Joshua Fonseca Rivera
I study character training and model welfare: whether a model's consent matters, whether training removes preferences or only behavior, and whether self-reports can be trusted.
Before this, I was a software engineer for six years, most recently lead mobile engineer at Taekus, a fintech company.
- Role
- AI Safety Research Fellow, Anthropic Fellows Program
- Location
- Berkeley, CA
- Local time
- Pacific Time
Papers
-
Steering Awareness: Detecting Activation Steering from Within
Models learn to detect steering injected into their own activations (95.5%, 0 false positives), yet become more steerable.
COLM 2026
-
Item Response Theory for AI Safety
Largest psychometric study of safety evals (8 benchmarks, 192 models); cuts eval cost 97–99% and flags sandbagging.
arXiv 2026
-
Measuring Activation Control in Large Language Models
Most LLMs can steer their own activations enough to partially evade latent-space monitors.
arXiv 2026
-
Touch&Fold: A Foldable Haptic Actuator for Rendering Touch in Mixed Reality
Best Paper Honorable Mention (top 5%).
CHI 2021
Experiments
-
A Simple Method for Accelerating Grokking
Let the model overfit first, then apply Frobenius norm regularization. Tested on modular arithmetic.
Half the steps of Grokfast
-
Replacing Animal Crossing's Dialogue with an LLM
Reads and writes GameCube memory in the Dolphin emulator. No game code is modified.
319 stars
-
Training AI to Find Austin's Most Dangerous Streets
An XGBoost model trained on 2,757 bike crashes, with SHAP values to explain each prediction.
AUC 0.670
-
Fine-tuning GPT-4o for YouTube Titles
Trained on scraped and cleaned titles of viral science videos.
Click-through rate 5% → 15%
-
AI vs. The Sorites Paradox
Asked several LLMs which pile sizes of sand count as a heap.
No sharp threshold
Projects
-
Neuron Persona Library
Every MLP neuron in nine open language models, pushed up and down on its own, then asked “Who are you?”. Search the replies, or open any neuron to read what the model said.
4.4M neurons
-
Artblock
A browser extension that replaces ads on web pages with fine art from public museum and NASA collections.
340 stars
-
golf.vim
A daily Vim challenge: edit the starting text into the target text in as few keystrokes as possible. Scores go on a live leaderboard.
297 stars
-
vim-bujo
A bullet-journal task manager for Vim.
205 stars