Tom
Yeh

In early 2024, Tom Yeh posted a hand-drawn exercise on LinkedIn showing how a transformer works - not with code, not with diagrams borrowed from a research paper, but with actual numbers written in colored pen on graph paper. He expected mild interest. What he got was a movement.

Within weeks, 25,000 people were following along. Within months, the number was 200,000 across LinkedIn and X. The AI by Hand newsletter on Substack crossed 62,000 subscribers - ranking in the top 100 technology publications on the entire platform. None of it was planned. All of it was earned, one calculation at a time.

Yeh is an Associate Professor of Computer Science at the University of Colorado Boulder, where he has been teaching and researching since 2012. He holds a PhD from MIT, where he built some of the early foundations of what we now call computer vision-driven automation. His Sikuli project - which let computers automate GUIs by looking at screenshots the way humans do - was downloaded over 200,000 times and spawned one of the most-cited HCI papers of the 2010s.

But the part of Tom Yeh that matters most to most people right now sits at the intersection of a Feynman quote and a box of colored markers. His guiding philosophy: "What I cannot create, I do not understand." In an era where AI has become a vending machine - put in a prompt, get out a result - Yeh insists on understanding the mechanism. The math. The actual numbers flowing through an attention head, a feed-forward layer, a normalization step.

His AI by Hand series does not treat the reader as a passive consumer. You work through it. You fill in the cells. You multiply the matrices by hand - or in an Excel spreadsheet if you prefer, since Yeh also built an entire GitHub repository of Excel-based deep learning exercises that garnered 1,300+ stars almost immediately. The exercises span everything from basic matrix multiplication and gradient descent to full transformer architectures, diffusion models, and the Mamba state-space model.

What makes this genuinely unusual is that Yeh drew every single one of those exercises himself - by hand, in color, on printed paper. Three hundred individual exercises across twelve workbooks. He did not outsource the aesthetics. He did not generate them. He sat down and created them, the same way he asks his students to engage with the material.

The response has surprised even Yeh. When students find errors in his solutions - and occasionally they do - he is reportedly delighted. It means they understood the problem well enough to catch the mistake. That is the point. Not passive absorption, but active engagement. Not trust in the black box, but mastery of what's inside it.

In 2026, Yeh launched AI by Hand Academy: a structured curriculum of video-based courses covering introductory and advanced AI topics, starting with a module on Agentic AI that walks through LLMs, RAG, agents, and tool use. The Feynman method, now with video. Over 130 founding members - recognized personally by their initials on the Academy page - joined in the early months.

Beyond the newsletter, Yeh leads the Imagine AI Lab at CU Boulder and co-directs the Center for the Brain, A.I., and Child (BAIC). His research sits at the junction of human-centered AI and interpretability - studying how people build, inspect, and reason about AI systems. His academic citation count sits at 7,488, with an h-index of 41, and funding from NSF, NIH, DARPA, and two private foundations.

He did not arrive at AI education through a single clean narrative arc. He grew up in Taiwan, completed his BSc at Simon Fraser University in Canada, earned both his MS and PhD at MIT, spent three years in postdoctoral training at the University of Maryland, and eventually landed at CU Boulder. The range of his research - computer vision, 3D printing, assistive technology, brain imaging, citizen science, AI ethics - makes more sense when you understand that he has always been more interested in what AI can do for people than in AI for its own sake.

The viral newsletter is not a pivot. It is the same instinct - toward legibility, toward inclusion, toward making the inaccessible accessible - that drove VizWiz in 2010, when he built a system that let visually impaired users photograph an object and receive a real-time spoken answer. Years before GPT-4V. Years before anyone used the phrase "multimodal AI" in casual conversation.

Tom Yeh is what happens when deep expertise meets a stubborn commitment to clarity. He is not simplifying AI. He is insisting that it can be understood - and then proving it, cell by spreadsheet cell.