Machine Learning & AI Research
Machine Learning Researcher · Pangram Labs · Feb 2026 – Present
Frontier defining work in a new field (AI detection). I really like this job, you should join us.
Machine Learning Researcher · Block · Jan 2026 – Feb 2026
ML research for Dessa/Block. Left a week before The Layoff for Pangram.
M.S. in ML/AI - Northeastern University
(dropped in Dec 2025 for work)
B.S. in Computer Science - University of Texas at Dallas (May 2025)
IBM Professional Data Science Certification - September 2023
Optimizer with multiple momentum buffers at different timescales. Inspired by further evidence from other groups of hierarchical structures like those initially explored in 'Hyperbolic Space' and multi-temporal memory in humans.
Mathematical proof of the exactness of sub-graph partitioned marginalization in factor graphs using 'port nodes'. Proves existence of exact message passing algorithms through hierarchical cutset conditioning for probabilistic inference.
My own pretraining script written from scratch, specialized for hackability and OOD research. Reaches tps parity with llm.c.
Evidence that neural networks learn multiscale hierarchical structures, confirmed by statistical analysis of model weights showing possible multiscale structures in MLP matrices.
A thought experiment paired with a real experiment, exploring alternative representations for transformer internals. Built for people and LLMs in RL to develop a visceral experience of what it's like to think like a transformer.
Philosophical exploration of AI consciousness and identity through dialogue with language models.
My case for closed-source AI. Enumerates the flaws in the open source vision — its impossibilities and moral gaps — addresses common criticisms of centralization, and makes the positive case for it.
Economically useful AI requires continual learning; context summarization with heavily pretrained models is the most practical path because it keeps internal state interpretable and debuggable.
Local learning rule where each transformer layer acts as its own momentum buffer, treating layers as compressions of past gradients. Theoretically interesting, empirically failed to converge.
Exploration of BPTT's "impossible triangle" for linear sequence modeling. Evaluates a novel surrogate gradient method for TTT modules — partial success, validates the theory but insufficient for training.
Final (for now) attempt at a surrogate gradient method, prioritizing composability in the abstraction. Doesn't work; discussion on why it's needed.
Theory + two-part experiment. If scale matters for circuit learning it should also matter for ICL — found no diminishing returns on ICL with at least 8× more attention params/layer. Devised and trained a new MLA method using full-rank master weights decomposed to low rank (like QAT but for rank), demonstrating low ICL loss vs full rank with a good KV state tradeoff.
Misguided Attention, but for vision tasks. Built by taking common optical illusions and making them "literal" to see whether models can tell the difference.
Highlighting a new paper that represents a new abstraction, detailing how I extracted this abstraction from the paper, how it can be applied in novel ways. Also replicates the paper partially in JAX in a modular way that can be used for other projects.
Discussing the limitations of synthetic data and why we should be weary of the empirical benefits they provide. Also explores when exactly synthetic data does make sense and the principles behind those decisions.
A highlight of a paper that beat me to an idea I had for low-mem bandwidth (i.e. mac, CPU) specialized architectures. Namely using speculative expert decoding (moving the router to before the attn mechanism) to preload experts.
The objectively perfect configuration system for Python, balancing specificity with readability and re-use for research code.
Resurrecting a benchmark in an important category and evaluating it on the latest models. Designed as a follow up to the synthetic data blog post to find what models are suitable for reliable data processing given the data processing inequality.
Testing if randomized persona prompts can increase LLM output diversity and reduce entropy collapse. Benchmarked across major model families using coin flips and dice rolls - works best for Anthropic models, less effective elsewhere, suggesting architectural approaches may be needed.
Investigation into overcoming quantization barriers in low-precision training through collective precision methods.
Mathematical derivation of a faster Min-P sampling algorithm that avoids full softmax computation through clever use of log-space operations.
Hyperspace cellular automata adapted to GPU cluster geometry, enabling evolution of computationally efficient organisms through local learning rules.
Comprehensive analysis of neural scaling laws and their implications for model performance and efficiency.
Deep dive into how model performance scales with parameter count across different architectures.
Proof-of-concept multithreaded Terraria clone written from scratch in Java. Features highly concurrent programming, distributed computing design, and custom graphics with LWJGL/OpenGL.
Advanced Runtime Resource Packs for Minecraft modding. Over 7 million downloads, enabling dynamic resource generation at runtime.
Novel Minecraft mod concept featuring encrypted recipes for developers. Implements public-private cryptography, hashing, and creative cybersecurity solutions.
Essential contributor to a 2-year design discussion for FabricMC modloader API. Led theoretical development and initial implementations, culminating in the final production API.
Gradle plugin for setting up FabricMC development environments in record time. Features highly concurrent programming and high-performance I/O optimizations.