Simple Step by Step Guide to Build Your Own AGI
Building AGI has not been so simple before. Follow this simple 3 step guide to build your own AGI.
AGI is real and AGI is coming. OpenAI, the front runner of building AGI, has even started a large effort called super alignment, where they are exploring how we dumb humans can make sure super intelligence aligns with human values. However, so far nobody has publicly shared a practical guide to build AGI. In the post, I am going to dismantle the mystery of AGI. Follow the simple to follow, step by step instructions, and you can build your own AGI in an unknown time.
Believe it or not, it just takes 3 steps.
Step 1, build a continuously trained world model, a model that takes the physical world around as input, predicts what the world around would look like in the near future and what the agent (the AGI that we are building) should do next, and correct the predictions if they turn out to be wrong decisions. This is what the human brain is doing and for true intelligence, they shouldn’t rely on humans to provide highly processed knowledge, because important information that humans don’t yet understand are lost in the production of knowledge.
Challenge of this step:
there are no such models for a slightly complex physical world. The closest example is self driving cars. But currently they are not fully autonomous and self driving is a pretty controlled environment with a very limited number of actions.None
Step 2, decentralize the world model. The model built in step 1 will work in a small controlled environment, but the physical world is just too big and too complex for a single system to process autonomously. If the training & decision making is in a centralized system, who can take care of maintaining and replacing sensors that are thousands of miles away from the central processing system? Who can take care of maintaining the connection between censors & the central processing system? Who can extend the system to explore previously unknown territory? A centralized system is just too fragile and too inflexible.
Decentralization means there will be lots of local agents, each of them carrying their own censors, making their own predictions, making their actions and correcting their own models. But since one agent only sees part of the reality and has only limited processing capabilities & memory, how can they discover & build new things?
Challenge of this step:
Putting sensing/training/prediction/acting into a single small system has lots of engineering challengesIt is unclear how we can make it better, or at least having similar quality as humansNobody has tried it yetNone
Step 3, let the models learn how to abstract & communicate. One agent only sees and learns from part of reality, but if the agents can communicate, share their ideas and learn from others’ ideas, then as a whole they become as capable as a huge centralized system, but without the fragility and inflexibility of the centralized system.
Abstraction is the ability to compress experience into broadly applicable knowledge. Abstraction is essential to effective communication because each agent has only limited capacity and if there is no abstraction, the communicated information will get lost because of the capacity limit.
Through abstraction & communication, the agents can now build on top of each other’s ideas to develop new ideas, make new things, and eventually form their own civilization.
Challenge of this step:
It is unclear how models can learn to abstract & communicateIt is unclear how we can make it better, or at least having similar quality as humansNone
That concludes this simple to follow, 3 step guide to build AGI. I hope you all enjoy it and feel confident building your own. If you run into problems, feel free to comment below and I am more than happy to help.