What if we set GPT-4 free in Minecraft?

https://twitter.com/DrJimFan/status/1662115266933972993

by ecliptik • 3 years ago

268 99 3 years ago

Imnimo2• 3 years ago

It's always fun to look at the prompts used by these projects. Here are a few snippets from this one:

You are a helpful assistant that tells me the next immediate task to do in Minecraft. My ultimate goal is to discover as many diverse things as possible, accomplish as many diverse tasks as possible and become the best Minecraft player in the world.
8) Tasks that require information beyond the player's status to verify should be avoided. For instance, "Placing 4 torches" and "Dig a 2x1x2 hole" are not ideal since they require visual confirmation from the screen. All the placing, building, planting, and trading tasks should be avoided. Do not propose task starting with these keywords
7) Use `exploreUntil(bot, direction, maxDistance, callback)` when you cannot find something. You should frequently call this before mining blocks or killing mobs. You should select a direction at random every time instead of constantly using (1, 0, 1).
9) Do not write infinite loops or recursive functions.

You can really imagine the sorts of pitfalls the agent fell into that induced the authors to add these stipulations.

qayxc• 3 years ago

In the end it all looks like classical programming with extra steps to me :)

Imnimo2• 3 years ago

Even more than the complex prompts, they give the bot access to a variety of functions implementing primitive actions, which are crafted to give helpful hints if the bot is struggling. For example, the code for mining a block:

https://github.com/MineDojo/Voyager/blob/main/voyager/contro...

This function keeps a global count of how many times it's been called to mine a block that doesn't exist nearby, and will warn the bot that it should try exploring first instead.

There's nothing necessarily wrong with all that - it's an important research question to understand how much hand-holding the agent needs to be able to do these sorts of tasks. But readers should be aware that it's hardly dropping an agent into the world and leaving it to its own devices.

uw_rob• 3 years ago

Some would call this Software 2.0 https://karpathy.medium.com/software-2-0-a64152b37c35

doctoboggan• 3 years ago

9) Do not write infinite loops or recursive functions.

did GPT-4 just solve the halting problem?

geoelectric• 3 years ago

Seems slightly easier for me to guarantee something I wrote halts. The halting problem is more about a process analyzing someone else's arbitrary code.

narrator• 3 years ago

If the brain is like a computer, how do humans even solve the halting problem? Maybe it's because humans feel anxiety at the unproductive passing of time while computers will just do what they're doing forever.

moffkalast• 3 years ago

    if(going_to_halt)
        dont();

redox99• 3 years ago

It's trivial to avoid infinite loops by adding some exit condition in case it takes too long, which is probably what you want here.

And obviously trivial to avoid recursion too.

bentcorner• 3 years ago

Every program halts because no computer can run forever

travisjungroth• 3 years ago

One description of the halting problem is that there are not just two categories of programs (halts and doesn’t), but a third of “undecidable”. To limit yourself to the category of “halts” isn’t solving the halting problem.

nr2x• 3 years ago

Probably the best programming advice any anyone could get.

jhanschoo• 3 years ago

As others mentioned, it's trivial* to write useful programs that provably halt; the impossible task is to determine if an arbitrary program halts.

*: Trivial in a mathematician's sense

hgsgm• 3 years ago

Is there a name for the phenomenon where people who understand a concept don't talk about it because it isn't important or relevant, but people who don't understand think it's a big deal?

tmountain• 3 years ago

How did you get the prompt?

cjbprime• 3 years ago

https://github.com/MineDojo/Voyager/blob/main/voyager/prompt...

kodah• 3 years ago

It would be really cool if they simplified an NPCs world to include GPT prompts for an RPG-like experience.

