DeepJudge CTO and co-founder, ML researcher and YouTube educator
Yannic Kilcher
Profile
If you have ever tried to actually read a machine learning paper — the dense notation, the buried assumptions, the “we leave this to future work” hand-waving — then you have probably ended up watching Yannic Kilcher explain it instead. For years, Kilcher has run what is effectively the largest YouTube channel dedicated to walking through ML research papers line by line, and he does it without dumbing anything down. His format is unfussy: pull up the PDF, scroll through it, and reason out loud about what the authors actually did, what is genuinely new, and where the claims get shaky. For a developer trying to separate real progress from press-release hype, that skeptical, math-literate running commentary is worth more than a dozen breathless threads.
Kilcher earned his PhD in machine learning from ETH Zürich, working on the robustness of deep networks, NLP, reinforcement learning, and generative models, and spent time as a researcher at Google (in the AI/Language org in Zurich) during those years. That combination — academic depth plus industry exposure — is what makes his explanations land. He is not a science communicator borrowing other people’s understanding; he is a practitioner who reads the same papers the authors’ peers do and tells you which parts are load-bearing. Alongside the paper reviews, his recurring “ML News” segments became a genuinely useful pulse-check on a field that moves faster than anyone can follow.
Today Kilcher is co-founder and CTO of DeepJudge, a Zurich-based legal-AI company he started in 2021 with fellow ETH PhDs Paulina Grnarova and Kevin Roth (all three had done stints at Google). DeepJudge builds AI-powered search and document intelligence for law firms — helping lawyers find and reuse the institutional knowledge buried in their own archives. The company has raised real money on that thesis: a seed round led by Coatue and a $41 million Series A in late 2025 (led by Felicis alongside Coatue) reportedly at a $300 million valuation, with elite firms among its customers. It is a reminder that Kilcher is not a full-time explainer — he ships production ML for a living, which is exactly why his teaching stays grounded.
He is also a mover in open-source AI. Kilcher was one of the original initiators of OpenAssistant, the LAION-backed community effort to build an open, instruction-following ChatGPT alternative by crowdsourcing human-feedback data — work that culminated in a widely cited NeurIPS 2023 paper. That project matters to developers precisely because it tried to democratize the RLHF-style data that proprietary labs keep locked up. Kilcher is not without controversy (see below), and he can be provocative to a fault — but if you want to actually understand what changed in a paper this week rather than how to feel about it, he remains one of the most reliable voices to learn from.
Key Articles & Papers
OpenAssistant Conversations — Democratizing Large Language Model AlignmentControversies
GPT-4chan (2022). In June 2022, Kilcher fine-tuned the open GPT-J model on millions of posts from 4chan’s notoriously toxic /pol/ board, then deployed bots that posted roughly 15,000 unlabeled messages back onto the board over about 24 hours, and released the model publicly on Hugging Face. He framed it as a “prank” and a demonstration, and was genuinely interested in the model’s benchmark behavior — but the reaction was severe. Hundreds of researchers signed an open letter condemning the project as unethical, citing the unconsented human experiment on live users and the release of a model built to generate hateful content; Hugging Face restricted access. It remains a useful case study in AI-ethics-versus-hacker-culture tension: even a technically interesting result can cross a line when you deploy it on real people without disclosure. Kilcher engaged with the criticism publicly but largely stood by the technical framing while acknowledging the deployment was the contentious part.
Spotify Podcasts
YouTube
