Builder · Investor · Pitcher
Landon Hershberger
16. High school junior. I build AI models, manage money, and have been training baseball for almost 1000 days in a row.
Featured project
A pitching-coach AI, built on quality data
Democratizing access to the best pitching coaching in the world
Fine-tuning Qwen3.8-27B (open-weights, Apache 2.0) on ~5,000 rows of curated coaching data.
Research
Open-access, peer-reviewed papers. CC-licensed full texts.
Pitch data
MLB Statcast / Baseball Savant pitch-level data.
Biomechanics
OpenBiomechanics motion-capture data.
Coaching content
Paraphrased from free public articles and coaching videos, plus coaching dialogues.
Safety rules
High school pitch-count rules for all 50 states plus DC, and Pitch Smart youth guidelines.
My own training logs
My own training logs are part of the data behind the pitching-coach AI. The training and the building feed each other.
Eight benchmarks, run against six frontier models.
I am still in progress on building this model, I've been working on it for 6 weeks and should have some SOTA results in the coming weeks.
-
Step 1
Gather data and build benchmarks -
Step 2
Fine-tune the 27B model with different hyperparameters each time -
Step 3
Run all 8 benchmarks against top frontier models -
Step 4
Compare the benchmark scores, and rerun the loop
What was hard about the AI
Gathering quality data
This is the most important, yet the hardest step. The quality of the data can make or break a model and is the only way that a small model can compete with frontier models that have 5-10 trillion parameters.
Benchmarking
Benchmarking is one of the longest tasks because each question and answer has to be hand written by a professional. This part can't be automated because then it is capped by the model you choose to write the benchmarks.
Choosing the right model
I had to find the model that was on the Pareto frontier for capability and parameters. Qwen 3.8 27B happened to be perfect because it scores a 34 on AA intelligence index which is comparable to GPT 6 Luna, while being small enough to run on a MacBook.
Fine-tuning
Fine-tuning was one of the most fun parts of this project. Adjusting the hyperparameters when training a model can make a big difference on its performance, so finding the right configuration is crucial and can only be done through trial and error.
Data center land deal
I negotiated a data center land deal from $5M to $20M.
Though I am not able to talk about many details of this endeavor, this achievement is one I am most proud of. I spent many nights going back and forth with Codex on the value of this specific property and data centers, and found out that the price that was offered ($5M) was not worth the real value. I compiled a packet of why the land was worth $20M, and game planned how to negotiate with this new price target. I was in the room for the negotiations, and a few weeks later, the papers were signed for $20M. This taught me how to negotiate in a high-pressure situation, as well as how to value items that may seem to not be worth much on the surface.
Investing ~$310k AUM
I manage about $310k in assets, about $95k of it my own.
Investing started for me when my dad would buy me and my siblings stocks instead of gifts for our birthday and Christmas presents. The power of compounding and exponentials always fascinated me, so I’ve been learning about stocks and investing since I was 6. My own portfolio has grown to ~$95,000 and I’ve been managing my older brother's account (~$215,000) for 4 months. In the last 3 years, my annualized rate of return has been ~26%. Managing other people's money has taught me how to make decisions that affect others and how to balance risk and reward.
~1000 days · Right-handed pitcher
I have trained every single day for almost 1000 consecutive days.
I started taking my baseball journey seriously on February 18, 2024. I set out to become the best pitcher I could physically be and I am still on that journey. It has been the hardest thing I have ever done due to the consistency and physical fatigue that training demands. I have had multiple injuries so far, opening my eyes to other opportunities such as building AI and getting into the weeds on markets. Even if my baseball journey doesn’t pan out, it has taught me how to outwork everybody and what it means to sacrifice EVERYTHING to accomplish a mission. I wouldn’t have the mindset or work ethic if it wasn’t for playing a competitive sport, and in many ways it is the reason why I am applying for the academy.
Writing
I write about AI and markets on Substack.
How will frontier labs make a profit?
June 29, 2026
As open-source models get good enough for everyday use, I argue frontier labs will earn their money in three places:
- The Longevity Dividend: Royalties on breakthroughs in drugs, materials and health.
- The Red Queen Effect: Enterprises can't afford a weaker model than their competitors.
- Labor as a Service: Charging per task instead of per token.
Why the Academy