Lucidity Sciences says algorithmic changes can make AI more accurate and efficient

Lexi Posse, CEO and co‑founder of Lucidity Sciences, told the Deep Tech panel that many machine‑learning pipelines default to collecting vast amounts of training data and effectively memorizing patterns, which limits out‑of‑sample accuracy and costs substantial compute.

“We've taken a very different approach,” Posse said, describing work to rework the mathematical and algorithmic underpinnings of machine learning so models can discover fundamental structure and be both more accurate in unfamiliar situations and more efficient to train. She cited a tabular‑learning benchmark and said Lucidity matched or outperformed top models that required hundreds of hours on a cluster, while doing comparable tasks “in hours on a gaming laptop.”

Posse emphasized that large language models are one class of AI that pattern‑match language but that many valuable AI applications can bypass a linguistic filter and operate directly on mathematical representations of systems such as finance, healthcare or grid telemetry. She argued design choices that prioritize discovery of underlying structure reduce reliance on large training corpora and on expensive compute.

In discussing human‑centered AI, Posse said machines and people excel at different tasks and that the best systems will combine machine insight with human expertise rather than attempt to emulate human cognition. The panel did not present independent benchmarks or peer‑reviewed results onstage; Posse’s claims describe her company’s internal performance comparisons.