☆ you've reached the homepage of ☆

Joshua Fonseca Rivera

I'm an Anthropic Fellow trying to figure out what's going on in language models. My latest paper shows that models can be fine-tuned to detect and name steering vectors injected into their own activations (COLM 2026). Before that, I spent many years as a software engineer.

me.jpg
Josh Fonseca Rivera smiling in a woven poncho in front of Machu Picchu
me @ machu picchu ⛰

publications.htm

ai_experiments.htm

An experiment in making the grokking phase transition — where models suddenly generalize long after memorizing — arrive sooner.

training dynamics · generalization

A scrape-clean-finetune pipeline on viral science videos that took a channel's click-through rate from 5% to 15%.

fine-tuning · data pipelines

Testing LLM reasoning against ancient paradoxes of vagueness. Do models understand "heaps" of sand?

reasoning · philosophy

engineering.htm

Artblock ⭐ 340

A browser extension that replaces every ad on the internet with fine art from public museum and NASA collections.

JavaScript · browser extension
golf.vim ⭐ 295

A keystroke-golf game inside Vim: transform the starting text into the target in as few keystrokes as possible, scored live against a leaderboard.

Vim script · daily challenges
vim-bujo ⭐ 206

A minimalist bullet-journal task manager that lives in Vim.

Vim script · CLI tooling

guestbook.htm

Sign the guestbook and I'll email you when a new paper or post goes up. No spam — pinky promise.

····.··.·· ·:·· V:······