Wian Stipp

Technical

Working notes and math.

  1. The Mathematics of Hidden Markov ModelsHidden Markov Model mathematics from first principles: forward filtering, forward-backward smoothing, and Baum-Welch EM for parameter estimation.
  2. Self-taught reasoning via post-rationalization fine-tuning approximates policy gradient optimizationSTaR bootstraps system 2 reasoning by fine-tuning on self-generated rationales and post-hoc rationalizations, which approximates policy gradient RL on filtered correct trajectories.
  3. Variational autoencoders learn structured low-dimensional manifolds by optimizing the evidence lower boundVAEs force the encoder to map inputs to latent distributions rather than points, balancing reconstruction against a Gaussian prior via the evidence lower bound.
  4. Conway's Game of Life is embarrassingly parallel, so one CUDA thread per cell gives a 70x speedup over PythonPorting Game of Life from Python to C to CUDA at the AGI House Hardcore CUDA Hackathon: one thread per cell, a block-wide barrier to swap states, and why copying every frame back to the host halves the speedup.
  5. Dr Claude: Monte Carlo tree search plans the diagnostic interview whilst Claude conducts itHow we combined AlphaGo-style MCTS over a disease–symptom probability model with Claude 2 to ask the fewest questions on the way to a diagnosis. Third place at Anthropic's Claude 2 hackathon, built in 24 hours.
  6. Embedding-based entity resolution turns noisy LLM triples into a queryable knowledge graphKnowNet crawls a news frontpage, extracts triples with an LLM, merges near-duplicate entities in embedding space, and answers questions over the graph with cited source URLs.
  7. Coffee bean weighing scale from scratchBit-banging the HX711 from a Raspberry Pi Pico, mapping 24-bit ADC counts to grams with two-point calibration, and displaying live weight on a small SPI LCD.
  8. Monte Carlo dropout is approximate Bayesian inferenceKeeping dropout active at test time samples an approximate posterior over network weights, yielding uncertainty estimates and better predictions without changing the architecture.
  9. Tobit models handle left-censored data by combining cumulative probability mass with continuous normal densitiesDeriving the Tobit log-likelihood: a discrete probability mass at the censoring threshold for clamped observations, a normal density for the rest, joined by an indicator into one differentiable objective.