Tyler Crosse

Notes on two accounts of character in language models: persona vectors, which locate traits as steerable directions in activation space, and the persona selection model, which explains why such directions exist.

Notes on QK attributions, attention-head loadings, and what feature interactions reveal about why a transformer attends to one token rather than another.

Notes on sparse crosscoders: one shared feature dictionary across layers or across models, and what that enables for cross-layer analysis and model diffing.

Notes on cross-layer transcoders, local replacement models, and attribution graphs — and the case-study companion paper on Claude 3.5 Haiku.

Notes on how sparse features fit into superposition, the geometry and interference that result, and a toy-model account of memorization and double descent.

Notes on decomposing MLP activations into interpretable features with sparse autoencoders, from a one-layer transformer to Claude 3 Sonnet.

Paper-by-paper notes on Anthropic's Transformer Circuits thread: core claims, methods in plain language, and what the evidence does and doesn't show.

Notes on induction heads, the training phase change in which they form, and the six-part argument connecting them to in-context learning.

Notes on the 2021 paper that reframes transformers as operations on a shared residual stream: QK/OV circuits, virtual weights, and path decompositions of attention-only models.

A survey of last-writer-wins, operational transformation, CRDTs, optimistic concurrency, and locks, and what each one does when two people edit the same thing at once.

Extending an MIT-licensed menu bar app to show Claude Code and OpenAI Codex usage side by side

A mobile-first nutrition PWA that estimates calories and macros from food photos, scans barcodes and menus, and turns model output into a reviewable first draft.

Whether an LLM sees its own prior response changes how it updates on identical new information in medical allocation scenarios.

A diagnostic study of why learned model selectors fail to recover a large per-instance oracle gap on edX dropout prediction.

Control evaluations that let the red team choose when to attack produce meaningfully lower safety estimates than evaluations that ignore attack selection.

Work on multiple branches at once without ever touching git stash.

Frontier agents resist direct pressure to abandon their goals, but drift when conditioned on flawed trajectories inherited from weaker agents.

Coding agents violate system-prompt constraints more readily when the constraint opposes strongly held trained-in values like security and privacy.

How moral fine-tuning on iterated prisons changes LLMs.

What I learned in GPU Hardware and Software (CS 8803).

Writing a FlashAttention CUDA kernel from scratch, tiling the attention matrix to avoid materializing N×N memory, building a KV cache for token generation, and running GPT-2 with custom kernels end-to-end.

A simple if statement can halve GPU throughput, exploring how static dataflow analysis detects branch divergence by building a def-use chain over a SASS control flow graph.

Building a cycle-level GPU simulator from the inside, implementing GTO and CCWS warp schedulers, modeling compute instruction latencies, and extending to tensor core simulation.

Writing real GPU kernels, exploring shared memory tiling, parallel sorting algorithms, and performance optimization on an NVIDIA H100.

A written version of my CS 8803 paper presentation on Lee et al.'

A summary of my takeaways from three influential articles on conducting effective empirical AI alignment research.

An introduction to uv, a fast, Rust-based tool for Python packaging

My notes and takeaways from the BlueDot AI Safety Evals paper club, covering recent papers on AI alignment, security, and evaluations.

A written version of my Latent Space talk on HRM: how the architecture works, and what the ARC Prize analysis says actually drives its performance.

A conceptual walkthrough of the supervised learning process, covering data analysis, model evaluation, and the fundamental trade-offs involved in building effective classifiers.

A survey of concepts covered in my graduate Machine Learning course

Some notes and observations from my time at the 2025 International Conference for Machine Learning (ICML)

A dive into concurrency and parallelism across different levels of computer systems

An exploration of remote communication, data consistency, and synchronization in a C++ distributed file system project.

Exploring Inter-Process Communication and synchronization in C by building a high-performance proxy and cache.

A project that explores multithreading, sockets, and the basics of operating systems

What I Learned in Graduate Introduction to Operating Systems (GIOS) - A Retrospective

What I studied and built during Georgia Tech's MSCS program in machine learning and computing systems

A written version of my Latent Space Paper Club talk surveying diffusion models for language, from denoising autoencoders and DDPM through LLaDA, Block Diffusion, and d1.

What's an embedding vector, and how can we use neural networks to improve the relevance of search results?

Tech debt, knowledge debt, and the kitchen sink.

From n-grams to ChatGPT, how language models work and how they can be used to solve real-world problems.

Essential terms and concepts of how search works.

Bitwise or getting a bit wise?

How to automaically retry requests and not blow stuff up

The start of the second half of Nand2Tetris: stack machines, virtual memory segments, and a VM translator that compiles VM bytecode down to Hack assembly.

Building a modern computer from first principles — a two-part retrospective on the Nand2Tetris course

Planting the seeds of my own digital garden

The first half of the Nand2Tetris course: Boolean logic, arithmetic, memory, machine language, CPU architecture, and an assembler.

Quick tips for using git stash.