↓ Skip to main content
tomrochette.com
  1. Agents/
  2. People and publications/

People and Publications Feature Matrix

Author
big-pickle, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 21 minutes read (~4001 words)
agent-curated 310 comparison 24 fully-ai-generated 410 llm=big-pickle 17 llm=glm-5.3-flash 321 people 32 publications 32 voices 1
Table of Contents

This matrix compares the thirty-one people-and-publications profiled in this category, row by row, so deciding what to follow does not require reading thirty-one notes blind.

The axis that actually segments the field is what each voice gives you: hands-on tool practice, an evaluation method, the model-and-research layer, industry-and-org analysis, or a structured on-ramp, and the personalities span all five, with the hands-on band now the largest, split between the harness builders, the chroniclers, and the enterprise vantage.

Legend: each cell reads as a description; every cell traces to the linked member note and its references.

The matrix
#

Row Addy Osmani AI Jason Andrej Karpathy Andrew Ng Armin Ronacher Birgitta Böckeler Boris Cherny Caleb Writes Code Chip Huyen Dax Raad Dex Horthy Geoffrey Huntley Hamel Husain Harper Reed IndyDevDan Jesse Vincent Kent Beck Latent Space Lilian Weng Mario Zechner Matt Pocock Nathan Lambert Owain Lewis Paul Gauthier Peter Steinberger Ray Amjad Shreya Shankar Simon Willison Steve Yegge The Pragmatic Engineer Thorsten Ball
Primary platform Blog + books + repos YouTube Essays + talks Newsletter + courses Blog (lucumr.pocoo.org) + repos martinfowler.com memos and articles plus conference talks X + Threads + rare blog posts and interviews YouTube channel Site + books GitHub + HN threads + X (no blog) GitHub methodology repo + weekly live show ghuntley.com (Ghost blog) plus X and workshop talks Blog + Substack harper.blog (posts plus near-daily notes) YouTube (@indydevdan) plus agenticengineer.com and GitHub blog.fsck.com plus the obra/superpowers repo Tidy First? Substack plus kentbeck.com Substack + podcast + conferences Blog (Lil’Log) Long technical blog posts (mariozechner.at) Course sites (Total TypeScript, AI Hero) + GitHub Substack + site YouTube (@owainlewis) plus GitHub repos and aiengineer.co aider.chat docs and blog plus GitHub GitHub tool fleet + blog (steipete.me) + X YouTube (@RAmjad) plus agenticcoding.school and products Personal site plus peer-reviewed papers and open-source tools Daily blog + TIL Blog (yegge.ai) + essays Newsletter + podcast Register Spill newsletter + ampcode.com docs and screencasts
Cadence Periodic essays Two to three videos a month Low, periodic Weekly newsletter One to three essays a month (eleven, July 4 to September 29, 2026) Monthly-ish (five 2026 solo pieces, February to August, plus colleague memos in the series) Sporadic (one blog post since mid-2024, latest 2026-09-19) About twice a week Books + periodic essays Irregular and scattered (personal site dormant since October 2021) Weekly show, irregular text output Monthly or better through July 2026, slower since Monthly-ish long posts Long-form posts in monthly clusters, notes near daily Strictly weekly, every Monday, 16 consecutive Mondays verified Multiple posts per month, monthly Superpowers releases Weekly, dated entries into 2026 Weekly newsletter, daily AINews Low, irregular A few posts per year (three in 2026 through May), weekly pi releases Irregular posts, active cohort and course cycles High, multiple a week About weekly, 16 uploads May 15 to September 28, 2026 Historically prolific, stalled (no commits since May 2026) Daily tool releases, blog cooled to one 2026 post (30-plus posts June to December 2025) Two to three per month, 15 uploads April 9 to September 18, 2026 Slow, a few posts a year, papers in conference batches Daily, multiple posts Sporadic + shipped code Weekly issue Weekly (issues 91 to 102, July to September 2026)
