Latest

A world model is a learned simulator. Give it a state and an action and it predicts what happens next. This write-up follows the idea from Ha and Schmidhuber's V+M+C agent through Dreamer, MuZero and TD-MPC, to Genie, Cosmos and V-JEPA, the drift that comes with imagining too far ahead, LeCun's case for predicting representations instead of pixels, and why robots need all of this.

Post-training has moved from imitating human labels to learning from verifiable outcomes. This write-up covers the RLVR objective, GRPO and the choice of KL penalty, advantage collapse and the methods that recover the lost gradient, reward hacking and noisy verifiers, outcome versus process reward models, the production data flywheel that turns failures into weight updates, long chain-of-thought and latent reasoning, and the OpenRLHF stack that runs it all.

Modern LLM inference is memory-bound, and the KV cache is the reason. This report walks through why the cache exists, how large it gets, the attention changes that shrink it (GQA, DeepSeek MLA), how engines like vLLM and SGLang manage it, how quantization and eviction compress it, and how disaggregated serving moves it between machines.

Sixteen questions about sparse models, each one the thing you would ask after hearing the last answer: what the router really is, why experts do not learn topics, why sparsity stops paying once you batch, and what it costs to serve one.

How I would design the pipeline that moves a legacy database onto Azure: bulk load and CDC, a medallion lakehouse with a quality gate that actually stops things, star schemas and SCD2, a feature store, and the security, monitoring and cutover plan around it.

How to tell whether an LLM system works, and how to keep it working: test cases, graders, judges, groundedness, CI gates, live signals, traces, latency, schemas, injection defense, and PII masking.

An agent that gets better on its own has to change something about itself, and something has to check the change. This post walks the six things it can change, the checker that decides whether any of it worked, the ceiling it runs into, the ways it breaks, and the loop teams actually run in production.

What happens inside a serving engine: why the GPU fills up with something other than the model, what paging the cache buys you, why reading a prompt and writing an answer fight each other, and which knobs settle the fight.

What a senior data scientist must know cold: statistics, experiment design, causal inference, metric craft, modeling judgment, and evaluation, plus one section on the 2026 market and one on interview prep.

Forty production questions on agentic systems, answered plainly: evals, guardrails, quality measurement, observability, security, and token economics.