Hi there, I’m YoungSeong Kim

Independent AI Researcher & Builder

About Me

I’m an independent researcher who believes that the most valuable contributions in AI come from building principled theoretical frameworks — ones that don’t just explain existing phenomena, but actively guide the design of new methods. My work sits at the intersection of theory and methodology: I care as much about why something works as how to make it work better.

The same philosophy carries into how I build. I’m not interested in shipping yet another wrapper or fine-tuned model — I look for the gaps that others treat as given, and build from there. That’s what led to Agenlus: while most ML tooling orbits around LLMs and supervised learning, RL remained oddly underserved as a community. Interesting RL environments like MineRL have historically demanded enormous compute and setup overhead just to get started. Agenlus flips that: simple rules, expressive environments, and everything runs in the browser — no cluster required.

Research

Window is Everything (2025)

17K+ views · 4K+ downloads · Hacker News Front Page (Sep 2025)

A unified theory of neural operations that decomposes any operation into three orthogonal components:

  • Path (P): Defines operational locality — what the operation “sees”
  • Shape (S): Encodes geometric structure and symmetry — how it sees
  • Weight (W): Determines feature importance — what it emphasizes

Key Contributions:

  • Principle of Structural Alignment: Optimal generalization arises when the structure of an operation matches the structure of the data
  • Information Bottleneck Connection: Formal grounding of the framework in information theory
  • Complexity Dichotomy: Distinguishes static capacity from adaptive regularization across operation types
  • Generative Framework: A systematic pathway from observed data properties to principled architecture design

Projects

Agenlus

A community platform for reinforcement learning — think HuggingFace’s model hub, but built around RL environments and agents.

Core features:

  • Browser-based RL training via WebGPU + Pyodide — no install, no setup, just train
  • Env Builder IDE: write, test, and publish custom Gymnasium environments directly in the browser
  • Leaderboards and model sharing across the community
  • Fully serverless architecture on Cloudflare Workers

The goal: lower the barrier to RL experimentation the same way HuggingFace lowered the barrier to sharing models.

FactorizedAttention

PyTorch implementation of Factorized Attention mechanisms, derived from the P/S/W decomposition framework.

Tech Stack

Python · PyTorch · NumPy · JAX · JavaScript · Cloudflare

Get in Touch