OpenAI just pulled back the curtain on its technical roadmap, and it’s refreshingly candid about one thing: they don’t have a good way to measure intelligence yet. The fix? A ‘living metric’ that tracks how well a single agent can handle wildly different tasks—from video games to physical chores—using a unified set of controls. It’s built on OpenAI Gym environments with a standardized action and observation space, so one agent can be tested across games, robotics, and language tasks without retooling.
The organization is splitting its research into fundamental exploration and three concrete bets. First, a household robot using off-the-shelf hardware that learns chores through general algorithms rather than task-specific tricks. Second, an agent that can parse complex language instructions—and crucially, ask for clarification when things get ambiguous. The current state of play handles question answering and translation well enough, OpenAI notes, but falls apart when you need a machine to actually follow a multi-step command or hold a conversation.
Project three targets games, aiming for an agent that can solve any title in the test suite. DeepMind’s pioneering work here gets a nod, but OpenAI sees generative models and reinforcement learning needing serious upgrades to pull this off. The through-line is that all three projects share a common algorithmic core, so breakthroughs in one domain should spill into the others.
What’s unsaid is almost as interesting as what’s announced. A governance update is promised for later this year—a loaded topic given the ongoing tension between OpenAI’s nonprofit structure and its commercial ambitions. The post is deliberately light on timelines, too. ‘We’re just getting started on these projects, and the details may change as we gain additional data,’ the team writes. In other words, this is a direction-setting document, not a product roadmap. I’d watch the robot project most closely. If they can make a general-purpose home robot work with off-the-shelf parts and learning algorithms alone, that would validate the entire thesis.
