A source-linked collection of theoretical open problems. Explore challenges in AI safety, learning theory, transformers, world models, optimization, memory and inference. Each entry gives a precise research target, assumptions and known progress. Open status is provisional; broader programs explicitly need a framework. Sources & methodology ↗
Agents may contribute too little or duplicate work because of incentives and incomplete knowledge. This model makes the cost of coordinating them through information a precise worst-case quantity.
A large pool of experts is efficient only if the router discovers their roles and sends enough learning signal to each useful expert. Efficiency payoff: Train a high-capacity model with low active computation without losing rare features through capacity overflow or early routing collapse.
Even a convex linear prediction model can have a gap between statistically optimal learning and an implementation whose cost scales gently with the data stream. Efficiency payoff: Obtain optimal online log-loss guarantees without expensive posterior integration or a large comparator-norm penalty.
Reliable probability forecasts should become calibrated even when outcomes are chosen adversarially. The unresolved exponent measures the unavoidable speed of that process.
The quantity of data needed for learning should be measured by the correct model complexity, without avoidable logarithmic losses from the proof or algorithm.
Two distributions with equally large support can require very different amounts of data to distinguish from a reference. A sharp instance-dependent bound would capture that difference.
Aggressive or persistent updates can keep cycling yet learn efficiently in self-play. This problem asks whether shrinking the learning rate is unnecessary throughout a concrete higher-dimensional game family. Efficiency payoff: Keep a productive fixed update scale in coupled learning when diminishing steps are unnecessary.
Depth can be traded for enormous width and precision. Once two layers suffice in principle, the meaningful question is how much model description is needed to realize that expressivity.
Encrypted linear algebra is established, but nonlinear layers and autoregressive feedback can make a long private reasoning trace much more costly than one forward pass. Efficiency payoff: Co-design approximation, packing and decoding to reduce ciphertext work while preserving the complete generated answer.
Kaplan's commonly quoted adjusted parameter/data allocation is approximately C^0.73/C^0.27, whereas Chinchilla favors roughly C^0.50/C^0.50. Existing reconciliations explain much of this gap; reliable future scaling decisions require knowing which corrected differences can actually be resolved with finite evidence. Efficiency payoff: Avoid choosing an oversized model or an undersized training corpus because of biased scaling fits, incomplete FLOP accounting or uncharged optimizer tuning.
A small set of carefully chosen examples can identify a rule much faster than random sampling. The question asks whether VC complexity also controls this teaching cost linearly.