About
Stephen Welch is the creator of Welch Labs, a YouTube channel that explains AI, mathematics and science through animation rather than conversation. He started it in 2014, frustrated with the material available on neural networks, by making his own: Neural Networks Demystified, a short series that builds and trains a network in Python. He holds engineering degrees from Georgia Tech and UC Berkeley, led deep-learning vision work for driver-assistance and autonomous products as VP of Machine Learning at Autonomous Fusion, and has taught computer vision at UNC Charlotte. He ran the channel on the side for a decade before going full time, and loved math as a child but fell out of love with it in college — he blames the teaching, not the subject, and the channel is his answer to it.
Key Contributions
- Started Welch Labs in 2014 with Neural Networks Demystified, building and training a complete neural network in Python on screen
- Made Imaginary Numbers Are Real, the channel's most successful series, and published it as a book
- Wrote The Welch Labs Illustrated Guide to AI (2025), a 376-page first volume running from the perceptron to AI video generation, with code and exercises
- Led a team building deep-learning vision for driver-assistance and autonomous products as VP of Machine Learning at Autonomous Fusion
- His videos in this collection draw the arguments other channels talk through: AlexNet and the moment we stopped understanding AI, the compute-optimal frontier, sparse autoencoders, LeCun's JEPA, and why gradient descent does not get stuck
6 conversations they hosted
Connections
Grant Sanderson
CollaboratedMathematician & Educator (3Blue1Brown)
In the summer of 2025 Sanderson handed his own channel to Welch: the 3Blue1Brown video on how diffusion models actually generate images and video was produced by Welch Labs, with Sanderson giving editorial feedback over several months. He introduced Welch as 'a true kindred spirit of 3blue1brown on YouTube'. The pairing is the obvious one, made official — two channels that explain mathematics by animating it until the idea can be seen rather than taken on trust.
3blue1brown.substack.com · x.com
Yann LeCun
In conversationExecutive Chairman & Co-founder, AMI Labs; Professor, NYU
Welch's two-part series on LeCun's billion-dollar bet against language models was made with LeCun's help and carries his voice: Part 2 opens with Welch asking whether JEPA-based approaches will overtake the vision-language-action models now driving robots, and LeCun answering 'Oh, absolutely… VLAs are doomed.' Welch closes hoping to interview him again in a few years and see how JEPA has fared, which makes the series a dated bet as much as an explanation.
youtube.com · youtube.com
Chris Olah
KindredCo-founder & Interpretability Lead, Anthropic
Read as a pairing: Olah's life's work is making a network's insides visible, and Welch's channel keeps returning to the pictures Olah's research made possible. His AlexNet video opens on an activation atlas, and his interpretability video takes its title from Olah's own analogy — the features no one has yet extracted are the dark matter of interpretability, and sparse autoencoders are the telescope. One builds the instruments; the other shows a general audience what they see.
youtube.com · youtube.com · distill.pub
