Software engineering:
I have a strong software background. I have spent about 15000 hours total writing software: 7000 hours for Python, 5000 hours for C and C++, 1000 for Typescript and Javascript, another 1000 for other languages (Rust, Ruby, PhP, Web frameworks). Around 2000h doing devops work (CI, CMake, building packages, etc). I feel confident in developing good code and I can solve many sorts of unexpected and "this shouldn't be happening" issues, thanks to my systems engineering background. I have worked with LLMs as a contractor for several projects and have a good practical grasp of it. Currently, I'm doing an interpretability course in Cambridge's CamLAB that covers TransformerLens and by the end of it I hope to have a sturdier fundamental understanding of the area. I also believe I'm fairly good at explaining technical subjects to others, thanks to my experience in mentoring people in software.
AI safety and alignment:
I estimate to have spent around 2000h working with AI safety related topics, either by attending to courses and discussion groups, reading papers or writing my own experiments and ideas. Currently, I'm developing research in the area of Cooperative AI in which I model the interactions between AGIs and humans and intend to find paths for cooperation among both [1],[2]. This study has been accepted as a poster to the Machine + Behavior conference [3], happening later this month.
Paper re-implementation project:
I have re-implemented the paper "Simulacra of Human Behavior" applying my own ideas about how to better represent theory of mind to agents as well as drawing from other concepts from game theory, psychology and agent based modeling. I made the decision to keep the code closed-source due to concerns that it could potentially be misused by malicious actors. Nonetheless, I shared my insights through a comprehensive blog post [4], and this work was also featured in an art exposition in New York City [5] and a talk I delivered in Hamburg [6].
Code projects:
Code for my current research [2]. I have introduced the ability to model environments with agents with mixed observability. It's a MARL project using Jax.
For a project where I was a member of a group you can look at my work at Anaconda in the PyScript software [7], it's a library that aims to integrate Python to the web browser.
For a project I fully developed by myself you can take a look at a kernel I wrote for Jupyter Lab [8], that allows users to directly run SQL code in the notebook and immediately plot graphs with their queries.
[1]
https://docs.google.com/document/d/1D8Iqgs8bGY8IHxKLb9XfsbAenFGPpgf9gsUjABqxW1c/edit?usp=sharing
[2]
https://github.com/marimeireles/MARL-mixed/tree/master
[3]
http://marimeireles.com/Navigating-the-Complexity-of-AI-Cooperation.pdf
[4]
https://techforgoodresearch.substack.com/p/yoasi-a-generative-llm-based-universe
[5]