Research at the Edge of Reliability

Tuesday, Feb 3, 2026 | 5:30 PM – 8:30 PM
Axiom's Office, San Francisco

Research for AI that ships. Talks on the unglamorous problems that kill products: slow accelerators, silent hallucinations, context limits, and metrics that lie.

The Lineup

Charles Hong · PhD Candidate, UC Berkeley (Shao Lab)
Autocomp: Turning LLMs into Performance Engineers
Hardware accelerators power modern AI, but programming them is slow and specialized. Autocomp turns LLMs into automated optimizers — 5.6× faster than vendor libraries, 1.9× over expert-tuned code on AWS Trainium. The schedules are human-readable and transfer to similar workloads.
MLArchSys 2025 Outstanding Paper · IEEE LAD Fellow · Apple Scholar

Mathew Vanherreweghe · Member of Technical Staff, Logical Intelligence
Kolmogorov-Arnold Geometry: Detecting Hallucinations via Network Structure
Neural networks develop geometric structure during training. Mathew's research analyzes this self-organization to detect out-of-distribution inputs and predict hallucinations — before they reach users.
Previously Pebblebed

Jialian Wu · Senior Research Scientist, AMD GenAI
VideoMarathon & VideoSeek: Hour-Long Video Understanding
Current video models fail beyond a few minutes. VideoMarathon processes hour-long video at 1 frame per second — enabling training and inference at unprecedented scale. VideoSeek takes it further: an agent that actively seeks relevant content via tool use, hitting SOTA across benchmarks.
NeurIPS 2025 Spotlight · PhD, SUNY Buffalo

Jack Langerman · Independent Researcher*
When Benchmarks Lie: Aligning Evals with Human Judgment
How do you measure whether a 3D reconstruction is "good"? Thousands of expert preference judgments reveal that complex graph-based metrics correlate worse with human preferences than simple alternatives — and off-the-shelf VLMs perform at chance. Practical frameworks for evals that actually predict what humans care about.
ICCV 2025 Highlight · Co-leads CVPR Workshop on Urban Scene Modeling
*Now at Apple

Mandi Zhao · PhD Student, Stanford
[Topic TBD]
Details coming soon.
Stanford · Previously BAIR, Meta AI, Weights & Biases

Agenda

5:30 PM — Doors
6:15 PM — Talks + Q&A
7:45 PM — Discussion + Networking

Capacity strictly limited. Application required.

Hosts

🐉 Eragon
The AI Operating System built for companies — a reasoning engine that understands, learns, and acts across an entire organization. We embed deeply into the flow of business and connect data, people, and processes into one intelligent, context-aware workspace. AI built for billion-dollar decisions, not busywork.

🏢 Axiom
An early-stage venture firm that brings together the best AI practitioners in-market today, to invest in founders using AI to change the world tomorrow. We believe in AI for the Real World, and back visionaries that are building insanely useful and usable products for 8 billion people around the world, not just the elite 80 million. Backed by Khosla Ventures, a16z, General Catalyst, and Iconiq.

Host: Sandhya is a Silicon Valley veteran operator and investor building with and investing in AI since 2011. She sold her first company to Cisco for $1.5B, was VP of Product at Skype when it was acquired by Microsoft for $8.5B, and was the first investor in Groq and AI unicorns including FirefliesAI and HelmAI.

The AI Collective is a global non-profit building the human layer for the AI age. We unite 150,000+ leaders, builders, and stakeholders across 100+ forums worldwide to democratize the frontier, build trust, and coordinate how society navigates the rapid acceleration of technological progress.

All attendees and organizers at events affiliated with The AI Collective agree to our privacy policy and are subject to our code of conduct.

Eragon x Axiom Research at the Edge of Reliability