Gianluca Truda

This is a synthesis of ideas I’ve overheard, read, discussed, and learned first-hand. It draws on over three years exploring LLMs for shaping code. My initial notes for this were part of an internal workshop I hosted for our engineering team at TORTUS Health, where we use and improve these techniques every day as we work to eliminate error in medicine. The shifting bottlenecks With the explosion in LLM capabilities, producing syntax got cheap. But good judgement did not. In many ways, AI makes the easy part easier and the hard part harder. ...

A deep learning algorithm I developed for training diffusion models on tabular data under differential privacy guarantees.

This project applied statistical learning techniques to an observational Quantified-Self (QS) study to build a descriptive model of sleep quality. A total of 472 days of my sleep data was collected with an Oura ring. This was combined with a variety of lifestyle, environmental, and psychological data, harvested from multiple sensors and manual logs.

Winning EF’s Bio x AI hackathon with a multimodal LLM for lab protocol automation.

We all use LLMs to make code, but there are at least two distinct modes of working with them. Both are powerful, but conflating them is dangerous. It leads to burnout, loss of agency, security breaches, and disastrous outages. Instead, we want to enable better ideas, executed faster, at higher quality.

As a learning exercise, I recently implemented simple-jax-nn, a simple neural net for the MNIST dataset, written with JAX. This post is my notes on using JAX (as a PyTorch / NumPy user) from working through the JAX docs and a NN tutorial. Simple JAX neural net I managed to get it to use my MacBook’s GPU/Accelerator through the Metal support in jax-metal. It’s a little bit tricky, but will prove useful in future transformer projects that can make use of the hardware. ...

This week, whilst troubleshooting video encoding discrepancies for a film, I was reminded of the immense wizard power of debugging: the art of systematically bisecting your way to the root cause of a problem, then devising the cleanest fix. This is an art that cannot be taught and which few professions lead you to learn. Even among those that do, such as computer programming, only the most diligent practitioners come to master it, because it is at once unnatural and primordial. We can be sure that many humans will never in their lives enact this discipline, but also that it is fundamental to our success as a species. ...

A sorcerer’s guide to wielding chaotic and powerful magicks.

Embracing the exponentials and vibecoding an entire 3D game in a few days with Aider, o3-mini, and Claude 3.7 sonnet.

Notes on the interconnection of generative AI’s two leading paradims.

The development stories and roadmap for Orbital, my physics-based 3D simulator for realistic orbital maneuvers that runs in your browser with WebGL and Three.js.

A byte-sized guide to developing your taste and agency at the world’s best programming retreat.

How I spent my 12 weeks at the Recurse Center annihilating my skill issues, shipping projects, and having fun.

How I learned Svelte and outsmarted Vercel to make a delightful frontend to OpenAI’s DALL-E … just for my mom.

Enter richify.py: a real-time Markdown rendering tool that supports streaming input. Built with Rich and shipped effortlessly with uv.

Links to the projects, games, blog posts, and tools I shipped in my 12 weeks at RC.

Week 44 of 2024. Week 12 of RC.

Week 43 of 2024. Week 11 of RC.

Week 42 of 2024. Week 10 of RC.

Week 41 of 2024, Week 9 of RC.

How I built a command-line LLM on Apple Silicon for RC’s ‘Impossible Stuff Day’

Week 40 of 2024, Week 8 of RC.

Week 39 of 2024, Week 7 of RC.

I started writing weeknotes as part of learning generously when I started at Recurse. But the distinction between Recurse and non-Recurse has increasingly blurred and I’d like to continue making weeknotes after I never graduate. To that end, I’m trying out a new format for my weeknotes. Tell me what you think! See previous weeknotes via the weeknotes tag. Updates Decided to extend to a 12-week batch at the Recurse Center. You can continue following my progress by filtering for posts with the Recurse tag. “QS Rubicon”: After 8 years of doing extensive self-tracking and ML-powered Quantified Self projects, I’ve decided to just stop collecting most data. I have plenty to say about the whole experience, so it should probably be its own blog post or talk. But the immediate feeling of being “out of control” is still something I’m trying to reflect on. Inputs / outputs Wrote and published 4 posts Make games people play: What Paul Graham doesn’t tell you about delighting 7-year-olds. AImong Us: a reverse Turing Test game: Can you outwit a gang of LLMs? What OpenAI’s o1 really means: The AGI is dead. Long live the AGI. RC Weeknotes 05: Move fast and make games. AAAHHH! The Creative Coding prompt at RC this week was “Aaaahhhh!” I whipped up something between a click trainer game and a psychological torture device. It also works on mobile if you like torturing yourself on the go. Participated in an intense Rock Paper Scissors tournament at RC. It involved multiple rounds of submitting a Python bot (with no dependencies) that played 200 rounds against some simple bots as well as the other participants’ bots. I managed to finish 3rd using a third-order Markov model with a stop-loss heuristic to switch back to the Game Theory optimal random play style. Made massive improvements to my dotfiles for better macOS package management with advanced Homebrew wizardry, and better LSP configuration for Python and Markdown in NeoVim. Improved PR to llm with tests. Try AAAHHH full screen at aaahhh.vercel.app ...

The AGI is dead. Long live the AGI.

Week 5 of Recurse: Move fast and make games.

What Paul Graham doesn’t tell you about delighting 7 year olds.

Can you outwit a gang of LLMs?

Week 4 of Recurse: Rust, AI tips, TDD, and finally cooking with gas!

Rustling up some skills and making addictive games in week 3 at the Recurse Center.

Projects and flaming hot takes from my second week at the Recurse Center.

Notes from my first week of Recurse or: How I learned to stop nerdsniping and love Rust.

I’ve joined the Recurse Center (RC) for the Fall 1 ‘24 batch.

A dynamic realtime raytracer in pure JavaScript you can run in your browser.

Initial thoughts on transforming bioengineering into software engineering.

An open source newsreader app I built using LLM embeddings to semantically prioritise arXiv and Github updates.

The past 18 months in AI have been a whirlwind-clusterfuck of acceleration, innovation, and chaos.

We used LLaVA and Claude to build a crazy multimodal LLM health assistant. OpenAI made it redundant the next week.

Why some sail and others drift.

There is a trifecta of variables that just about everyone building generative AI systems cares about — cost, quality, and latency. As is often the case in computing, there is a trade-off between these properties. And all three are bought with the currency of performance.

Making 95% annual return on the hardest data science tournament in the world, for fun and profit.

My MSc thesis and what it aims to do for biomedical science.

How I almost outsmarted 3 pro forecasters using Bayes rule.

Trying to out-forecast Vox in 2020. What could go wrong?

A selection of my older projects (pre ~2019) transferred from an old website for posterity.