Jayson Lynch

  • Jayson Lynch is Head of Algorithms in the FutureTech Lab at MIT, working on predicting the future progress of algorithms development and understanding the fundamental limitations to improving computing performance. They earned their PhD from MIT under Erik Demain working on the computational complexity of motion planning problems, computational geometry, and games and puzzles. Afterwards Jayson continued researching computational geometry at the University of Waterloo. Some of their other work includes reversible algorithms to help unlock future highly energy efficient computers, as well as cache-oblivious and cache-adaptive algorithms.

  • We're planning three distinct projects this summer. Selected candidate will work on one of these projects. Given that each project requires different skill sets, we do not expect any candidate to be qualified for all three. Please refer specifically to the Desired Fellow Qualifications section to see the skills relevant to each individual project.

    1. The Progress in Algorithms Project seeks to understand the historic rate of algorithmic progress and use it to better understand current and future trends in improving computing performance.

    2. We want to use Mechanistic Interpretability to help understand LLM code generation. I'm particularly interested in program control flow (eg if, for, while, goto statements) and how control logic clusters in embedding space, whether it is the same across different programming languages (which have the same semantic notions while having different names or syntax), how programming is impacted when these features are removed from models, and how these features might relate to natural language conditionals.

    3. We have been studying the ability of language models to adapt to esoteric programming languages. The ability for a model to code in an unfamiliar language with minimal or no retraining is both potentially useful for software engineers but also worrisome as a component of self-improvement automation and rapid takeoff.

  • Check the mentor topics page to see the corresponding projects.

    1. We need candidates to assist both with algorithms literature surveys and with data analysis. For data collection the candidate should have a strong background in algorithms and algorithmic analysis, and this summer we are particularly interested in people with experience in quantum algorithms or in algorithmic statistics (including topics like sample complexity and lower bounds).

    2. Candidates should be familiar with techniques in mechanistic interpretability, in particular the ability to identify and ablate features in language models.

    3. Candidates should have some familiarity with LLM code generation and benchmarking. Ideal candidates will also have familiarity with a variety of standard and esoteric programming languages as well as programming language theory.

Head of Algorithms, MIT FutureTech