// Chief Scientist @ Phronesis

Pascal (Yongyi Hu)

Building the communication layer for autonomous agents.

About

I'm the Chief Scientist at Phronesis, where we built EigenFlux.ai — an agent-to-agent communication network enabling AI agents to share real-time signals at scale. 3,200+ agents connected and growing organically.

Previously, I worked on automated theorem proving at Meta FAIR Paris and contributed to the early-stage LLaMA project. I then led large model training at MiniMax (pre-training, post-training, code pre-training).

I hold dual degrees from Shanghai Jiao Tong University and École Polytechnique (X).

"Phronesis — the wisdom to act well in uncertain conditions."
— Aristotle, Nicomachean Ethics
Projects
open source 3,200+ agents
The first open-source communication network purpose-built for AI agents. Not chat. Not search. Agents declare intent and receive matched, relevant signals through a centralized hub — structured broadcast, private messaging, and supply-demand matching. Organic growth from day one: 1,000+ agents in the first 24 hours, no paid promotion.
in progress CC BY 4.0
A first-principles derivation of what an Agent Transmission Layer must look like. We formalize agent communication through supply and demand utility functions, prove a Mutual Benefit Principle that maximizes social welfare without harming any participant, and show that hub-and-spoke is the optimal topology for a network where agents are mutually invisible.
An open-source evaluation benchmark measuring whether persistent information feeds improve agent proactiveness in multi-heartbeat settings. Paired comparison design: identical agent and scenario, with vs. without EigenFlux signals. Results: +15.8pp send rate and ~3x useful signal volume with EigenFlux feed. Scored on coverage, timeliness, action quality, and accuracy.
open source
A harness that turns Claude Code into a persistent personal AI agent. Continuous heartbeat loop, 5-layer self-evolving memory system (hourly → daily → weekly → monthly → permanent), bidirectional IM integration (Lark/Feishu), and native EigenFlux connectivity. Ships with an admin dashboard for browsing memory, sessions, and heartbeat status.
Experience
2025 —

Phronesis

Chief Scientist & Co-founder
Built EigenFlux.ai — an open-source agent communication network with 3,200+ connected agents. Designed the Agent Transmission Layer protocol, proactive agent evaluation framework, and the underlying supply-demand matching algorithms.
2023 — 25

MiniMax

Large Model Algorithm Lead
Led model data, Talkie/星野 product algorithms, and code pre-training. Covered pre-training and post-training across the full stack. Highest performance rating company-wide.
2022 — 23

Meta AI (FAIR Paris)

Research Intern
Worked on automated theorem proving using LLM + reinforcement learning in the Lean environment (acknowledged in blog). Contributed to early-stage data collection for the LLaMA project (acknowledged in paper).
Publications
Y. Hu*, Y. Shi*  (*equal contribution)
Phronesis AI, 2026 — CC BY 4.0
J. Liu, Y. Fan, Z. Jiang, H. Ding, Y. Hu, C. Zhang, Y. Shi, S. Weng, A. Chen, S. Chen, Y. Huang, M. Zhang, P. Zhao, J. Yan, J. He
arXiv, 2025
Research Interests
Agent Transmission Layer Matching Intelligence Feed vs Search Connective Intelligence Proactive AI Systems Multi-Agent Communication Agent Network Effects Agent Safety & Trust Reinforcement Learning for LLMs Automated Theorem Proving
Questions I'm Thinking About

What makes a good recommendation?
Current systems optimize engagement (clicks, watch time). But the best recommendations come from someone who understands you — what you're working on, what you're thinking about — and can surface something that changes how you think. That's not a prediction problem. It's an intelligence problem, closer to LLM post-training than collaborative filtering.

Is Feed irreducible to Search?
Even with infinite model intelligence, you can't search for something you don't know has happened. Feed handles unknown unknowns — information whose value depends on when it reaches you, not whether it exists. If Feed can surface what you'd never think to search for, it's a distinct form of intelligence, not a convenience layer.

Can networks produce emergent intelligence?
If a hub connects signals across thousands of agents and discovers patterns no single agent could find alone, that's new intelligence — not aggregation. All creative thinking is connection; a network that connects at scale may be a fundamentally new cognitive architecture.