OpenAI’s mission is to build safe AI, and ensure AI's benefits are as widely and evenly distributed as possible. We’re trying to build AI as part of a larger community, and we want to share our plans and capabilities along the way. We’re also working to solidify our organization's governance structure and will share our thoughts on that later this year.
Defining a metric for intelligence is tricky, but we need one to measure our progress and focus our research. We're thus building a living metric which measures how well an agent can achieve its user’s intended goal in a wide range of environments.
The metric will consist of a variety of OpenAI Gym environments with a unified action and observation space (so a single agent can run across all of them), including games, robotics, and language-based tasks. Our implementation will evolve over time, and we’ll keep the community updated along the way.
A significant fraction of our research bandwidth is being spent on fundamental research. We’ll always be developing and testing new ideas, especially those that don’t fit neatly into our current worldview. This is important — our current ideas will not be enough to achieve our long-term goal.
We’ve also formed teams around specific projects. The intention isn’t just to solve these problems, but to develop general learning algorithms in the process. These algorithms will, in turn, help us build agents that are more capable according to our metric. These projects are:
We're working to enable a physical robot (off-the-shelf; not manufactured by OpenAI) to perform basic housework. There are existing techniques for specific tasks, but we believe that learning algorithms can eventually be made reliable enough to create a general-purpose robot. More generally, robotics is a good testbed for many challenges in AI.
We plan to build an agent that can perform a complex task specified by language, and ask for clarification about the task if it’s ambiguous. Today, there are promising algorithms for supervised language tasks such as question answering, syntactic parsing and machine translation but there aren’t any for more advanced linguistic goals, such as the ability to carry a conversation, the ability to fully understand a document, and the ability to follow complex instructions in natural language. We expect to develop new learning algorithms and paradigms to tackle these problems.
We aim to train an agent capable enough to solve any game in our initial metric. Games are virtual mini-worlds that are very diverse, and learning to play games quickly and well will require significant advances in generative models and reinforcement learning. (We are inspired by the pioneering work of DeepMind, who have produced impressive results in this area in the past few years.)
Our projects and fundamental research all have shared cores, so progress on any is likely to benefit the others. Each captures a different aspect of goal-solving, and was chosen for its potential to significantly move our metric.
We’re just getting started on these projects, and the details may change as we gain additional data. We also expect to add new projects over time.