Hello! I’m Ed, I’m a Member of Technical Staff at Anthropic. Previously I worked as a Software Engineer at Apollo Research where I built software for evaluations of LLM agents. I have also worked on the problem of faithful chain-of-thought reasoning as an external contractor for Anthropic and as part of the MATS programme.
Prior to that I was a Senior Machine Learning Engineer at Speechmatics where I led a team working on representation learning in the audio domain. We think we have the best speech recognition in the world (you can try it out free here).
Posts & Papers
Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought
Culmination of my research during my time at SERI MATS / as an Anthropic contractor. We investigated bias-augmented consistency training (BCT), an unsupervised fine-tuning scheme that trains models to give consistent reasoning across prompts with and without biasing features.
Transformer Memory Arithmetic: Understanding all the Bytes in nanoGPT
How much memory is really used when training a transformer? In this post I take a first principles approach to calculating the steady state and peak memory usage during training and by deriving estimates of memory usage for the different components see how close I could get to the...
My Favourite Terminal Workflow Tricks
The following are my 4 favourite terminal scripts and shortcuts that I have found useful for optimizing processes and procedures that arise frequently during my day job as a Machine Learning Engineer. These tricks are mostly applicable for zsh / tmux users although the Command line copier is more...
Gaussian Processes - Interactive Demo
This is an interactive implementation of a gaussian process written in javascript that runs in the browser. To use simply click on the main chart to add 'observations' and watch the model update its predictions! Three of the most common kernels are implemented: the radial basis function kernel, the linear kernel and the periodic kernel. It is...
