Unlocking the Future: How Chat-GPT Revolutionizes Software Development through Minecraft Exploration

Large language Models (LLMs) seem to be the "Holy Grail" for the future of software development.

In a recent mind-blowing discovery by a team working at NVIDIA. Emerged a new way of solving some of computer science's most challenging problems. And it all began with Minecraft, yes...the child friendly building block game.

But hear me out!

On May 25th 2023, the team published a paper called "Voyager: An Open-Ended Embodied Agent with Large Language Models"[1]. In it detailed the inner workings of turning a vast ever changing multi-dimensional environment into something more digestible by OpenAI's GPT-4.

It gathers information about the game through an API. Allowing it to see the current state of the game to then make decisions based on that information. For example if it sees the users hunger is low and has food in the inventory. It's next objective will be to eat. Otherwise, it will make the intelligent decision to scavenge for food. Then continue on its quest to achieve the next goal. If that be building a house, exploring the land, or mining for diamonds. It is able to make these high level decision trees which would mimic that of a human brain.

It's no surprise GPT-4 has been called the beginning of AGI. Microsoft's Sr. Principal Research Manager Sebastien Bubeck even wrote a research paper suggesting just that[2]. What was found was, "GPT-4's performance is strikingly close to human-level performance".

Side note: Microsoft is a lead investor in OpenAI, and invested $10 billion at the start of 2023[3].

Back to Minecraft, there are very few solutions currently known that would be able to accomplish a feat like this. One of the primary ways video games were solved by AI in the past was through reinforcement learning. Simply, a machine learning strategy of trial and error, and moving the "fittest" to the next generation. A process similar to that of evolution.

An extremely great video on this here:

While this strategy works on video games like the original Super. Mario Bros where the game is exactly the same every time its played. With the constantly changing environment in Minecraft, reinforcement learning wouldn't work in this case. Minecraft's world is procedurally generated and is essentially random. What might "work" on one randomly generated world won't work on the next. Each would require completely different inputs to achieve the exact same goals.

The paper even reference's this major flaw in reinforcement learning stating "Classical approaches employ reinforcement learning (RL) and imitation learning that operate on primitive actions, which could be challenging for systematic exploration, interpretability, and generalization."

So, why is this a big deal? Sure a computer can now play Minecraft. But it doesn't just end here. If this technology keeps advancing and LLMs are able to solve these extremely abstract and complex problems in a similar way. We could very well see this same solution applied to so many different aspects of our lives. From autonomous vehicles to fully automated factories.

Large language models are still in their infancy and I believe might be the single greatest discoveries of our lifetime.

If you would like to read more about the Minecraft playing AI: https://voyager.minedojo.org/