Writing
-
Labels are the expensive part. Why models that generate their own supervision, learning by observing rather than by being told, are the route to systems that actually generalise.
-
A read of the RLTA paper: an RL agent that steers an LLM to a chosen output, not just a forbidden one, and why that means model output can never sit on the trusted side of a boundary.
-
AGI research and MLOps practice barely talk to each other. What the pipelines, monitoring, and data discipline would have to look like to carry something more general.
-
A ground-up primer on what AI actually is and why it matters, for people who have been nodding along in meetings and would rather understand it.
-
An RNN and a transformer built from scratch in PyTorch, trained on the same data with the same budget. They finish in a dead heat, and the reason why is the useful part.
-
A full MLOps pipeline on a laptop: CIFAR-10 in PyTorch, experiment tracking with MLflow, data and model versioning with DVC and GCS, and CI through GitHub Actions.
-
Written after OpenAI’s first Dev Day. Where scaling deep LLMs meets reinforcement learning, and why the reward function is the part nobody has solved.
-
Gradient descent on pixels rather than weights. What the Gram matrix throws away to turn a painting into a texture, and why the two losses can never both be satisfied.
-
Twin networks that learn similarity rather than classes. Measured on digits it was never trained on, against the raw-pixel baseline almost nobody bothers to run.