Toward more explainable, more human AI — a research direction
What this is
This describes an active research direction and its aims. It is deliberately high-level: it does not disclose the underlying methods, and it makes no claim of finished results or benchmarks. It is here to show what I'm working toward, not to announce that I'm done.
The problem
Today's most capable models are powerful but opaque. When they explain themselves, the explanation is usually reconstructed after the answer rather than drawn from how the answer was actually produced. In low-stakes settings that gap is tolerable. In law, finance, medicine, and governance, it is not — those fields need reasoning that can be inspected, not just outputs that are usually right.
The approach
I'm developing an approach grounded in mathematical and linguistic structure rather than scale alone. The aim is a system where the representation itself carries meaning, so the path to an answer can be followed and checked — explanation that is intrinsic to the computation instead of bolted on afterward.
What I'm aiming for
Three properties guide the work: generalization that feels more human and needs less data to find the pattern; explanations that are faithful to the actual computation; and results an expert can audit and reason about rather than take on trust.
Why it matters
The market is converging on larger black boxes. The harder, more useful question is whether intelligence can be built so that understanding comes with it — where being able to explain a decision is a property of the system, not an add-on. That is the direction I think is worth pushing.
Status
This is early-stage research, not a product. Methods and results are kept private while the work matures. I'm open to the right collaborations, conversations, and funding partners.
If this is your field
If you work on explainable, neuro-symbolic, or otherwise interpretable AI — or you fund this kind of research — I would like to talk.