@Fuqua

Projects, team wins, and moments from my MBA at Duke Fuqua.

Screenshot of the GraspGrade site

Robots can learn a lot from watching people, and human egocentric video is the cheapest robot data there is. But most of it is not useful for a robot hand. Fingers are hidden, objects are too big, or you cannot tell when the hand actually touches the object. Today teams pick clips by eye.

So Joonse and I built GraspGrade at the Meta Global AI Developer Hackathon. It watches a hand video, finds the moment the hand touches an object, cuts the clip there, and gives it a score from 0 to 100 for how well a robot hand could copy the grasp. It runs on video only, using Meta's Muse Spark and SAM 3.1.

We did not want to judge our own work by eye. We built two ground truths from real sensors: touch-sensing gloves for when a touch starts and ends, and Meta HOT3D motion capture mapped onto a robot hand for whether a grasp can really be copied. Over 42 experiments, we wrote down every test before looking at the results, and we reported what did not work too. We also released the sensor labels as an open benchmark so other teams can test their own filters.

A few things I learned:

Check it against something real

Building ground truth from sensors changed which method looked best. Without the glove and the motion capture, we would have picked the wrong one.

How you ask matters more than training

Asked directly if a robot hand could copy a grasp, the model was no better than a coin flip. Asked through a short checklist about the object, the same model got clearly better. Learning extra weights on top added almost nothing.

Write the rules before you look

We fixed every comparison and pass rule before scoring the test data. It kept us honest, including about the parts that failed, like our soft object filter.

Touch is still the missing piece

Our score can tell which objects are hard for a robot hand, but not which of two grasps on the same object is better. That is where touch data comes in, the same gap I heard about at Long Horizon.

See it live: Project site · Demo video · Try it · Code

Thanks to Meta and Cerebral Valley for the models, the credits and a great 10 days, and to Joonse for building it with me. Always happy to chat about robot data and physical AI.

Attendees at Long Horizon, Encord's Physical AI summit in San Francisco

I went to Day 1 of Long Horizon, Encord's first Physical AI summit, at Fort Mason in San Francisco. I sat in on six sessions with speakers from Skild AI, Physical Intelligence, General Intuition, Amazon Robotics, Google DeepMind, Boston Dynamics and more. Then I turned my recordings into a long Medium post.

A few takeaways from the summit:

Your answer depends on the data you hold

Everyone asked whether robots will get a ChatGPT moment, and everyone answered it differently. Each answer lined up closely with what data that company owns and how it makes money.

The next data frontier is touch

Video, games and simulation are slowly solving scale. But almost every speaker pointed at the same gap: force and touch, the things a camera can't capture.

99% is the starting line

In the real world, 99% still means millions of failures. The hard part is everything outside the model: evaluation, rollout, and how fast you find and fix failures.

Writing it up is how I learn

Going back through the recordings and digging into the details taught me more than the day itself. It's also why I keep writing my Physical AI data pipeline series.

Full write-up on Medium: Will Robots Ever Get Their ChatGPT Moment?

Thanks to Encord for putting together such a great lineup. Always happy to chat with anyone working on Physical AI.

Industry Education cabinet tile naming Inseok Lee on the Duke MBA Tech Club 1Y Cabinet

I've been selected to the Duke MBA Tech Club 1Y Cabinet, taking the Industry Education portfolio. My focus this year is simple: raise the overall tech capability of Fuqua students.

The gap I keep seeing

Most of my classmates come from finance, consulting, ops, or the military, not tech. They can name the AI headlines, but far fewer can explain how a tech company actually makes money, or what a model, an API, and a data pipeline mean for a business decision. That isn't a talent problem. It's an exposure problem.

What Industry Education means to me

Not more lectures. Practical reps. Workshops where students build something small with AI and ship it, sessions that walk through how a tech company really runs (the P&L, the org, the roadmap), and honest conversations with people who already made the move from consulting or finance into a tech role.

What I want to see by May

A first-year who can sit across from an engineering team in a customer interview and hold their own. More Fuqua students treating tech as a realistic destination, not a distant one. And a club culture where asking a basic question about tech is normal.

If you're a Fuqua student who wants to understand tech better, or someone in industry who'd be willing to talk to a room of curious MBAs, I'd genuinely love to hear from you.

Fuqua MBA students on the Bay Area tech trek

After Summer Term wrapped up, a group of us, first-years and second-years, put together our own tech trek to the Bay Area. We met 20+ people across Big Tech and startups during the trip, Duke Fuqua alumni and first-time connections alike, and came away with advice that was genuinely valuable.

A few takeaways from the tech trek:

What do I actually want?

It's easy to keep sprinting through a busy MBA schedule without stopping to ask this. This trip gave me the space to picture where I want to be in the next two years and beyond, and that made the path feel more real.

Surviving the AI era

Things are moving fast. Building AI-powered products, using AI in daily work, and staying on top of new developments feel like the baseline skills to keep up going forward.

Bias toward building, not just thinking

AI has lowered the barrier to building software so much. I want to keep shipping side projects and writing down what I learn along the way. Actually building something for Fuqua students right now, more on that soon.

Share your thinking, consistently

I've been writing on my Medium blog and LinkedIn, but I need to do it more often. Writing things out helps me think more clearly, and it's also how new connections happen. (I should move my content from Medium to Substack!)

English is a never-ending project

Even living and studying in the US, my English still has plenty of room to grow. Competing here as a non-native speaker means I can't let up on this.

Huge thanks to everyone who supported us and shared their advice so generously. I'll be back in the Bay Area often, and would love to meet more people there, alumni, MBA classmates, and future Fuqua students alike. Always happy to be introduced to anyone in your network.

Screenshot of the Fuqua Daily Tech News newsletter channel

Every morning before class, I break down the one tech story that actually matters. 150+ Fuqua MBAs read it.

Phase 1: The Gap I Saw

Most of my classmates come from finance, consulting, or ops backgrounds, not tech. Every day there's a flood of AI, robotics, and Big Tech headlines from TechCrunch, Bloomberg Tech, WSJ Tech, and The Robot Report, but no time (or context) to figure out which ones actually matter for business. I started reading all four every morning anyway, out of habit from my PwC days. I just started sharing what I found.

Phase 2: Curate, Don't Report

I don't summarize everything. I pick one or two stories, explain what happened, and more importantly, why it matters for business, strategy, or the industries we're studying. No jargon, no hype. Just enough context so someone with zero tech background can follow the conversation at lunch.

That's the value. Not the headline itself, but knowing why it's worth caring about.

Winning team celebrating after the Fuqua 3D Printing Case Competition

Our team of 6 had 48 hours to design, print, and pitch anything.

Phase 1: Decide, Don't Debate

We started with a detachable heel for open-toe sandals. Sounded clever, until we held the prototype. No one would actually buy this. We killed it, pivoted to a foldable clip-on cup holder, and moved on. The ideation was messy: six people, strong opinions, a whiteboard full of pros and cons. But endless discussion is just another way to avoid commitment. We made the call and shifted all our energy to execution.

Phase 2: Capture the Need, Then Make the Case

Most teams led with market sizing and financial projections. We did that work too, but kept it in the background. In the room, we made one bet: if the audience doesn't feel the problem, the numbers don't matter. So we opened with the product itself, let people hold it, and made them want it before we told them why it made sense.

That's what won. Not the spreadsheet: the moment someone thought, "I'd actually use this."