austxu.dev

Austin Xu / field notes 01—04

I chase the edge cases where models meet machines and markets.

Austin Xu is a researcher at StarAI, UCLA, building faster inference paths, bluffing agents, and tools for uncertain systems.

Explore the work Read the AMD story
uncertainty∞
01 / inference02 / deception03 / volatility
Los Angeles / UTC−07Research × systems × play
Signal / audited+3.095%geometric mean decode gain on an RX 6700 XTtrace the evidence

Selected work / three technical worlds

Small systems, sharp edges.

Each project starts with a constraint: a consumer GPU, hidden information, or a market that changes its mind.

01 / GPU inference / 2025

What happens when a consumer GPU becomes an inference laboratory?

Three percent, honestly.

A Vulkan decode investigation on an RX 6700 XT, where the result is less a record than a repeatable hot-loop improvement with its failure modes intact.

Open case study

02 / Strategic deception / Gen5

Can an agent learn when to bluff?

The tell is in the timing.

A playable 1v1 Coup bot trained to make decisions with incomplete information, pressure, and a very small action space.

Open case study

03 / Market volatility / research demo

Reading volatility as a regime, not a number.

A surface that changes its mind.

A live calibration lab for Heston-style dynamics, designed to make parameter uncertainty legible before it becomes a chart.

Open case study

Working principle

Make the uncertainty inspectable.

Benchmarks are arguments. Interfaces are arguments too. I build the instrument panel alongside the experiment so someone else can see what moved, what held, and what did not survive contact with reality.

A little more context

Open channel / 04

Have a strange system
worth looking at?

[email protected] github.com/austxu

The bear is a personal reference, not a logo. No IKEA affiliation.