Today I’m writing to offer a brief peek into the GoodAI R&D roadmap to AGI (general artificial intelligence), a document we plan to release to the public in the coming weeks. It outlines our vision and methodology for developing universal artificial intelligence. The roadmap will focus on our plans for the next few months.
To generate our roadmap, we started with a set of simple game environments where our AI agent will learn in a gradual and guided way (what we call a “School for AI”). We then analyzed which abilities the AI agent needs in order to perform successfully in these environments. This analysis led us to a very specific list of functional requirements for an AI system, which we will implement and test before moving to more advanced development stages (and further iterations of our roadmap).
We plan to implement all these requirements into one universal algorithm that will be able to successfully learn all designed and derived abilities just by interacting with the environment and with a teacher.
The roadmap document will be aimed at both the general public and AI researchers.
The purpose of this document is to:
- Define our research and development roadmap for the next few months
- Agree on approaches to development
- Specify a list of very detailed abilities and requirements for a first-stage AGI system
- Unify our design, architecture, algorithms, and terminology to make sure we understand each other within the team
The current roadmap covers a first-stage AGI system. It should be able to complete learning tasks with gradually increased complexity, such as playing Pong/Breakout and Mastermind games, predicting movement and shape-switching (VisualPredictions), or completing a Watermaze task.
Example functional requirements include (not all):
- Being able to learn to detect a new/changed pattern
- Being able to learn additively, not forgetting or erasing existing knowledge
- Being able to predict on different time scales and only relevant events
- One-shot learning
- Generalization first, then specialization
- Pattern and sequence generation (later add hierarchies)
- And more
All remaining abilities/requirements will be addressed in future roadmaps – including higher-level abilities such as natural language, reasoning, creativity, curiosity, advanced exploration, goal alignment (safety, ethics), and so on. We also expect to adjust our assumptions and thinking as we progress in our experimentation – the roadmap is a working document, and will certainly develop with us as we move forward.
So please keep an eye out for our roadmap in the near future! We’re looking forward to feedback from fellow researchers and anyone interested in the progress of general artificial intelligence.
Marek Rosa
CEO, CTO & Founder of GoodAI & Keen Software House
🙂
—
And don’t forget to follow us on social media!
Facebook: www.facebook.com/GoodArtificialIntelligence
Twitter: @GoodAIdev
Forum: forum.goodai.com
www.GoodAI.com
Interesting! Are you developing your AI with the Goal of rapid exponential self-improvement in mind? Or will the aforementioned "teacher" Limit the Speed at which your AI will improve?
The speed of the AI will be limited to 104.4M/S
What do you mean by "one-shot learning"?
It only needs to see something once to have it memorized, aka what you wished to be able to do in school.
This comment has been removed by the author.
A famous roboticist in a blogpost on The Origins of Artificial Intelligence is quoting this Roadmap and evaluating it.
Thanks! It's a nicely written essay, with good arguments, but unfortunately wrong conclusions, and lacks practical usefulness for AGI research.