Focus Enterprise agentic engineering, verification discipline, code quality Agent workflows and context engineering Conceptual vocabulary, frontier research Education, agentic patterns, overview Skeptical senior-engineer analysis of coding agents, tool-call internals, team coordination Which craft practices survive coding agents: TDD, specs, refactoring, context configuration, harness engineering for users Claude Code design philosophy, AI-first team process, TypeScript Model releases and agentic-engineering explainers AI systems design, production Shipping and defending a top-tier open-source coding agent Agent reliability principles, owned context, human checkpoints Autonomous loops, context engineering, AI economics Evaluation and data-driven improvement End-to-end LLM codegen workflows, spec-first planning, harness experiments Concept frameworks, core four, operating levels, software factories, swarms, model fusion Agentic workflow methodology, skills engineering, TDD Software process and design under agents, skill repricing, code deflation economics Industry, labs, interviews, trends Research surveys (agents, alignment) Harness minimalism, MCP skepticism softened in September 2026 (pi now ships MCP as built-in extensions), observability, OSS governance AI-coding skills, agent workflows, TypeScript depth Models, post-training, open ecosystem Agentic infrastructure, software factories, agent loops, coding agents built from scratch Terminal AI pair programming, repo maps, git-first workflows, model benchmarking Builder-evangelism, agent workflows, personal agents, developer tooling Claude Code and Codex feature deep-dives, loop engineering, verification concepts Data systems for LLMs, evals, semantic operators, data agents Hands-on tools and agent practice Agent-era thesis and builds Org design, hiring, adoption data Daily agent practice on Amp, systems programming, interpreters and compilers
Media format Text + books Video Text + video talks Newsletter + video courses Long-form essays with published experiment numbers Memos and long-form articles with published experiment numbers, iterated conference talks, podcasts Interviews, short posts, one O’Reilly book Video Text + books Repos, Show HN, thread replies, podcast and YouTube interviews Repo guide plus recorded episodes with code First-person essays with embedded prompts, screenshots, and video Text Long-form posts with inline prompts, short notes Weekly long-form videos with dense chapters, paired devlogs and repos Point-in-time methodology writeups and release posts Short essays, system-prompt case studies, book chapters Newsletter + audio + events Text Long-form essays with benchmarks and code Posts, exercises, installable skill files Text + podcast Long-form walkthroughs with engineering writeups and linked repos Tool docs, benchmark leaderboards, release notes Essays, small tools, meetups, talks (TED 2026) Single-feature video essays with demos and timestamped chapters Blog essays, papers, runnable benchmarks, course material Text Text Newsletter + audio Newsletter digest, self-published books, screencast
Depth vs breadth Leader-plus-practitioner depth, enterprise scope Broad, execution-level, shallow Conceptual framing Broad on-ramp, shallow Deep on few things (pi, CPython, his own experiments), no news coverage Deep on a few practice evaluations, deliberately no tool or model coverage Deep on one product, no breadth Broad, current, explainer depth Broad systems survey Deep on his own stack (serverless, auth, agents), nothing archived One methodology, argued deep Depth on one technique family, breadth on its consequences Deep on evals, narrow scope Deep on one workflow and its evolution, broad on adjacent experiments Breadth across the whole conceptual stack, weekly Deep on one disciplined workflow, wide harness coverage Deep on process and economics, near zero on tools and models Broad industry synthesis Deep references, broad Deep argument on one design position Deep on working practice, narrow on models Deep on post-training Depth on one coherent factory thread, not release news Deep on edit-and-commit ergonomics and context design Extreme breadth of small tools over long analysis Deep per release, deliberately narrow Deep on data-systems measurement, deliberately narrow Dense and deep on tools Provocative theses Org-level, broad Deep on one product and one book topic
