The Mechanism — What a Transformer Actually Computes
Part 1 of a five-part series on how LLMs work. The forward pass step by step: attention, the residual stream, and where the parameters live.
43 min read · Aug 28, 2026
Courses
Interactive courses I've put together with AI tools, mostly on AI and engineering.
How models work
The Training Stack — What It Costs to Make a Model
Part 2. Scaling laws worked out as unit economics: training compute, memory per parameter, parallelism, and why Chinchilla's headline fit didn't replicate.
40 min read · Aug 28, 2026
Post-Training — How a Document Completer Becomes an Assistant
Part 3. SFT, RLHF and DPO, with DPO's closed form derived step by step, and what that shortcut costs.
36 min read · Aug 28, 2026
Evals — Error Analysis First, and Judges You Have Validated
Part 4. Read traces and label the failures before picking a metric, then check an LLM judge like any other classifier before trusting its pass rate.
40 min read · Aug 28, 2026
Inference Economics — Why Decode Is Bandwidth-Bound and What It Costs
Part 5. Why decoding is limited by memory bandwidth, why the KV cache caps how many users a server can hold, and what that does to serving costs.
37 min read · Aug 28, 2026
Building with AI
The New Rules of Context Engineering for Claude 5 Models
Based on Thariq's July 2026 article on how prompting changes for Claude 5 models, with applied modules on CLAUDE.md, skills and tool design.
71 min read · Jul 26, 2026
The Context Budget — Tokens, Skills and Memory as an Enterprise Decision
Skills, CLAUDE.md, memory, retrieval, MCP and subagents compared as different ways to decide what the model sees, and what each one costs per turn.
60 min read · Updated Sep 15, 2026
Claude Mods — Reading a Feature Before It Ships
How Claude Code's extension model is moving from hooks that vote on an action to mods that wrap it, read from a public issue and published source before release.
26 min read · Updated Sep 15, 2026
What to Do About Slop — The Cherny Exchange, Examined
A close reading of the email exchange Boris Cherny published about AI-written code in an old codebase, and what each side actually argued.
30 min read · Updated Sep 15, 2026
The Attention Factory — Reading OpenAI's Agentic Pipeline
OpenAI's nine-stage agent pipeline, read as a system for deciding what a person should look at rather than a system for writing code.
47 min read · Sep 15, 2026
The AI-Native SDLC, Audited
Anthropic's playbook for running each stage of software delivery with agents, checked against METR, DORA 2025 and other published evidence.
40 min read · Aug 29, 2026
AI progress
Price and Pace — Measuring AI Progress from the Outside and the Inside
Epoch AI's estimate of how fast the cost of AI is falling, next to Anthropic's proposal for measuring progress from inside a lab.
20 min read · Sep 26, 2026
We Must Pace the Frontier — Amodei's Argument, Examined
Dario Amodei's September 2026 essay laid out as a chain of claims, then tested against its sources and the main objections.
67 min read · Updated Sep 15, 2026
Jev — AI That Doesn't Talk
TypeSafe AI's Jev returns typed answers with probabilities instead of chat. What its calibration does and doesn't guarantee, and the economic bet behind it.
19 min read · Sep 27, 2026
Suncatcher — The Engineering Case for AI Compute in Orbit
The engineering in Google's paper on putting TPUs in orbit: power, laser links between satellites, and radiation, with measured results kept apart from projections.
14 min read · Sep 26, 2026