Neural + Symbolic
Symbolica is an artificial intelligence research laboratory pioneering the application of category theory and type theory to design architectures that are capable of structured reasoning.
Our research program is focused on designing a fully general-purpose program synthesis engine, from the ground up, on our proprietary non-LLM categorical architecture.
We believe that future progress in artificial intelligence depends on the unification of the reliability of symbolic program execution with the adaptability of neural optimization. The way we build this bridge is by leveraging the language of category theory.
Program Synthesis
At Symbolica, we believe that the narrow domain most likely to yield generalized reasoning capability is program synthesis and theorem proving. A general purpose program synthesis engine or, alternatively, a sufficiently advanced theorem prover, can be configured to perform many different kinds of combinatorial optimization problems. A programmable symbolic engine capable of producing general solutions to these problems would have industry redefining consequences at almost every level.
Semantics
The formal semantics of our architecture are ensured across every level of construction. We design systems such that generated programs are provably correct. This is important for the construction of programs that run, and proofs that type check. At every level, from the high level user specification through to learned, latent representations our architecture remains mathematically self-consistent.
Computation
We integrate the ability to run arbitrary, recursive computation directly into the model architecture. The ability to execute programs is a fundamental building block of symbolic systems. By leveraging categorical semantics (see 01) we ensure that learned representations are always informed by symbolic execution.
Search
Our architecture provides a single, unified representation through which the system can propose a plan, test it (see 02), and ensure its correctness (see 01). Feedback is able to propagate through all of these components uniformly to enable continuous learning and self improvement.
Semantics
The formal semantics of our architecture are ensured across every level of construction. We design systems such that generated programs are provably correct. This is important for the construction of programs that run, and proofs that type check. At every level, from the high level user specification through to learned, latent representations our architecture remains mathematically self-consistent.
Computation
We integrate the ability to run arbitrary, recursive computation directly into the model architecture. The ability to execute programs is a fundamental building block of symbolic systems. By leveraging categorical semantics (see 01) we ensure that learned representations are always informed by symbolic execution.
Search
Our architecture provides a single, unified representation through which the system can propose a plan, test it (see 02), and ensure its correctness (see 01). Feedback is able to propagate through all of these components uniformly to enable continuous learning and self improvement.
General Program Synthesis
We have not yet solved general purpose program synthesis, and success is not guaranteed. This challenge remains one of the most difficult open research problems in the field of artificial intelligence today.
We have been hard at work leveraging state of the art research and engineering techniques to build pieces of the solution incrementally. To see our first steps in this direction, try Agentica.