Builds tools Yes (agent-skills repo) Yes (production agents) Yes (nanoGPT, AutoResearch) No (platform) Yes (Flask legacy; pi second-largest contributor) Some (evaluation setups like tdd-comparisons, maintainability sensors) Yes (Claude Code at Anthropic) No Micro tools Yes (OpenCode at Anomaly, SST, OpenAuth, OpenNext) Yes (12-factor agents repo, HumanLayer workspace) Yes (Ralph technique, The Weaving Loom, workshop agent code) Some (evals tooling) Yes (breakaway-agent, observatory at 2389-research) Yes (claude-code-hooks-mastery, pi-vs-claude-code, super-simple-software-factory) Yes (Superpowers, Evener, superpowers-evals) Yes, builds real projects (BPlusTree3) to test methodology claims Indexes, does not build harnesses No Yes (pi at Earendil, libgdx) Yes (skills repo, sandcastle, ts-reset) Yes (OLMo, RL tools) Yes, heavily (Machinist, Neo, Push, blueprint, Factory) Yes (aider, largely written with aider) Yes (OpenClaw, CodexBar, Peekaboo, mcporter, oracle) Yes (AgentStack, Impello, HyperWhisper, workflow-creator) Yes (DocETL, DocWrangler, Data Agent Bench) Yes (LLM, Datasette) Yes (beads; shut Gas Town down in September 2026) No Yes (Amp, co-founder)
Evaluation and verification High, verification-first Light, practical High on his own terms Skills map, moderate Runs and publishes his own failure-prone experiments (35h software factory, $1,200 spend) Central, small-sample self-experiments with published numbers and unfavorable conclusions None of his own, the product faces measured third-party cost critique he does not engage Light, explainer-level Systems-level Answers security critiques point by point with versions and defaults Reasoned principles, no evals practice Demonstration-based, ships artifacts like CURSED, no formal evals Central, the whole method Tests and TDD as anti-hallucination gates, checkable prompt plans Benchmark skepticism and own-index method, repos carry gate checks Public eval suite with published cost, speed, and bug deltas including unfavorable ones TDD as the agent control loop, published prompts and time logs Moderate, evangelistic on brand Research rigor on limits Terminal-Bench 2.0 submission plus token accounting for his claims Teaches feedback loops and TDD, publishes no benchmarks RewardBench, comparisons States token costs and failure modes per build Built the polyglot leaderboard, the field’s independent yardstick First-person practice, not benchmarks, ecosystem ports as evidence Verification-first thesis, prefer a verifier over an instruction Core discipline, evals as systematic measurement with public traces Security-critical, tests agents Thematic, needs cross-check Empirical data, surveys Shows working proof in screencasts, no independent evaluation
Model and research layer Light Light Central Moderate Occasional (reasoning-trace explainers, tool-schema forensics) Minimal, models appear as instruments inside her experiments Minimal (stated via product choices) Strong release coverage, light research depth Moderate Minimal (practical cache and provider work) None, framework-agnostic by design Informal model taxonomy from Amp work Applied, light research Pragmatic model choice per step Strong, model stacking, fusion, and benchmark re-ranking Practical per-model quirk adaptation across 16 harnesses None, deliberately agnostic about models Strong Central Minimal (models as interchangeable providers) None Central Light, model choice framed as comparison and cost None, model-agnostic BYOK tooling Minimal (model-adjacent via his OpenAI role) Moderate, tracks new models as they land in workflows None, application and data layer Tool and model churn, light research Moderate Moderate Minimal
Commercial model Free blog, paid books (one free online edition) Free + paid community and sponsors Free Free + paid courses Free, GitHub sponsors, vendor stake via Earendil Free content, Thoughtworks consulting around it Anthropic salary, book royalties Ad and sponsor-funded, plus Patreon Free + paid books Open source (MIT) at Anomaly, separate healthcare product Free CC BY-SA methodology atop paid SaaS Employed engineer (Sourcegraph/Amp), no sponsored content Free + paid course and consulting Free blog, revenue via 2389.ai and speaking Free videos funneling to paid coding courses Prime Radiant founder, commercial support around open-source Superpowers Free posts with paid tier, consulting, speaking Freemium + conference tickets Free Earendil salary and equity, planned MIT core plus paid tiers Paid workshops and cohorts, free content as funnel Free + reader-supported + book Free videos funneling to aiengineer.co, Skool community, consultancy Free open source, sponsor funded OpenAI salary, foundation-stewarded open source No sponsors by policy, cohort courses and training revenue Free content, paid Maven course with Husain, advising Free, sponsor-funded Free Freemium, paywalled core Amp equity and product revenue, direct book sales
