Animated research Watch the mathematics move. Twelve animated explainers on ML security, distributed systems and the mathematics behind them. Published research, engineering practice and open directions are labelled separately. Each film states its sources and assumptions.
Who trained a model, and whose is it? SecurePoL is the research. The two films after it each go deeper into one question it raises.
01 SecurePoL: Proving How a Model Was Trained A training record and a watermark checked together, so a forged training claim has to pass both. Published research: IEEE Access 2025 and my dissertation. 02 Could This AI Be a Copy? Deeper: how a faint mark is read out of noise, and why the threshold needs calibrating. A teaching model, not the paper's detector. 03 Signing an AI Model Deeper: three kinds of model watermark, and what access each one needs to check. From the comparison in my dissertation and SecurePoL. Research, design studies and practice Each card says what it is: published research, a design experiment, or engineering practice.
04 Who Improved the AI? Five organisations train one model: how to evaluate an update, reward the contribution and keep data private, with a blockchain as one mechanism among several. Systems reviewed in my survey, IEEE Access 2023. 05 AI Meets the Deadline Where should inference run? Test asynchronous proposals, expiry, fallback and contention around a 60 Hz loop. Design experiment, not yet published. 06 Inside a Level D Flight Simulator How a simulator earns a pilot's trust, and what AI would have to keep to run inside one. Engineering practice; every rule cited from the regulation text. 07 Security Events in Turkish Text Turn local-language posts and news into an investigation queue without a labelled dataset. Published research: ICISSP 2021. Open research notes Questions I am working on rather than results I have shown. Feedback and collaborators welcome.
08 Can a Contract Trust an AI Answer? A proof shows a model was computed correctly; it does not show the answer is right. 09 Can a Market Coordinate AI Agents? When agents propose the work they are paid for, how do you tell useful work from reward hunting? Foundations Teaching films on results that belong to other people, animated because the derivations are worth seeing move. The papers are cited on each page.
10 Gradient Pinball How neural networks learn: the learning-rate cliff, the √κ momentum speedup, and saddle points. After Polyak, Nesterov and Dauphin et al. 11 Block Race Why "6 confirmations" is a probability, not a promise. After Nakamoto, §11. 12 Redundancy Reactor Voting makes independent faults rare; a shared cause sets a floor no number of copies removes. Classical TMR and the Ariane 5 inquiry. 