Placeholder
Daily Planning
A tiny operating board for today, this week, and the thing that matters most.
The cost of creating is falling.The utility of understanding is rising.
Learning a wider set of mental models of the world, and what we can build next.
ABOUT
Hello! You're here early
I’m Jack Min Ong. I was a founding research engineer at Prime Intellect, and was there from its $10K valuation on paper to the most recent valuation of $1B.
My work has spanned language model pre-training and post-training, embedding and diffusion models, compute verification, and site reliability engineering.
I am currently on sabbatical exploring many different topics such as immunology, pharmaceuticals, robotics, agriculture, shipping, aviation and mining.
This site is still being drafted, so a few things are placeholders for now. Feel free to ping me, tell me you caught the early version, and let me know what you think!
SF / KLTennisPokerModel TrainingInferenceGPUs
Talks
Papers
Arcee Trinity Large Technical Report
INTELLECT-3: Technical Report
PRIME-RL: Async & Decentralized RL Training at Scale
TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference
SYNTHETIC-2: Four Million Collaboratively Generated Reasoning Traces
INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning
SYNTHETIC-1: Two Million Collaboratively Generated Reasoning Traces from DeepSeek-R1
INTELLECT-1 Technical Report
OpenDiLoCo: An Open-Source Framework for Globally Distributed Low-Communication Training
Jina Embeddings 2: 8192-Token General-Purpose Text Embeddings for Long Documents
Image-based plant identification with taxonomy aware architecture
Awards
- Gold Medal
Google AI4Code: Understand Code in Python Notebooks
- Silver Medal
Benetech — Making Graphs Accessible
- Silver Medal
Lux AI Season 2
- Champion
Programming League National 2022
- Champion
Deans Cup Coding League 2020
- Gold Tier
USACO 2018–2019 Season
- Gold Medal
Malaysia Computing Competition 2018
Writing
Notes from the frontier, sorted like a small newsroom.
Fused MoE + LoRA: Fitting Heterogeneous Adapters in a Single Forward Pass
Designing a fused kernel that applies per-expert LoRA adapters inside the MoE dispatch/combine cycle.
Notes on Blackwell SM100: What Actually Changed for Kernel Writers
A practical breakdown of what SM100 means if you're writing CUDA kernels — TCGEN05, UMMA, TMEM, and what you can ignore.
KV Cache Recompute vs. Offload — When Does Each Win?
Benchmarks and intuition for choosing between KV cache recompute and CPU offload in LLM serving.
Shipping INTELLECT-3: Lessons from Async RL on 512 H200s
What we learned training INTELLECT-3 with prime-rl's disaggregated async RL architecture.
Mini Apps
Small tools for planning, charts, and explainers.
Live
World Clocks
Find meeting overlap across SF, New York, Europe, Malaysia, and whatever timezone gets added next.
--:--
SF
--:--
NY
--:--
DE
--:--
MY
Placeholder
Chart Studio
Quick visual sketches for throughput, money, habits, and systems that need a shape.
Placeholder
GPU Explainers
Little interactive diagrams for kernels, memory, inference, and cluster behavior.
Placeholder
Field Notes Index
A future shelf for daily logs, public scratchpads, and visual breadcrumbs.
Placeholder
System Pulse
A small dashboard for live-ish status, rituals, or the machines that keep the site moving.
Uptime: 0d 0h 0m
Mem: -- / 141 GB HBM3e
Lore
Manhwa chapters for the arc underneath the work.
The polished version can become illustrated panels. For now, these are chapter placeholders: a loose timeline with room for mystery.
Roots and Rooms
Family, friends, and the first version of a life that feels designed on purpose.
Prime Labs
Training systems, GPU clusters, and the strange joy of making throughput visible.
Curiosity Engine
A chapter for questions that refuse to stay tidy.
Unknown Future
The page where the map ends and the work keeps going.
Long Arc
A placeholder for the far chapter, still private, still being earned.