It’s hard to leave Microsoft Research. Or OpenAI. Or Yoshua Benjio’s lab.
And yet, Edward Hu has done all three. (To be fair, he makes noises about finishing his research in Benjio’s lab.)

At Near Horizon, we talk a lot about edge: what gives a founder an unfair advantage. Edward’s edge is his restless mind and curiosity, and his ability to find simple and elegant solutions to complex problems. Think: LoRA.

Edward invented LoRA in 2021 at Microsoft, collaborating with OpenAI on GPT-3. The core problem: fine-tuning big models was slow, expensive, and increasingly untenable. So he looked to the math. The result: low-rank adaptation, a now ubiquitous and extraordinarily efficient way to adapt an LLM (or other large pre-trained model) to a new context. 

Same story with µTransfer. Edward dove deep into the math of scaling laws and infinite-width neural networks to develop a framework for predicting how training dynamics scale across model sizes. µTransfer enables scale-agnostic hyperparameter tuning, making the world’s largest models less brittle and more efficient to train. It’s become a critical tool for companies looking to save millions on large-scale training runs by transferring learnings from small models.

Another clear edge is his willingness to walk away from a sure thing (e.g., a massive job at OpenAI) and follow his passion. So now Edward is building his own thing – not a single company, but a portfolio of projects.

What Edward's Building Now:

Two projects in particular stand out:

• The Compute Exchange – Think NYSE, but for GPU auctions. Weekly bids, spot pricing, a real-time marketplace for the world’s most valuable commodity: compute. The early signs are good. Buyers are getting better prices, suppliers are competing in real time, a growing marketplace.
  
  
• AI IP Infrastructure – We won’t say too much (yet), but the thesis is simple: you can’t keep extracting value from human creativity without compensating the creators. AI doesn’t get to be a parasite. The world needs an attribution layer. Something durable. Something fair.
  

The Founder Archetype: Rick Rubin in Researcher Form

What makes Edward different isn’t just his technical depth. It’s that every time he solves something hard, he moves on. From µTransfer to LoRA to… whatever’s next. He mentioned Rick Rubin’s The Creative Act as the book that best captured his thoughts about personal reinvention, and it fits. Novelty-seeking and cross-disciplinary curiosity are a feature, not a bug.

This jibes well with what we’re seeing from our best founders. They bring a combination of broad curiosity that allows them to learn from everywhere, with at least one superpower. In Edward’s case, it’s the math chops to build elegant solutions to extraordinarily complex problems, with others it might be distribution, or a product insight.

If that sounds like you, find us here at Near Horizon.

‍