1270018080• 3 years ago

That's the only application I see for black box language models that hallucinate text.

neuronexmachina• 3 years ago

Something like this? https://arstechnica.com/information-technology/2023/04/surpr...

nitwit005• 3 years ago

I searched for the word "infinite", and my suspicion was quickly proven correct:

9) Do not write infinite loops or recursive functions.
Sometimes GPT-4 will write an infinite loop that runs forever.

csours• 3 years ago

do not write infinite loops

lsy2• 3 years ago

I can only imagine the desperation of the researcher pleading with a computer to do its best to solve the halting problem.

tehsauce• 3 years ago

Earlier discussion: https://news.ycombinator.com/item?id=36085936

ChrisArchitect• 3 years ago

[dupe]

https://news.ycombinator.com/item?id=36085936

smcl2• 3 years ago

That opening sentence is a very funny statement when you take into account that "...in Minecraft" is a way some YouTubers hide hyperbolic/unserious statements (to skirt TOS violations). Like after 100 deaths to Malenia in Elden Ring: "Oh fuck me, I might as well kill myself ... IN MINECRAFT"

opan• 3 years ago

I thought the same thing, although I think of the meme as being used to avoid feds and legal problems, like saying you'll kill someone "in Minecraft" to make it not a realistic-sounding threat.

TylerLives• 3 years ago

It doesn't seem to be very effective - https://www.kotaku.com.au/2023/03/man-arrested-after-making-...

IIAOPSW3• 3 years ago

Its the kiddie version of "asking for a friend".

lsy2• 3 years ago

It's not playing in-context of minecraft, it's playing in-context of an API to minecraft. You can see one of the limitations in its error condition when it tries to craft an "acacia axe" out of acacia planks and sticks, fails, and then replaces all the references to "wooden axe". Of course in the real world it doesn't matter what you call the axe you made, and it's pretty clear what an acacia axe is. Even if it did matter, you could also easily keep the function name and output message, and just make an "wooden axe" behind the scenes. The fact that the GPT is so tightly bound to the formalism of the API is an indication that this is a task the GPT can likely do quite well as this API is well-used and documented.

thesuperbigfrog• 3 years ago

At first glance this just seems like an alternate approach at building expert systems.

The Minecraft videos are impressive.

Nethack (https://www.nethack.org/) has been used for AI development in the past and more recently:

http://shelf2.library.cmu.edu/Tech/9997774.pdf

https://portfolios.cs.earlham.edu/wp-content/uploads/2018/12...

https://arxiv.org/abs/2211.00539

https://proceedings.neurips.cc/paper/2020/hash/569ff987c643b...

https://github.com/facebookresearch/nle

https://ojs.aaai.org/index.php/AIIDE/article/view/12923

I am curious how well Voyager would do in Nethack.

slg• 3 years ago

And yet it still digs straight down.

neura• 3 years ago

Yeah... was wondering if it eventually learns that digging straight down is more likely to kill you than any other direction or if it could ever learn to down down only when standing on the edge of another block, instead of the one you're breaking.

yazaddaruvala• 3 years ago

They don't explicitly talk about curriculum improvement based on negative reinforcement. It would be relatively straight forward to do tho.

Perform self-reflection every time damage is taken (the iterations of self-reflection can depend on % health lost). Something like "Please output as a list of general guidelines / best-practices that any Minecraft player can use in the future. For each guideline add a risk profile that is introduced if the guideline is neglected. Based on the following game log over the last 100 steps, what should the Minecraft player have done differently to avoid taking damage? \n <game log>"

And keep storing those guidelines over multiple iterations. If there start to become too many best-practices, ask GPT-4 to "Please output as a list of general guidelines / best-practices that any Minecraft player can use in the future. For each guideline add a risk profile that is introduced if the guideline is neglected. Take the guidelines below and summarize them. Prioritize retaining more detail about the guidelines with a higher risk profile, and do merge guidelines if possible and appropriate. \n <old guidelines>"

js8• 3 years ago

It also fights an Enderman with a shovel, how smart.

Geee• 3 years ago

Very interesting. Here LLM writes code that plays Minecraft. If software 2.0 is neural networks, then software 3.0 is code written by neural networks.

ablyveiled• 3 years ago

And software 4.0 is sticks and stones. har har.

jmugan2• 3 years ago

I wish there was a summary of how this worked. I see the abstract and lots of figures and movies, but I still don't get a good sense of what exactly the algorithm is. I even skimmed the whole paper.

ShamelessC2• 3 years ago

https://twitter.com/DrJimFan/status/1662117785487704067

You just scroll down a tiny bit on the twitter page and get this nice video and summary from the author.

Voyager has 3 key components:
1) An iterative prompting mechanism that incorporates game feedback, execution errors, and self-verification to refine programs;
2) A skill library of code to store & retrieve complex behaviors;
3) An automatic curriculum to maximize exploration.

jmugan2• 3 years ago

Right, but I wanted to know what each of those things was.

codeulike• 3 years ago

This is like its writing a bot to play minecraft?

