All writing

123 long-form pieces on architecture, training, inference, agents and safety.

All Model Architecture19 Training & Alignment22 Inference & Serving18 Agents & Orchestration11 Reasoning & Evaluation26 Safety, Security & Governance7 Platforms & Practice20

Platforms & Practice 7 min

Five Doors Into Enterprise Agentic AI: Copilot Studio, Joule Studio, Azure AI Foundry, Databricks, and LangGraph

Every major enterprise vendor now sells a way to build AI agents, and they are not competing on the same axis. The real choice is not which product is best. It is how much control you are willing to trade for how much convenience.

agents enterprise langgraph copilot-studio

Inference & Serving 21 min

FlashAttention and the Memory Wall: Why Attention Was Never Compute-Bound

An A100 can execute 312 trillion half-precision operations per second but can only pull about 2 terabytes from memory in that same second. FlashAttention made attention fast not by computing less, but by refusing to touch memory.

flashattention gpu attention transformers ∑ ◫

Inference & Serving 21 min

Four Bits Per Weight: How Low-Precision Quantization Stopped Hurting LLMs

A 70B model in FP16 needs 140 GB of memory it spends most of its time waiting to read. Dropping each weight to four bits cuts that to 35 GB, and for years that cut also broke the model. Here is what changed.

quantization inference gptq awq ∑ ◫

Platforms & Practice 27 min

Four Ways an A/B Test Lies: Power, Peeking, Sample Ratio Mismatch and Interference

At Microsoft, roughly one third of tested ideas move the metric in the intended direction. That number is only meaningful if the tests are trustworthy, and the four failure modes that destroy trustworthiness all produce a clean, …

experimentation ab-testing causal-inference statistics ∑ ◫

Model Architecture 21 min

From 4K to a Million Tokens: How RoPE Scaling, YaRN, and Ring Attention Stretch the Context Window

A model trained on 4,096 tokens can be coaxed into reading a quarter-million without retraining from scratch. The trick is not bigger attention; it is lying to the model about position, and splitting the sequence across a ring of GPUs.

long-context rope yarn ring-attention ∑ ◫

Reasoning & Evaluation 23 min

From Features to Circuits: What Attribution Graphs Explain, and the Fraction They Do Not

Swap the Texas features for British Columbia and Claude answers Victoria instead of Austin. That single intervention is the strongest evidence yet that a language model performs genuine multi-step reasoning inside one forward pas…

interpretability mech-interp safety alignment ∑ ◫

Platforms & Practice 9 min

From Macros to Agency: A Short History of Automating Knowledge Work

The dream of automating office work is older than the personal computer. Tracing the line from VisiCalc through RPA to LLM agents reveals a recurring lesson: the tools that record steps break, and the tools that pursue goals are …

history automation rpa agents

Model Architecture 42 min

From Nano Banana to Vision Banana: How Google DeepMind Turned an Image Generator into a Generalist Vision System

Training a model to generate photorealistic images teaches it geometry, semantics, depth, and object relationships. Google DeepMind's Vision Banana proves that a lightweight instruction-tuning pass over an image generator can bea…

nano-banana vision-banana image-generation computer-vision ∑ ◫

Model Architecture 24 min

From Normalising Flows to Flow Matching: How the Change-of-Variables Idea Survived Its Own Architecture

For five years, normalising flows bent every layer around one number: the log-determinant of a Jacobian. Flow matching kept the invertible transport map, stopped computing that number during training, and ended up inside Stable D…

variational-and-flow-models flow-matching diffusion-models generative-ai ∑ ◫

Inference & Serving 24 min

Goodput, Not Throughput: The Metric That Decides Whether Your LLM Deployment Works

Two servers run the same model on the same GPUs. One reports 4,200 tokens per second and is unusable; the other reports 2,600 and feels instant. Throughput is a property of the server, latency is a property of the request, and th…

inference serving latency goodput ∑ ◫

Reasoning & Evaluation 23 min

Guarantees Without Calibration: Conformal Prediction and the Limits of LLM Confidence

A language model's stated confidence is a number, not a probability. Conformal prediction offers the opposite trade: it promises nothing about any single answer and something exact about the long run, from any scorer, with one as…

uncertainty evaluation calibration safety ∑ ◫

Model Architecture 20 min

Half Mamba, Half Attention: Why Hybrid State-Space Models Took Over

Pure Mamba was supposed to replace attention. Instead the most efficient open models in 2025 are roughly seven-eighths Mamba and one-eighth attention. The reason is the KV cache, and what each layer can and cannot remember.

state-space-models mamba transformers long-context ∑ ◫

Browse by topic