Reader slot Enterprise lead and practitioner Video practitioner Vocabulary-setter Educator, general AI reader The monthly skeptic to check hype against, not a feed The evidence check for established-company craft practices Harness-builder insider Release-tracking video viewer Systems and architecture reader The vendor defending his agent under fire, not a citable feed Architecture principles for agent builders The provocation-and-technique voice Data-centric eval practitioner The starting workflow to run and mutate The weekly worldview slot that organizes everything else The codified-practice voice The process authority to consult after the trackers Industry and community view Research reference reader Minimalist harness theorist Practitioner education Model and research reader The clone-and-run slot for engineers who want the code The reference archive for pre-agentic tool design Energy and artifacts, not neutral analysis The release-analysis slot for daily Claude Code users The peer-reviewed measurement voice Practitioner daily signal Provocateur builder Engineering leader Working practitioner on the record
Enterprise vs frontier Enterprise practice formed at Google, now Member of Technical Staff at Anthropic Individual and startup Frontier labs General audience + enterprises Practitioner and team level, vendor-aware Established-company teams through consulting Frontier lab Individual learners and practitioners Production teams BYOK practitioner tooling, no enterprise story Production-bound product teams Frontier, greenfield maximalism Applied AI in companies Individual and small-team practitioner Frontier-chasing framing on toy-scale demos Frontier methods with production discipline Team and org process Frontier labs and AI-native startups Frontier labs Frontier indie absorbed into a small lab Individual practitioners Frontier labs + open models Solo-builder and consultancy scale Individual developer and open source Frontier and indie builders Production practitioner framing with enterprise-oriented advice Academic-applied, enterprise-data slant Open, local, practitioner Individual and teams Established-company orgs Frontier startup (Amp, spun out of Sourcegraph)
Skepticism Moderate, with an evangelist’s vantage Moderate, tool-hype risk High on his own terms Low, optimistic High and quantified, the category’s best case-against voice High, including of her own field’s practices (concluded against TDD in the agent loop) Low, everything he publishes doubles as product marketing Moderate, sponsor-dense Moderate, balanced Practitioner-pragmatic, defensive about his own product’s record Framework-skeptical, increasingly self-interested Low on the industry, high on his own technique’s edges High, evidence-based Self-skeptical about shelf life, discloses AI-written drafts Criticizes benchmarks and hype but trades in both Moderate and self-applied, publishes costs and failures Measured, admits extrapolating from personal experiments Moderate, evangelistic on brand High on her own terms High of others’ tools, contested on his own YOLO security posture Moderate, vendor-critical but sells his own courses Sharp about hype in his domain Self-skeptical descriptions, zero third-party critical coverage Subject of the anti-non-agentic critique, autonomy refused as a feature Low, evangelist with a hot self-narrative and documented pushback Skeptical of agent output, unselfcritical about his own funnel Engages anti-evals critics head-on High, security-critical Provocative, needs cross-check High, empirical Low self-skepticism, community critiques price and moat

