OpenAI Drops 7 New Unsolved AI Research Problems

OpenAI just published a fresh set of seven unsolved research problems, a spiritual successor to their original ‘Requests for Research’ that previously led to several published papers and even job offers. Think of it as a public challenge list—designed to get newcomers into the field and give seasoned practitioners genuinely hard problems to chew on. The problems aren’t about tweaking hyperparameters on a standard benchmark; they explicitly warn that many will require inventing entirely new ideas.

Leading the charge are a handful of graded challenges. A one-star starter asks you to train an LSTM to solve the XOR parity problem on binary strings of variable length, probing just how well the model generalizes when sequence length isn’t fixed. It scales up quickly from there: a two-star problem demands you build a multiplayer Snake clone (‘Slitherin’’) in a Gym environment and solve it with self-play, wrestling with the notorious instability that makes it feel like training a GAN. The suggestion to train your current policy against a distribution of its past selves points directly at the core difficulty of multi-agent dynamics.

The real meat is in the three-star problems. One asks researchers to explore parameter averaging schemes in distributed RL, hunting for ways to slash communication bandwidth while using divergent agent parameters as a built-in exploration strategy—a nod to algorithms like EASGD. Another tackles transfer learning between Atari games via generative models. The goal is to pre-train a Transformer on trajectories from 10 games and quantify the benefit when fine-tuning on an 11th, especially under severe data constraints like a 100x reduction.

But the final challenge might be the most quietly provocative. OpenAI wants someone to make a Transformer work with linear attention instead of the standard softmax, turning it into an RNN with fast weights suitable for RL rollouts over massive contexts. They then undercut the entire premise with a single sentence: ‘Only one caveat: this may turn out to be impossible.’ It’s a refreshingly honest signal that they’re not just looking for incremental results—they want people willing to hunt for a breakthrough that might not exist.