it can string those functions together, but doesn't do it correctly. https://github.com/MineDojo/Voyager/tree/main/voyager/control_primitives https://github.com/MineDojo/Voyager/tree/main/voyager/contro... 20 minutes of light reading through the repository pretty much dispels any notions that this is a self-learning system that can reason and think. It's the same minecraft automation we have been seeing for a decade now, with a chatbot text search builtin.
- asperous 3y ago
  While I do believe LLMs can perform some reasoning, I'm not sure this is the best example as all the reasoning you would ever need for Minecraft is well contained in the data set used to train it. A lot has been written about minecraft. To me, it would be more convincing if they developed an enterly new game with somewhat novel and arbitrary rules and saw if the embodied agent could learn this game.
- notamy 3y ago
  Looking at the paper, as I understand it they're using Mineflayer https://github.com/PrismarineJS/mineflayer https://github.com/PrismarineJS/mineflayer and passing parts of the state of the game as JSON to the LLM that are used for code generation to complete tasks. > I still don't understand it and it blows my mind - how such properties emerge just from compressing the task of next word prediction. The Mineflayer library is very popular, so all the relevant tasks are likely already extant in the training data.
lsy 3y ago
I'm not sure how the authors arrive at the idea that this agent is embodied or open-ended. It is sending API calls to minecraft, there's no "body" involved except as a symbolic concept in a game engine, and the fact that minecraft is a video game with a limited variety of behaviors (and the authors give the GPT an "overarching goal" of novelty) precludes open-endedness. To me this feels like an example of the ludic fallacy. Spitting out "bot.equip('sword')" requires a lot of non-LLM work to be done on the back end of that call to actually translate to game mechanics, and it doesn't indicate that the LLM understands anything about what it "really" means to equip a sword, or that it would be able to navigate a real-world environment with swords etc.
- FeepingCreature 3y ago
  Though, I also don't individually track muscle fibers, and there's strong indications that a lot of my own behaviors are closer to API calls than direct control.
osalberger 3y ago
Just checked by talking to the free version of ChatGPT, and yes, the MineFlayer api docs are indeed in its training set. It can give me detailed instructions on how to build a minecraft bot. And of course, it also knows the entire minecraft tech tree very well. So this isn't really open ended work, its just making it do something it is already trained on, by connecting it to an API that it has learned the docs of.
- osalberger 3y ago
  However, the skills library that it writes with live feedback from runtime errors and that it retrieves with a vector DB is really interesting. In that sense it looks like a very interesting code generation application.
ivanblagdan 3y ago
Semantics of how this works aside, take a moment to appreciate how easy it is to remap the variable “zombie” to “human” in a prompt without the model altering its behavior. It instantly makes you realize the immensity of the AI safety & alignment problem.

96 points by jedixit 3y ago

TLDR. An AI system with an IQ of 110-130, with some careful prompting can generate code to play Minecraft through an API.

andai 3y ago
>130 IQ >mines straight up and down
- startupsfail 3y ago
  I’m not sure what is the point that you are making. GPT-4 does tend to pass various IQ tests with the scores in the range of 110 to 130, with outliers between 90 to 150.
  - notamy 3y ago
    It's a joke about playing the game "right." Mining straight up/down is a rather suboptimal strategy, as: - mining straight up means you either seal your path behind you, or are limited how high up you can go - mining straight down likely traps you in a pit - mining straight down far enough can drop you straight into lava, as many Minecraft players learn early on

jamilton 3y ago

I wonder how many prompts this uses in a minute. Interestingly, they mod the server so that the game pauses while waiting for a response from GPT-4. That's a nice way to get around the delays.

tehsauce 3y ago

This is very cool despite the most important caveat: “Note that we do not directly compare with prior methods that take Minecraft screen pixels as input and output low-level controls [54–56]. It would not be an apple-to-apple comparison, because we rely on the high-level Mineflayer [53] API to control the agent. Our work’s focus is on pushing the limits of GPT-4 for lifelong embodied agent learning, rather than solving the 3D perception or sensorimotor control problems. VOYAGER is orthogonal and can be combined with gradient-based approaches like VPT [8] as long as the controller provides a code API.”

bartwr 3y ago

This is kind of amazing given that obviously, GPT-4 never contained such tasks and data. I think it puts an end to the claim that "language models are only stochastic parrots and cannot do any reasoning". No, this is 100% a form of reasoning and furthermore, learning that is more similar to how humans learn (gradient-less). I still don't understand it and it blows my mind - how such properties emerge just from compressing the task of next word prediction. (Yes, I know this is oversimplification, but not a misleading one).

chinchilla2020 3y ago
I read through the code and tried it out for 15 mins. It's a hard-coded program that can do a text search for it's own hard-coded, human-implemented functions. Apparently it can string those functions together, but doesn't do it correctly. https://github.com/MineDojo/Voyager/tree/main/voyager/control_primitives https://github.com/MineDojo/Voyager/tree/main/voyager/contro... 20 minutes of light reading through the repository pretty much dispels any notions that this is a self-learning system that can reason and think. It's the same minecraft automation we have been seeing for a decade now, with a chatbot text search builtin.
asperous 3y ago
While I do believe LLMs can perform some reasoning, I'm not sure this is the best example as all the reasoning you would ever need for Minecraft is well contained in the data set used to train it. A lot has been written about minecraft. To me, it would be more convincing if they developed an enterly new game with somewhat novel and arbitrary rules and saw if the embodied agent could learn this game.
notamy 3y ago
Looking at the paper, as I understand it they're using Mineflayer https://github.com/PrismarineJS/mineflayer https://github.com/PrismarineJS/mineflayer and passing parts of the state of the game as JSON to the LLM that are used for code generation to complete tasks. > I still don't understand it and it blows my mind - how such properties emerge just from compressing the task of next word prediction. The Mineflayer library is very popular, so all the relevant tasks are likely already extant in the training data.

lsy 3y ago

I'm not sure how the authors arrive at the idea that this agent is embodied or open-ended. It is sending API calls to minecraft, there's no "body" involved except as a symbolic concept in a game engine, and the fact that minecraft is a video game with a limited variety of behaviors (and the authors give the GPT an "overarching goal" of novelty) precludes open-endedness. To me this feels like an example of the ludic fallacy. Spitting out "bot.equip('sword')" requires a lot of non-LLM work to be done on the back end of that call to actually translate to game mechanics, and it doesn't indicate that the LLM understands anything about what it "really" means to equip a sword, or that it would be able to navigate a real-world environment with swords etc.

FeepingCreature 3y ago
Though, I also don't individually track muscle fibers, and there's strong indications that a lot of my own behaviors are closer to API calls than direct control.

osalberger 3y ago

Just checked by talking to the free version of ChatGPT, and yes, the MineFlayer api docs are indeed in its training set. It can give me detailed instructions on how to build a minecraft bot. And of course, it also knows the entire minecraft tech tree very well. So this isn't really open ended work, its just making it do something it is already trained on, by connecting it to an API that it has learned the docs of.

osalberger 3y ago
However, the skills library that it writes with live feedback from runtime errors and that it retrieves with a vector DB is really interesting. In that sense it looks like a very interesting code generation application.

ivanblagdan 3y ago

Semantics of how this works aside, take a moment to appreciate how easy it is to remap the variable “zombie” to “human” in a prompt without the model altering its behavior. It instantly makes you realize the immensity of the AI safety & alignment problem.