John Dunbar’s portfolio
I’m a 3rd year computer science undergrad at the University of Texas at Austin.
Email: johnjdunbar at utexas dot edu.
What I’m doing right now:
Currently doing research under Scott Aaronson on the theoretical hardness of finding backdoors.
Also running Austin AI Alignment.
Things I did in 2025:
Paper in progress: Mixing behavior in randomly initialized neural nets
Things I did in 2024:
Research Note: Finding backdoors is NP hard
A bunch of random machine learning experiements. e.g.
- How much self-repair is in GPT2 without layernorm? Answer: It’s like the same amount as normal but with higher variance. Kinda interesting since layernorm is the main cause of self-repair in normal models. (code)

- Played around with training SAEs. They can recover the inputs on the toy model of superposition, but the signal is very noisy. (code)
- How well does a random forest on pixels do on MNIST? They can get 97% test accuracy. (code)
Interned as an embedded systems engineer at Lutron. I made the telemetry system on their flagship device go from a hacky sidecar to a normal component.
Things I did in 2023:
Earned a Knuth reward check for finding an error in an one of his equations. MIT Technology Review dramatically calls it “among computerdom’s most prized trophies.”

Was a Software Engineering Intern at Lutron. I wrote software that replaced a $50,000/year DevOps product. It’s now used by our 600 developers. I also found and fixed a privilege escalation vulnerability.
Invented an ISA where hardware interprets a primitive but lisp-like language.
Ran a reading group for the UT Quantum Club
Made robot software to find people and serve drinks to them. We were a team of 3 working in a massive interconnected system of cameras, Linux, ROS, and thousands of libraries.
Wrote a neural network in Numpy with all the fancy stuff.
Things I did in 2022:
Made compiler instead of using my professor’s made-up language.
Was the hardware lead in a high-school robotics team (FTC 8565) and we won the world championship. Here’s a video.