I'd like to see a visual/language model/AI that learns to play minecraft as an actual inhabitant of the game. i.e. processing visual input, recognising objects, working out whats going on, learning how to move around. Learning how to make food and avoid monsters. It would be an 'Embodied AI' within the world of Minecraft.

The language part would allow us to talk to this being. You could ask it things like:

"Do you prefer to make a house, or dig a cave?"

"What are your hopes for the future?"

"Is there a recent achievement you are particularly proud of?"

etc

AlecSchueler• 3 years ago

You could ask it those things but it will tell you that it doesn't have feelings, preferences or hopes. You'd also need to give a reason to do anything. Eventually you get back to the point where you're "writing a bot" but with behaviour closer to how you imagine it should be.

krapp• 3 years ago

I'd like to see a visual/language model/AI that learns to play minecraft as an actual inhabitant of the game. i.e. processing visual input, recognising objects, working out whats going on, learning how to move around. Learning how to make food and avoid monsters. It would be an 'Embodied AI' within the world of Minecraft.

There is already an AI VTuber, Neurosama, trained to "play" games, including Minecraft (also OSU! and Among Us.)

I don't think she's learned lava is bad yet.

iinnPP• 3 years ago

I imagine this is being worked on for OSRS as well. Exciting times. Terrifying times.

drekk• 3 years ago

It's interesting you mention OSRS because some bots are using ChatGPT (or other LLMs) to respond to players and it's incredibly effective. Only having to deal with conversations seems like a perfect use-case for the technology

https://youtu.be/fu9OQRh6K8U

phendrenad2• 3 years ago

I'm imagining Twitch Plays Pokemon but even less coherent.

blooalien• 3 years ago

Where I think it's gonna start to get really scary, and much more closely approaching "real A.I" / AGI is when they start augmenting and wiring together various differing forms of "A.I." with each other. GPT-4, no matter how impressive it might appear on the surface is still "just" a large language model. Augment it with other types of learning models and at some point you might just hit on the right combination for it to start some form of actual "reasoning" thought or "creativity".

As long as they're all still "special" single-purpose systems (LLM is about processing and responding to language input for example, CV / Computer Vision models specialize in operating on visual or image inputs, etc.), that's all they'll ever be, no matter how good they get at pretending they're more.

hahajk• 3 years ago

Isn't GPT-4 multimodal? I remember it designing a website based off of a sketch during the initial demo.

ShamelessC2• 3 years ago

It is.

rst• 3 years ago

Theres a multimodal variant, but it's not widely available yet.

MagicMoonlight• 3 years ago

A LLM is a form of input like a keyboard. If they put it in front of something like siri and used it as input instead of a processor then you could make siri actually functional.

It’s very good at understanding text but by itself it can’t think. Turning text into commands it knows is doable.

fnordpiglet• 3 years ago

This is happening. These projects are examples of feedback loops. While the models aren’t playing directly they are receiving feedback and iteratively improving. Constraining and optimizing using classical AI is an obvious next step. I agree this is when the magic happens.

mensetmanusman• 3 years ago

It will set out to build a library that contains all the sources that it quotes but that only exist in the multiverse.

busseio• 3 years ago

I’d like to build something like this, but for Robot Odyssey.

mbgerring• 3 years ago

Why? Who wants this? To what end?

mkaic• 3 years ago

Why?

Because the researchers found it to be an interesting problem. Using an AI to beat Minecraft has been an active benchmark in the ML world for some time now.

Who wants this?

Me, as well as the many others who have liked and shared the paper.

To what end?

Furthering our ability to create autonomous agents. If we can get it to work in Minecraft, that's one step closer to getting it to work in real life.

sdenton4• 3 years ago

There's untold billions of tokens of Minecraft related content on the web - the controlling LLM personally has memorized every Minecraft strategy guide ever written.

thdespou• 3 years ago

Prompt engineering at it's finest form.

Fgehono• 3 years ago

Crazy how fast the peaces fall in place.

A big text corpus gives quite a huge amount of context.

SD also got better through this

What if we set GPT-4 free in Minecraft?

Related Stories