I'm a technical generalist.
I previously worked on eliciting the structure of representations in CLIP. I also built product and worked on financial infrastructure at AtoB and Stripe. I'm also an Emergent Ventures grant recipient.
You can email me here: [email protected].
Writing
-
Training on mistakes improves problem solving performance
April 25, 2024
-
Ireland's unique promise for nuclear energy
April 11, 2022
-
Fast information flows result in homogeneity
January 26, 2022
-
Babble mode in voice dictation
January 26, 2022
-
High volume beats almost anything else
December 30, 2021
-
Compounding rewards as an argument for moral virtue
December 26, 2021
-
Time perception
August 10, 2021
See all posts...
Interpreting and Steering Features in Images
I trained a sparse autoencoder on CLIP and found over 50,000 highly interpretable and steerable features.
Training on Mistakes Improves Problem Solving Performance
Research post on the effect of error correction in a maze solving agent
On the GPU Memory Hierarchy
Guide to modelling algorithm performance on GPUs
In the Wild West of Crypto Scams
Training a model to front run bad actors in crypto for fun and profit
Evaluating steering vectors
Experiments in mechanically finding activation vectors to influence how GPT2 responds
System Synthesis
Experimental user interface for AI feedback on system design
Tome
A transpiled educational programming language with English-like syntax