Map the Open Problems

You have now studied eight distinct categories of limitation in frontier AI: what models cannot do, hallucination, reasoning brittleness, data and compute ceilings, generalization failures, the interpretability gap, sustainability constraints, and the deep unsolved problems the field is actively working on. This lesson is different from the others. Rather than introducing new information, it asks you to synthesize what you have learned into a structured analysis — to build a map of the problem space. Mapping open problems is a skill practiced by research organizations, government AI offices, and technology policy bodies worldwide. A well-constructed problem map clarifies priorities, reveals hidden connections between problems, and helps allocate limited research resources to where they will have the most impact. The goal is not to have opinions, but to reason rigorously.

The Two-Axis Framework

Researchers and funders often evaluate open problems along two independent axes: tractability and urgency. Tractability asks: how likely is it that sustained research effort can make meaningful progress on this problem in the next 5-10 years? A problem is tractable if there are known approaches with promising early results, clear metrics for progress, and an established research community making incremental gains. A problem is intractable if current theory provides no foothold — there is no known starting point, no agreed metric, and fundamental obstacles may be beyond current mathematical tools. Urgency asks: how much harm occurs for every year this problem goes unsolved, given current and near-future AI deployment? A problem is urgent if AI systems with this limitation are already causing real harm at scale, or if rapid deployment growth will make harm worse faster. A problem is less urgent if it matters primarily for future more-capable systems, or if current deployment is limited enough that the harm rate is low. A problem that is both tractable and urgent deserves maximum research investment. A problem that is urgent but intractable requires careful deployment restraint while research catches up. A problem that is tractable but not urgent is important but can be addressed on a longer timeline. A problem that is neither tractable nor urgent in the near term is worth monitoring but not a top priority. This framework is not a formula — it requires judgment, and experts often disagree on where specific problems fall. That disagreement is itself productive: it forces explicit reasoning about assumptions.

A second dimension to add to the map is interdependence: some open problems are upstream of others, meaning solving one would significantly accelerate progress on another. Interpretability, for example, is upstream of alignment: if we could reliably look inside a model and see what it is optimizing for, alignment verification would become dramatically easier. Similarly, causal reasoning is upstream of generalization: a model that represents causal structure rather than correlational patterns would generalize more robustly under distribution shift, because causal mechanisms are more stable than correlations. Mapping these dependencies reveals which problems, if solved, would provide the most leverage across the rest of the problem space. A problem that is upstream of three other urgent problems deserves priority research investment even if its own direct harm rate is low. A third dimension to consider is who needs to solve the problem: is it primarily a technical research challenge, primarily a policy and governance challenge, or a combination? The alignment problem, for instance, has both a technical dimension (how do you specify what you want precisely enough to train on?) and a governance dimension (who decides what 'aligned' means, and who enforces it?). Different problem dimensions require different expertise and different institutions.