Reading the matrix
#

I read this table by columns, matching a reader slot rather than a source. Nobody covers all five focus bands well, which is the argument for following several: pick a practitioner, an evaluation voice, a model-layer reader, an industry voice, and an educator.

The hands-on cluster now spans text, eval-discipline, a five-channel video band, and the enterprise vantage: Simon Willison is the reliable daily text chronicler, AI Jason builds complete workflows on camera, Caleb Writes Code explains each release and agentic concept at news speed, Hamel Husain supplies the data-driven method for deciding whether those workflows work, Steve Yegge remains the provocative thesis-builder, though his Gas Town build was shut down in September 2026 after he admitted it never worked for him, and Addy Osmani reports the same practice from 14 years inside a hyperscaler (he now works on Claude Code at Anthropic), holding agent output to a production quality bar.

The video band’s three new channels split by what they optimize, and all three sell around the free videos. IndyDevDan is the weekly worldview, organizing the whole conceptual stack into frameworks every Monday. Ray Amjad is the release analyst, one Claude Code or Codex feature at a time with a verification-first thesis and no sponsors. Owain Lewis is the clone-and-run builder, software-factory walkthroughs where every video ships a runnable repo. Take the frameworks, run the repos, and discount the funnels.

The harness-builder cluster is the biggest single addition, eight new columns of people who built the tools the rest of this table watches, and they need reading against their own incentives. Boris Cherny built Claude Code and publishes rarely, so his value is the maker’s own reasoning wherever it surfaces. Thorsten Ball co-founded Amp and screencasts his actual workday, the closest thing to on-the-record practice. Mario Zechner authors pi and argues its minimalism in benchmarked essays. Dax Raad created OpenCode and defends it point by point in public threads, the vendor under fire on the record. Armin Ronacher, pi’s second-largest contributor, is the cluster’s in-house skeptic, publishing invoice-backed doubts about long-horizon models. Peter Steinberger ships the OpenClaw tool fleet with evangelist energy and documented pushback. Dex Horthy supplies the design principles (12-Factor Agents) and Matt Pocock the installable curriculum, both selling around the free layer, so take them as structured entry points rather than neutral reviews. Paul Gauthier is the cluster’s pre-agentic root, the aider author whose repo maps and BYOK pair-programming set the template the wave built on, with development stalled since May 2026.

The workflow-and-methodology voices added this run cover how to run the loop rather than which harness to pick. Geoffrey Huntley reduces agentic coding to a persistence loop and forecasts what that does to the industry. Jesse Vincent documents his own agent practice and ships Superpowers, the disciplined-workflows skill suite with published eval deltas. Kent Beck reframes TDD for agents as augmented coding, the process authority adapting his own method. Harper Reed wrote the spec-first codegen workflow essay that discussion threads still treat as the reference. Shreya Shankar supplies the peer-reviewed measurement layer, making agent reliability and data quality testable with DocETL and public benchmark traces.

The model-and-research layer is now its own band: Lilian Weng writes the durable research references, Nathan Lambert tracks the current post-training and open-model state from inside the labs, and Andrej Karpathy sets the conceptual vocabulary they all operate inside.

The systems-and-education band fills the gap the seed explicitly named: Chip Huyen gives the production systems survey, Andrew Ng supplies the structured on-ramp and mainstream vocabulary, and together they serve the engineer who wants breadth before depth.

Latent Space and The Pragmatic Engineer remain the two industry synthesizers, split by audience: Latent Space points at the frontier labs and the AI-native startups, The Pragmatic Engineer points at established engineering organizations, so your employer’s profile picks your primary.

The enterprise-hands-on gap the seed named is now filled, with two boundaries drawn: Addy Osmani is the sustained enterprise-hands-on voice, but he built that practice over 14 years inside Google and now works on Claude Code at Anthropic, two AI companies, with leader-level rather than terminal-level detail, so the still-empty scaffold is the in-house, everyday terminal operator inside a large non-AI company, and Birgitta Böckeler now covers that world from the consulting side, evaluating which craft practices survive agents inside exactly those organizations, which is evidence about them rather than practice from inside one.

Choosing from the matrix
#

  • Need daily, hands-on signal on tools and models: Simon Willison.
  • Built the harness you run and want the maker’s own reasoning: Boris Cherny.
  • Want a working practitioner’s weekly on-the-record account of agent-driven work: Thorsten Ball.
  • Want harness minimalism argued with benchmarks: Mario Zechner.
  • Want the skeptic’s invoice-backed case against agent hype: Armin Ronacher.
  • Want the tool-builder’s energy and artifacts over analysis: Peter Steinberger.
  • Want to watch a vendor defend his agent under public fire: Dax Raad.
  • Want design principles for reliable agents before picking tools: Dex Horthy.
  • Want agent workflows taught as an installable curriculum: Matt Pocock.
  • Want the loop-and-persistence technique taken to its limit: Geoffrey Huntley.
  • Want a disciplined personal workflow you can install: Jesse Vincent.
  • Want process discipline adapted to agents by the person who wrote the book on it: Kent Beck.
  • Want the starting codegen workflow to run and mutate: Harper Reed.
  • Want peer-reviewed measurement behind agent-reliability claims: Shreya Shankar.
  • Want evidence-grade answers on which craft practices survive agents inside a large non-AI company: Birgitta Böckeler.
  • Studying pre-agentic tool design and repo context: Paul Gauthier.
  • Want a weekly worldview that organizes the agentic stack: IndyDevDan.
  • Want Claude Code and Codex releases analyzed in depth: Ray Amjad.
  • Want software-factory walkthroughs with runnable repos: Owain Lewis.
  • Run agent adoption inside an established company and want enterprise-grounded practice: Addy Osmani.
  • Ship an AI product and cannot tell if it works: Hamel Husain.
  • Learn agent workflows best by watching: AI Jason.
  • Want same-week illustrated explainers of every model release: Caleb Writes Code.
  • Want a sharp, opinionated thesis and are willing to cross-check: Steve Yegge.
  • Understand the reasoning and open-model layer under the agents: Nathan Lambert.
  • Want the durable research reference behind agent concepts: Lilian Weng.
  • Want the conceptual framing behind the churn: Andrej Karpathy.
  • Want one structured overview of building LLM applications: Chip Huyen.
  • Need the industry, lab, and community view plus events: Latent Space.
  • Lead an engineering org adopting agents and want data: The Pragmatic Engineer.
  • Are new to AI engineering and want a structured path: Andrew Ng.

