I ran global R&D. Now I build with AI.
Twenty years of running global R&D organizations and portfolios at Samsung, written up as a handbook — plus what I'm learning as I build an AI-native way of working. I log my forecasts, and I say so when I get them wrong.
Running R&D in the AI era — lessons and thinking from the work
What happened, what I learned, and what I think now — from running global R&D organizations and portfolios, in six parts. Each entry opens with the decision it helps with and ends with where it doesn't apply.
0 · Start here
1 / 11 · Why global R&D sites
1 / 42 · Choosing and allocating projects
0 / 43 · From advanced tech to product
1 / 44 · Running a site
0 / 55 · Working with HQ and executives
0 / 46 · Running R&D in the AI era
0 / 4

What to give each lab — split by technology area, not by project
When you divide work across overseas labs, why splitting by technology area and letting one lab own it end to end tends to beat splitting by project — language, culture, retained knowledge — and what it costs.
What this handbook is, and who's writing it
What I mean when I say my expertise is global R&D — and the four promises this handbook makes: only what I've done, what it cost, where it won't work, corrections in the open.
My first design failed — what took commercialization from 10% to 50%
An advanced-tech program that started with open topics did not work as intended. Narrowing the target window and setting direction first took commercialization from 10% to 50% — and here is what it cost.
See the record
Latest writing

If not productivity, how do I explain who I am?
Watching a dad ask his little boy "Why did you do that?", my first thought was "that kid will do well at work." Using productivity as a tool and explaining yourself through productivity turn out to be two different things.
Motivation for people, verification for AI
Most leaders have asked themselves, before handing someone a task: can I just let them run with it, or do I need to walk them through every step? So where does an AI agent sit on that map?
I logged 230 times my AI was confidently wrong.
I caught 22 of them. "All 40 tests passed." "That file isn't used anywhere." "I fixed both copies." When you…
Who runs this site
I started as an engineer bringing up mobile chipsets and went on to run an overseas R&D center. This is where I keep what I try, learn and apply.
At Samsung, from systems software to global R&D operations — now with the Security PMO at Coupang
People working across the Mobile Business portfolio I led at Samsung Research Ukraine
Raised the commercialization rate of overseas R&D centers’ advanced-technology projects from 10% to 50%
Led the world’s first mobile security certification (MDFPP) at the Ukraine center, and in the Technology Strategy team brought first-in-mobile and best-in-class technologies into products
Latest briefing
- Broadening access to AI for scientists
By opening its frontier models free to 200,000 researchers, OpenAI aims to become academia's default choice of tool for writing papers and analyzing data.
- Hugging Face and OpenRouter join forces to make open-weight models easy to run
Developers can run open-weight models (AI released so anyone can download and run it) found on Hugging Face directly through OpenRouter's single API, making it easy to swap and compare models without setting up their own servers.
- Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics
With Lambda adopting co-packaged optics (CPO) switches, which attach optical components directly to the switch chip, NVIDIA's networking approach has begun to reach real cloud clusters.
AI Ecosystem Briefing
Daily AI signals, each rated by strength of evidence. Published in daily, weekly, and monthly editions.
Since Jun 15, 2026: 153 daily · 19 weekly · 2 monthly
AI methods you can try now
Picked from my AI Frontier briefing — each one checked before it goes up.
- How-to · 11.12 Gather your skills and AGENTS.md in one repo and point every other repo to it→ one session can then work with the whole system in view
- Opinion · 11.03 Treat video generation as next-frame prediction and judge it on physics and geometry, and you measure how well a model understands the world, not whether it suits someone's creative taste→ Gain an evaluation axis that scores model quality on physics, geometry and object behaviour instead of 'do I like it'
- How-to · 10.20 Don't settle for a decision model's default 0.50 threshold→ tune it on your own labelled data and the same model catches failures it used to miss