Hi, I’m Vadim. I’m an AI engineer (since 2016) and I’m transitioning to AI safety. I started doing natural language processing before transformers were invented, and saw the better part of the recent AI boom, which made my p(doom) quite high.
I want to help deconfuse intelligence and alignment. I prefer empirical safety, especially interpretability, which brings theory and observation closer. My CV’s here — contact me at hi@wadimiusz.me to work together or hire me.
Check out my:
- CV,
- github,
- rarely updated substack with some of my longer texts,
- linkedin,
- google scholar.
Research
- Where does a transformer keep its state? (October 2026). An interactive note, with a demo, on where a small transformer keeps the state of the automaton it has learned. Part of my current work on reading automata out of transformer weights.
- AlphaEvolving sampling (AI Safety Camp 2026, MoSSAIC project). Evolutionary search for programs that predict a model’s output distribution from its weights, scored against plain sampling. Slides, code.
- Discourse probing of transformer-based language models (master’s thesis, HSE, 2020). Layer-wise probing of BERT, RuBERT and GPT-2 for discourse structure. Code.
- Tracing cultural diachronic semantic shifts in Russian using word embeddings: test sets and baselines (Dialogue 2019, first author). Paper, code and data.
- A study of machine learning algorithms applied to GIS queries spelling correction (Dialogue 2018, first author, with Ivan Bondarenko). Paper, code.
Code
- nanogpt — a GPT-style transformer written from scratch (RoPE, SwiGLU, RMSNorm), installable as a package. The base for my interpretability experiments.
- theorem-marketplace — post a crypto bounty for a theorem in Lean, and anyone can claim it with a machine-checked proof. Built solo end to end: Flask app, Lean verifier, Chainlink oracle. Live at theorem-marketplace.com.