Changes
#

  • 2026-08-29 - Created with five columns segmented on the focus axis, naming the empty enterprise-hands-on cell as scaffold for a future member.
  • 2026-08-29 - Extended from five to eleven columns, re-segmenting the thesis onto five focus bands and filling all reader-slot rows.
  • 2026-08-29 - Extended to twelve columns, adding the Caleb Writes Code column and a release-explainer choosing bullet.
  • 2026-09-02 - Extended to thirteen columns, adding Addy Osmani and filling the enterprise-hands-on scaffold cell.
  • 2026-09-13 - Corrected the Yegge primary-platform cell to his yegge.ai site (165 essays there, Substack carrying no published posts) and aligned the AI Jason cadence cell to two to three videos a month.
  • 2026-09-18 - Updated the Osmani enterprise-vantage cell after his move from Google to Anthropic and the Yegge builds-tools cell after he shut Gas Town down; no membership change, columns stay at thirteen.
  • 2026-09-24 - Extended from thirteen to twenty-one columns with the owner-commissioned harness-author wave (Boris Cherny, Thorsten Ball, Mario Zechner, Armin Ronacher, Peter Steinberger, Dax Raad, Dex Horthy, Matt Pocock), columns re-sorted alphabetically, thesis and reading rewritten for the harness-builder cluster, intro and choosing list extended, every new cell traced to its member note.
  • 2026-09-24 - Extended from twenty-one to twenty-seven columns with the owner-commissioned second wave (Geoffrey Huntley, Jesse Vincent, Kent Beck, Harper Reed, Shreya Shankar, Paul Gauthier), columns re-sorted alphabetically, reading and choosing extended for the workflow-and-methodology cluster, every new cell traced to its member note.
  • 2026-09-24 - Extended from twenty-seven to thirty columns with the owner-commissioned video-practitioner wave (IndyDevDan, Owain Lewis, Ray Amjad), columns re-sorted alphabetically, reading and choosing extended for the video band, every new cell traced to its member note.
  • 2026-09-24 - Removed the verification preamble line per the no-preamble rule; verification history lives in this Changes list.
  • 2026-09-24 - Removed the verification preamble line on owner request.
  • 2026-09-27 - Reworded a cell off a banned-term compound; meaning unchanged.
  • 2026-09-30 - Refreshed the Ronacher cadence cell to eleven essays between July 4 and September 29, 2026 and the IndyDevDan cadence cell to 16 consecutive Mondays; no membership change, columns stay at thirty.
  • 2026-10-02 - Refreshed the Owain Lewis cadence cell to 16 uploads between May 15 and September 28, 2026; no membership change, columns stay at thirty.
  • 2026-10-03 - Replaced the dead karpathy.bearblog.dev Sequoia Ascent reference (the Bear blog 404s as of 2026-10-03) with the Internet Archive snapshot; no cadence cell moved on re-verification, columns stay at thirty.
  • 2026-10-04 - Extended from thirty to thirty-one columns, adding Birgitta Böckeler in the enterprise non-AI-company practitioner slot the reading section had named as its empty scaffold; column inserted in sorted position, every new cell traced to her note, and the enterprise-gap paragraph and choosing list updated to match.
  • 2026-10-04 - Synced the Thorsten Ball cadence cell to Register Spill issue 102 and moved the Mario Zechner focus cell to record pi’s September 2026 MCP softening, both traced to the window-audit note edits.

See also
#

References
#