Will any agent perform better on Minecraft (or comparable open world game) after being fine-tuned on a manual by 2027?
Plus
13
Ṁ12552027
72%
chance
1D
1W
1M
ALL
To clarify: the experiment is that there are two copies of an agent that runs on Minecraft (or some other open world game environment). The agent has the capacity to be fine-tuned with text. One version is passed a manual for the game as text (or text + images, but *not* video), the other runs without any finetuning. Will the former perform better than the latter (either better sample efficiency or better final reward)?
The agent can't have been trained on that env before, but it can be trained on other envs/data beforehand (e.g. it's okay if there's a pretrained LLM in the loop).
This question is managed and resolved by Manifold.
Get
1,000
and3.00
Sort by:
There is at least a paper claiming to do this for Atari environments : https://arxiv.org/abs/2302.04449
@MartinRandall I will accept essentially any text-based vaguely guide-like thing. The specific details of the text aren't what this question is getting at.
@April If nothing like this gets attempted I'll resolve it N/A. I'm not very interested in the probability the experiment is performed at all.
@JamesBabcock which means they can get manifold bux by posting their experiment, which should compensate them for any scientific reputation loss from posting a failure case, right?
Related questions
Related questions
Will we get a video of claude 3.5 Sonnet running a very single minded competent minecraft agent before December 2024?
25% chance
Before 2035, will there exist any AI that can perform arbitrary tasks in Minecraft?
74% chance
Before 2035, will there exist any AI that can perform arbitrary tasks in Minecraft?
80% chance
Will an AI Minecraft Agent defeat the Ender Dragon before 2025?
12% chance
When will a single agent beat Minecraft (defeat the Ender dragon)?
Will AutoGPT-style AI Agents mostly work before the end of 2024?
20% chance
By 2026 will any RL agent with learned causal models of its environment achieve superhuman performance on >=10 Atari environments?
81% chance
Will a video game be created where its agents are sentient enough to try to hack its own simulation by 2045?
32% chance
Will models be able to do the work of an AI researcher/engineer before 2027?
40% chance
Will an artificial agent ascend in NetHack (version 3.6.6 or later) before the end of 2024?
25% chance