Daniele Pace

Ceci n'est pas une pipe — personal motif

Currently, I'm Research Manager for the fifth MARS cohort, and finishing my research project on CoT (un)faithfulness detection using attention probes at LASR Labs, supervised by Noah Y. Siegel. I've also been part of the previous iteration of MARS, working on harmfulness representation in different training checkpoint stages, under the mentorship of Lorenzo Pacchiardi. Before that, I spent almost 4 years in industry research, as part of the Prediction for Autonomous Driving research team at Stellantis. I hold a bachelor's and master's degrees from Politecnico di Torino, with the final two semesters spent at Aalto University, where I worked on my final thesis under the supervision of Jaakko Lethinen.

Since the beginning of 2025 I've been interested in AI safety, eventually making it my full-time job. My research spans different areas, with mech interp at its core. I believe that understanding the structure of LLMs' internal representations, and how they produce intelligent behaviors, is essential for alignment; moreover, studying a kind of intelligence quite unlike ours is intellectually fascinating. I'm particularly interested in applying white-box methods to control and monitoring. I'm against every simplification of complex concepts like morality, alignment, feature.

Projects

  1. qweDante 2025

    Personal project mimicking Dante's writing style with Transformer language models.

    Full details GitHub code

  2. Mechanistic Interpretability with SAEs 2025

    Interpreting the internal activations of Gemma 3 with sparse autoencoders.

    Full details GitHub code

  3. Yla 2025

    A local AI assistant that runs entirely on your own machine.

    Full details GitHub code

  4. Computer-Assisted Solr Query Generation 2022

    Built during the Pi School fellowship: assisting users in composing Solr queries from natural language.

    Full details

  5. Nonlinear Climbing Video Indexing 2021

    Master's thesis with Aalto University: a web application retrieving fragments of climbing videos via nonlinear queries on spatial regions and body key points.

    Full details GitHub code

  6. CiccioNet 2021

    Controlling Wikipedia navigation through hand gestures recognised in real time.

    Full details GitHub code

  7. Few-Shot Neural Network 2021

    A few-shot classifier for character recognition, with a live in-browser demo.

    Full details GitHub code

  8. AI-driven Art 2021

    Images produced exploring the state of generative neural networks before the DALL·E era.

    Full details