Ash Vardanian, founder of AI infrastructure company Unum, published an essay arguing that large language models may be the "epicycles" of the current AI era — powerful universal approximators that achieve impressive empirical accuracy but lack fundamental explanatory depth. Drawing on the history of Ptolemaic astronomy, Vardanian observes that just as ancient astronomers stacked circles upon circles to model planetary motion with increasing accuracy but decreasing elegance, modern AI researchers stack transformer layers to approximate intelligence. The system works, sometimes remarkably well, but <a href="/news/2026-03-14-lm-head-gradient-bottleneck-llm-training">parameter counts balloon</a> without revealing any underlying structure of what intelligence actually is. His central thesis is that AI may be awaiting its own "Kepler moment" — a conceptual breakthrough analogous to Johannes Kepler replacing dozens of hand-tuned epicycle parameters with a small set of geometric relationships to describe planetary orbits, and Newton compressing further still into a single universal force law.

Vardanian identifies memory and optimization as the two primitive operations most likely to survive any future paradigm shift in AI. He argues these are the invariant load-bearing concepts regardless of what eventually replaces the transformer architecture — memory because it grounds behavior in past experience, optimization because it enables adaptation. That argument comes from someone deeply embedded in AI infrastructure: Vardanian describes LLMs as "the world's largest lossy compression contest," a skepticism unusual in an industry defined by enthusiasm for scale. The essay situates current LLM development in a historical arc that includes Ptolemy, Copernicus, Kepler, and Newton, acknowledging the real utility of today's models while questioning whether "intelligence" is even the right frame for evaluating them.

The philosophical argument is not disconnected from commercial interest. Unum's product suite maps directly onto the two primitives Vardanian identifies as durable. USearch, the company's single-header C++ approximate nearest neighbor search engine, implements high-performance associative retrieval — the memory primitive — at the infrastructure layer, independent of whatever model generates the embeddings it indexes. UForm, its compact multimodal AI library, offers models ranging from 79M to 365M parameters, a deliberate architectural bet against the dominant scaling logic. A world where compact, principled architectures outcompete billion-parameter transformers — the post-epicycle future Vardanian describes — is a world where Unum's infrastructure layer becomes more valuable, not stranded. The essay is simultaneously a philosophical provocation and a product thesis, with the two reinforcing each other in ways that give the argument unusual coherence and stakes.