hiloop›hiloop› offers Autoresearch

hiloop sells an autoresearch service that turns a measurable target and a scoring function into thousands of directed experiments, returning a better model or algorithm once improvements verify. The campaign spans data, post-training, continual learning, prompts, tools, and systems, and the company reports a state of the art on an autoresearch benchmark published by Andrej Karpathy at 0.9016 validation bits-per-byte, the median of 25 confirmation runs against the published 0.9109.

Campaigns run on hosted infrastructure or inside the customer cloud, with persistent memory, experiment lineage, and statistical verification built in. The offering comes from a two-person San Francisco company in Y Combinator's Summer 2026 batch, and early access is arranged directly with the founders at [email protected].