Practical, deeply explained guides to machine learning, statistics, deep learning, and the systems that put them to work.
In the intelligence-layer post, the SSP’s floor-price section reframes “what should the minimum price be?” as a stack of binary classification problems: given this ad request, will at least one bid...
Bid Landscaping series - Part 1: Learning the Win Curve from the Auctions We Lost · Part 2 (this post): From Win Curve to Bidding, Pacing, and Its Limits Part 1 built the object: for a given cont...
Bid Landscaping series - Part 1 (this post): Learning the Win Curve from the Auctions We Lost · Part 2: From Win Curve to Bidding, Pacing, and Its Limits Recently, while working for a client in t...
So far in this series, quantization has been a serving story - shrink a trained model so it fits on cheaper hardware and runs faster. In the linear quantization article we set up the uniform grid a...
In the previous article we saw how AWQ protects salient weight channels by pre-scaling them before INT4 rounding. The mechanism was direct: spend more of the grid’s resolution on the channels that ...
Most of machine learning runs on first-order information: the gradient tells us which way is downhill, and we take a step. That is enough to train enormous models. But a whole family of methods, fr...
In the previous article we saw why a uniform quantization grid wastes precision on bell-curved LLM weights. We also saw how non-linear methods (quantile quantization, NF4, k-means) fix the problem....
A 70-billion-parameter LLM stored in 16-bit floats needs roughly 140 GB of memory, more than most GPUs can hold. Quantization shrinks the model by replacing those 16-bit floats with much smaller in...
Why Quantization Matters for LLMs Modern LLMs are enormous, not just in parameter count, but in the memory and compute they demand at inference time. A model with billions of parameters stored in ...
The Generalized Linear Model as the common modeling language behind t-tests, ANOVA, proportion tests, and CUPED - and the design, estimand, and inference layers a GLM does not supply.