CV - Logan Thomson

logan@loganthomson.com | github.com/xycoord | loganthomson.com

Technical Skills

Core ML Engineering:

• Python [Projects 1,2,3,4,5,6]
• PyTorch [Projects 1,2,4,5,6]
• ML Experimentation [Projects 1,2,4,5,6]
  Experiment design, hyperparameter tuning, debugging training

Specialised Expertise:

• Transformers and Language Models [Projects 1,4]
• Reinforcement Learning [Projects 2,6,7]
  Focus on policy gradient algorithms & PPO
• Algorithm Optimisation [Project 3]

Active Exploration:

• Mechanistic Interpretability [Project 4,7]

Research Skills:

• Technical Writing [Projects 1,2,3,4,7]
  Clear explanations of complex topics (blog posts, teaching materials)
• Paper Re-implementation [Projects 1,2,4,5,6]
• Mathematical Rigour [Projects 1,2,4]
• Philosophical Reasoning [Project 7]

Projects

All independently designed and implemented:

1. Transformer Language Model [GitHub]
   From-scratch PyTorch implementation with KV caching and RoPE [Blog Post].
   Modular architecture emphasising code clarity and understanding.

2. Reinforcement Learning Course [Part 1] [Part 2] [GitHub]
   Teaching RL through rigorous mathematical derivations and implementations from first principles from policy gradients to GAE.

3. BPE Tokeniser [GitHub] [Blog Post]
   Optimised training implementation (hours → 13s) with systematic profiling.

4. Mechanistic Interpretability [GitHub]
   Reproduced “Toy Models of Superposition” experiments and trained SAEs (ReLU, TopK, BatchTopK) to extract their learnt features.
   Teaching Direct Logit Attribution with derivations and implementations (supports LayerNorm and RMSNorm) [Blog Post]

5. Master’s Research Project (Supervised by Ronald Clark) [GitHub] [Report]
   Fine-tuned diffusion models for image segmentation of transparent objects. Implemented and evaluated NeRF methods which learn how light bends in a scene.

6. PPO Implementation [GitHub]
   From-scratch PPO agent featuring key modern techniques, such as GAE and vectorised environments.

7. Wild Chameleons [Blog Post]
   Analysis of hypothetical Neural Chameleons (probe-evading models) that emerge in the wild as a consequence of an RL pressure.

8. DeepSeek Sparse Attention Deep Dive [Blog Post]

Team Experience

Co-founder at [The Grove], 2024

• Collaboratively designed and built recording studio/music venue to high technical standards.
• Translated technical requirements (e.g. fire safety regulations, acoustic design) into actionable tasks for the team.
• Successfully built thriving community space; exited to focus on AI safety.

Education

Oxford University, 2020-2024
Master’s in Computer Science and Philosophy (MCompPhil)
First Class

Relevant Courses - Ethics of AI, Computer Vision, Geometric Deep Learning, Machine Learning, Ethics, Philosophy of Mind, Philosophy of Cognitive Science, Law and Computer Science, Computers in Society