OpenAI Technical Goals (2016)

OpenAI's 2016 technical roadmap establishes a framework for building safe AI by focusing on a universal metric for intelligence and four primary technical goals. This strategy aims to develop general learning algorithms that can be applied across diverse environments to ensure AI benefits are widely and evenly distributed.

The Universal Intelligence Metric

OpenAI is developing a "living metric" to measure intelligence based on an agent's ability to achieve a user's intended goal across a wide range of environments. This metric is designed to provide a concrete way to measure progress and focus research efforts.

To implement this, OpenAI is utilizing a variety of OpenAI Gym environments. These environments feature a unified action and observation space, allowing a single agent to operate across multiple domains, including:

• Robotics
• Language-based tasks
• Games

Core Technical Goals

OpenAI has organized its research into specific projects intended to develop general learning algorithms. While these projects target specific outcomes, the underlying goal is to create algorithms that increase the agent's capability as measured by the universal intelligence metric.

Goal 1: Measuring Progress

OpenAI prioritizes fundamental research to test new ideas that may fall outside their current worldview, acknowledging that existing ideas will be insufficient to reach their long-term objectives.

Goal 2: Household Robotics

One of the primary technical targets is the development of a household robot, representing a physical application of general goal-solving capabilities.

Goal 3: Natural Language Understanding

OpenAI aims to build an agent with useful natural language understanding, focusing on the linguistic aspect of goal achievement.

Goal 4: Multi-Game Mastery

OpenAI intends to train a single agent capable of solving any game within their initial metric. Because games serve as diverse virtual mini-worlds, mastering them requires significant advancements in reinforcement learning and generative models. OpenAI notes that this goal is inspired by the pioneering work of DeepMind.

Research Integration and Evolution

Progress in any single project is expected to benefit the others because the fundamental research and specific projects share core components. Each goal was selected because it captures a different aspect of goal-solving and has the potential to significantly move the overall intelligence metric.

OpenAI acknowledges that these projects are in their early stages and that details may evolve as more data is gathered and new projects are added over time.