George Kailas launches Aethon Fund, an AI-powered hedge fund launching in August

Aethon Fund

George Kailas is launching Aethon Fund, an AI-powered hedge fund targeting an August close, with September as the latest fallback. Capital is already committed through a fund-of-funds agreement plus traditional LP capital, though no figures were disclosed.

Kailas has been building toward this for years. He started in value investing at 17, taught himself accounting to get his first job, then grew skeptical of fundamentals-based approaches and moved toward AI. A mortgage algorithm he built in Excel eventually parlayed into an AI company. In 2019 he founded Prospero AI on the thesis that data would prove more durable than models over time.

“We're basically taking these signal libraries from Prospero. We're expanding into my R&D backlog... we essentially feed it into some of the world's best LLMs. We create thousands of strategies with that. And then from those strategies, we figure out intelligent ways to combine them to work in many different markets to opportunistically allocate amongst them. That's what Aethon is.”

The core architecture

Aethon inherits Prospero's signal libraries and expands them into R&D Kailas says a startup environment didn't allow him to pursue. Those signals feed into large language models to generate thousands of distinct investment strategies, which are then combined and allocated across markets opportunistically. The goal is a system that can distribute capital across 10, 20, or more strategy combinations simultaneously, with the mix shifting as conditions change.

The design philosophy at both Prospero and Aethon prioritizes compression: complexity in the back end, simplicity at the user interface layer. Kailas also claims that public-facing testing of signals at Prospero generated a kind of feedback loop where high public trust in a signal's predictive pop created self-reinforcing follow-on behavior, which informs how Aethon thinks about strategy validation.

Alternative data

On the question of novel data sources, Kailas is skeptical that much edge remains in the celebrated examples: satellite parking lot imagery, real-time credit card swipes, camera feeds. He says most of that is exhausted. The more interesting development, in his view, is data containerization, a concept he attributes to CarbonArc, a firm run by someone named Kirk. The idea is packaging large alternative datasets into smaller components that can be accessed only in the ways relevant to a specific training iteration, reducing training costs while enabling broader data exploration.

Growth plan

Kailas describes a phased expansion: test strategies within existing signal libraries, expand those libraries into new asset classes and geographies, feed the results back into the compression system, and increase the number of active strategy waterfalls over time. The explicit goal is becoming more distributed, not concentrated, as the fund scales.