In our efforts to develop human-level artificial intelligence, one of the most fundamental challenges we confront is creating an AI that is able to acquire and reuse skills and knowledge in a gradual manner.

This is why we decided to start the General AI Challenge series with a warm-up round  related to the AI’s gradual learning capabilities (more rounds will be announced later).

The goal is to create an AI that would be able to acquire useful heuristics and strategies as it learns, and later reapply those strategies when learning completely new skills in new domains. The AI would be capable of “learning how to learn”, allowing it to innovate solutions to new, complex tasks in less time — something state-of-the-art AI and machine learning algorithms can’t do.

Gradual learning: implementation
Gradual learning:
evaluation &
results announced
AI Race and
Safety Assurance
Dec 2016 -
Jan 2017
Feb 2017 - 
Aug 2017
Aug 2017 -
Sept 2017
Nov 2017 -
Aug 2018


We encourage you to share your comments, suggestions, and ideas regarding this Challenge on our forum.

1st round – Gradual Learning: Learning Like a Human

Participants develop an AI agent that is able to learn gradually.

2nd round – AI Race and Safety Assurance

Participants come up with a proposal of what practical steps can be taken to avoid the pitfalls of the race between AI developers and advance the safe development of general AI.

Future rounds? – Open to discussion

Our plans and ideas include:

  • developing AI agents capable of gradual learning in more complex environments

  • solving AI safety, ethics, security, control and value alignment problems

  • developing a futuristic roadmap dealing with social, political, and economical aspects of our future

  • roadmapping general AI R&D efforts

  • “School for AI”: developing curricula for teaching AI

  • and many more (send us your suggestions for additional rounds)

Multiple rounds can run in parallel.

1st ROUND:

What is gradual learning,

and why is it important?

Gradual learning means learning skills one by one and building on those skills to acquire new, more complex ones. You can read more about how we define gradual learning in our Framework document (download here).

In order to develop general (also known as human-level) intelligence, an AI needs to acquire a certain set of skills. Our vision is that, while a limited set of skills will need to be programmed in manually, the AI will gradually learn further skills on its own.

We envision an AI that will learn numerous useful skills through a carefully designed curriculum. The intrinsic ability of an AI to learn gradually will be crucial, since it should enable:


  • learning without forgetting previously learned skills

  • building upon prior knowledge for solving new tasks

  • reuse of already learned skills and the ability to transfer them to a different domain

  • development of effective learning and problem-solving strategies

  • faster learning

  • self-improvement

  • learning through a human-designed curriculum that will teach the AI basic, useful, and non-harmful skills, ensuring that the AI bases further discovered skills on these safe and socially acceptable grounds.