Writing

Newest first, by the date each post was written.

20 posts 7 months
Latest Business Oct 4, 2026

Tech Job Titles: Same Job, Different Price

We read 3,060 job postings from 2026 for ten tech titles in the US and Canada. Titles that ask for almost the same skills can still sit 35% to 60% apart in posted base salary.

Ten tech titles, from Software Engineer and Forward Deployed Engineer to Business Analyst, measured two ways: the skills their 2026 job postings ask for, and the base salary those postings disclose, set against base and total compensation reported on levels.fyi. Analytics Engineer and Data Analyst, or ML and MLOps Engineer, share most of their skills yet differ by 35% to 60% in posted base salary. Forward Deployed Engineer turns out to be its own job, closest to AI Engineer. Includes a pay explorer that compares posted base with levels.fyi base and TC by country, level and title.

Read the post

October

2026 · 1 post

  1. Tech Job Titles: Same Job, Different Price

    We read 3,060 job postings from 2026 for ten tech titles in the US and Canada. Titles that ask for almost the same skills can still sit 35% to 60% apart in posted base salary.

    Business · Oct 4

September

2026 · 3 posts

  1. Frontier Frugality

    The best model costs 10x the cheap one, and agentic workflows re-bill your entire context every turn. How to spend frontier tokens only where the intelligence actually lives.

    Agents· Agent practice · Sep 29

  2. Jev: A System One Model

    TypeSafe's Jev answers with typed decisions and probabilities instead of text. What it is, where it sits next to an LLM, and what the first outside tests found.

    AI & ML· Frontier models · Sep 27

  3. Every Health Claim, Tied to the Marker It Moves

    225 papers, 316 findings, 14 health goals. Every finding names the paper behind it, the people it was measured in, and the condition it holds under.

    Business· Health evidence · Sep 13

July

2026 · 3 posts

  1. How Reasoning Models Learn to Think

    RL with verifiable rewards, GRPO, and the second scaling axis — from OpenAI o1 to DeepSeek-R1.

    AI & ML· Frontier models · Jul 4

  2. The Trimodal Nature of Tech Compensation in 2026

    Three tiers still shape tech pay — and in 2026, a fourth spike of frontier AI labs has appeared above them.

    Business· AI economy · Jul 3

  3. The AI Investment War

    An interactive map of who owns whom, who pays whom, and why a trillion dollars keeps going in circles.

    Business· AI economy · Jul 1

June

2026 · 1 post

  1. GLM-5.2: The Open-Weight Model That Beats GPT-5.5

    How Zhipu AI's IndexShare mechanism makes 1M-token inference practical — and why the open-source community is paying attention.

    AI & ML· Frontier models · Jun 25

April

2026 · 4 posts

  1. DeepSeek V4

    How hybrid attention, million-token context, and agent post-training push an open model close to the frontier.

    AI & ML· Frontier models · Apr 24

  2. Inside the Claude Code Leak

    What 512,000 lines of accidentally published source code revealed about AI agent architecture.

    Agents· Claude Code · Apr 1

  3. Manifold Constrained Hyperconnections

    How DeepSeek stabilized the next generation of residual connections by constraining weight matrices to live on the Birkhoff polytope.

    AI & ML· Frontier models · Apr 1

  4. LLM-Compiled Knowledge Bases

    How to use LLMs as a personal research compiler: ingesting raw sources, building a living wiki, and querying it with no RAG required.

    Agents· Agent practice · Apr

March

2026 · 2 posts

  1. AI Agent-First Engineering

    From agent non-determinism to production harness systems — a synthesized roadmap.

    Agents· Agent practice · Mar 7

  2. PageRank: How Google Brought Order to the Web

    The recursive, link-based algorithm that turned hyperlinks into votes — and built the world's most powerful search engine.

    AI & ML· Search & ranking · Mar 7 Paper · 1998

February

2026 · 6 posts

  1. 98 Years of the S&P 500: How Major Events Shaped Market History

    Every crash felt like the end. Every recovery seemed impossible. The market kept climbing.

    Business· Markets · Feb 7

  2. FlashAttention

    How IO-awareness and tiling make exact attention fast and memory-efficient.

    AI & ML· Foundations · Feb 4 Paper · 2022

  3. Vision Transformers Need Registers

    Why large ViTs light up random specks of background, what those specks are secretly doing, and how four extra tokens make them go away.

    AI & ML· Foundations · Feb 4 Paper · 2023

  4. Understanding BERT

    Pre-training deep bidirectional representations for language understanding. How masked language modeling enables a single model to master almost any NLP task.

    AI & ML· Foundations · Feb 3 Paper · 2018

  5. Mastering Claude Code

    Workflow patterns and configuration practices from the team that built it.

    Agents· Claude Code · Feb 3

  6. Understanding the Transformer

    Understanding the building blocks and design choices of the Transformer architecture.

    AI & ML· Foundations · Feb 3 Paper · 2017