Mike Wu
Ph.D. Computer Science, Stanford Co-founder, Built Different

I teach machines to see how people move.

Abstract.I'm a computer vision researcher turned founder. Before my PhD, I built infrastructure for training computer vision models at Meta. At Stanford I worked with Noah Goodman on how models learn visual representations without labels. Since then I've shipped ML at Sanas and Pointable (acquired by Apple). Now I'm building a vision model for human movement at Built Different, and pointing it at basketball, volleyball and soccer.

Release angle
—
Release speed
—
Apex
—
Entry angle
—
Result
—
Session FG
0 / 0
Fig. 1. Simulated pose and ball tracking across four sports, rendered live in your browser. Ball flight is real projectile physics: the readouts are measured from the tracked ball and the dashed curve is the predicted path. Long courts are drawn compressed past the break mark, with true distances labeled. Built Different pulls this kind of signal out of real game footage.

In the lab

2,000+citations
20+papers at NeurIPS, ICLR, AAAI
2best paper awards (AAAI, EDM)

On the court

3M+actions captured
5,000+hours of footage
30+countries with users
§1

Now: Built Different

Consumer sports products built on one shared model of human movement.

Built Different makes next-generation consumer sports tech. Underneath every product is the same vision model: it reads bodies, the ball and the court from everyday footage, then turns that into stats, highlights and feedback for players.

Sports are a hard, honest test for physical AI. Motion is fast, bodies overlap, cameras shake, and the ground truth is public: the ball went in or it didn't.

We're backed by a16z speedrun.

HooperBasketball. Shot tracking and highlights from your phone.
SpikerVolleyball. The same movement model on a new court.
StrikerSoccer. Taking the model to the pitch.
Hooper · pose + segmentation
Fig. 2. Hooper on an outdoor court: per-player masks, 2D pose and backboard detection running on real footage.
Mike Wu presenting on stage at the MIT Sloan Sports Analytics Conference
Fig. 3a. MIT Sloan Sports Analytics Conference
Mike Wu speaking with a microphone at a16z speedrun demo day
Fig. 3b. a16z Speedrun
§2

Trajectory

I've worked on getting cameras to understand motion since high school, starting with a blind-spot detector built from a camcorder.

LabBuild
  1. 2011–13

    Blind-spot detector · high school

    Learning to drive, I was scared of my blind spot, so I hacked together a car detector from a camcorder and optical flow. Follow-up work used edge-based optical flow and time-to-collision for autonomous navigation.

    Intel ISEF 3rd placeSiemens finalistXSEDE best student poster
  2. 2012–16

    Yale · B.S. Computer Science, with distinction

    Plus a term at Oxford in 2015, taking machine learning with Nando de Freitas.

  3. 2016

    Lattice Data · software engineer

    Weakly supervised classifiers built on the DeepDive project.

    Acquired by Apple
  4. 2016–17

    Meta · research engineer, Applied ML

    Tools and infrastructure for training computer vision models.

  5. 2017–22

    Stanford · Ph.D. Computer Science, advised by Noah Goodman

    Self-supervised visual representations, contrastive learning and multimodal generative models. Also built AI that gives educational feedback to students at scale, which landed in the New York Times.

    NSF GRFPStanford Interdisciplinary Graduate FellowAAAI outstanding student paperEDM best paper
  6. 2022–23

    Sanas · ML consultant

    Transformer-based generative speech models for voice conversion and accent translation.

  7. 2024–25

    Pointable · chief scientist

    A platform for building and optimizing custom RAG and LLM applications.

    Acquired by Apple
  8. 2025 → now

    Built Different · co-founder

    Using computer vision and AI to give every athlete the kind of analysis that only pros used to have. We train a visual sports model to understand basketball, volleyball, soccer, and all forms of movement.

    a16z speedrunHooperSpikerStriker
§3

Selected publications

Representation learning, generative models and vision, with applications in education and interpretability.

  1. Conditional Negative Sampling for Contrastive Learning of Visual Representations
    M. Wu, C. Zhuang, M. Mosse, D. Yamins, N. Goodman. ICLR, 2021.
    Vision
  2. On Mutual Information in Contrastive Learning for Visual Representations
    M. Wu, C. Zhuang, M. Mosse, D. Yamins, N. Goodman. arXiv, 2020.
    Vision
  3. Viewmaker Networks: Learning Views for Unsupervised Representation Learning
    A. Tamkin, M. Wu, N. Goodman. ICLR, 2021.
    Vision
  4. Improving Compositionality of Neural Networks by Decoding Representations to Inputs
    M. Wu, N. Goodman, S. Ermon. NeurIPS, 2021.
    Representation
  5. Foundation Posteriors for Approximate Probabilistic Inference
    M. Wu, N. Goodman. NeurIPS, 2022.
    Inference
  6. Multimodal Generative Models for Scalable Weakly-Supervised Learning
    M. Wu, N. Goodman. NeurIPS, 2018.
    Generative
  7. Zero Shot Learning for Code Education: Rubric Sampling with Deep Learning Inference Outstanding student paper
    M. Wu, M. Mosse, N. Goodman, C. Piech. AAAI, 2019.
    Education
  8. Variational Item Response Theory: Fast, Accurate, and Expressive Best paper
    M. Wu, R. L. Davis, B. W. Domingue, C. Piech, N. Goodman. EDM, 2020.
    Inference
  9. Beyond Sparsity: Tree Regularization of Deep Models for Interpretability
    M. Wu, M. C. Hughes, S. Parbhoo, M. Zazzi, V. Roth, F. Doshi-Velez. AAAI, 2018.
    Interpretability
  10. Differentiable Antithetic Sampling for Variance Reduction in Stochastic Variational Inference
    M. Wu, N. Goodman, S. Ermon. AISTATS, 2019.
    Inference

§4 · Correspondence

Working on physical AI or sports? Let's talk.