Who to Follow

The AI space has a noise problem. Every week there’s a new thread declaring AGI is six months away, a new framework that will change everything, a new guru selling a $997 prompt engineering course. Most of it is garbage.

This chapter is about the signal. People who have built things, studied things, and earned the right to have opinions through decades of actual work. Not marketers. Engineers, researchers, and analysts who tell you what works, what doesn’t, and what they don’t know.

I didn’t compile this from “top AI influencer” listicles. These are the people whose work changed how I think or how I build. Some I’ve followed for years. A few I disagree with on fundamental questions. They’re all here because they’ve earned it through work, not followers.

The software architecture legends

Grady Booch — co-creator of UML, IBM Fellow, five decades in software architecture. One of the most prominent voices arguing LLMs alone won’t produce AGI. His threads are pithy, precise, and unimpressed by scale-is-all-you-need arguments.

Martin Fowler — Chief Scientist at Thoughtworks, author of Refactoring. Calls AI “the biggest shift in programming” but approaches it methodically. His “Exploring Generative AI” series documents what actually works — the kind of boring, correct advice that saves you from shipping broken products.

Kent Beck — creator of Extreme Programming and TDD, co-author of the Agile Manifesto. Remarkably even-handed. His “Genie Sessions” series shows real work with coding agents and names the failure modes honestly.

Robert C. Martin (Uncle Bob) — author of Clean Code. His argument: AI-generated code makes clean code principles more important, not less, because syntactic correctness is not maintainability. He’s right about that.

The AI researchers who tell you the truth

Andrej Karpathy — founding member at OpenAI, formerly Director of AI at Tesla. The best AI educator alive. He coined “vibe coding.” His “Deep Dive into LLMs” talk is what finally made me understand why Claude sometimes hallucinates confidently. Once you see how token prediction works mechanically, the behavior stops being mysterious.

François Chollet — creator of Keras, co-founder of the ARC Prize. The most rigorous public critic of “scale is all you need.” When a benchmark creator tells you benchmarks are being gamed, pay attention.

Yann LeCun — Turing Award laureate. Argues LLMs are “a dead end” because they lack persistent memory, planning, and a model of the physical world. When a Turing Award winner says the current approach is wrong, you should understand his argument.

Simon Willison — co-creator of Django, creator of Datasette. The most practically useful voice in the entire AI space for working developers. His blog is the most-cited running reference on LLM behavior. I read it more than any other source on this list. It’s how I stay current on what changed since yesterday.

Sebastian Raschka — independent LLM researcher, author of Build a Large Language Model (From Scratch). Pure technical content. Almost zero hype. If you want to understand the technical differences between DeepSeek, Qwen, and Claude, start here.

Nathan Lambert — post-training research lead at the Allen Institute for AI. One of the few public writers who actually trains frontier models. His Interconnects newsletter is read by frontier-lab researchers. Tagline: “minus the hype.”

Jeremy Howard — co-founder of Answer.AI, founder of fast.ai. Transparent about his own ambivalence: “I used to enjoy programming. Now my days are typically spent going back and forth with an LLM and pretty often yelling at it.” That honesty is rare and valuable.

The practitioners who build and share honestly

Boris Cherny — creator of Claude Code at Anthropic. His practical tips (5 parallel instances, 20-30 PRs per day) are unusually direct about what works versus what sounds good in a demo.

Mitchell Hashimoto — founder of HashiCorp. His “AI Adoption Journey” is one of the best first-person accounts of a senior engineer working through skepticism. His journey mirrors mine in some ways — we both came around through hands-on use rather than theory, and both refuse to call it magic.

Thorsten Ball — engineer at Sourcegraph. His Register Spill newsletter is one of the most honest accounts of an experienced engineer’s journey with AI tools.

Kelsey Hightower — retired from Google. The “call out the emperor’s clothes” voice. “Technologists should feel dirty for falling for AI hype.” If your timeline gets too breathless about agent swarms, read Hightower for recalibration.

Hamel Husain — independent AI consultant, former GitHub ML engineer. The “LLM bullshit knife.” His core argument: most LLM products fail because teams skip systematic error analysis. If you’re building an LLM product and haven’t read his evals framework, you’re probably shipping without knowing what’s broken.

Dan Shipper — CEO of Every. Runs a 15-person company where engineers write virtually zero code by hand. If you’re a non-technical founder trying to understand what “agent-native” means in practice, start here.

Infrastructure and industry analysts

Dylan Patel — founder of SemiAnalysis. His analysis of DeepSeek’s true training costs ($1.6 billion, not $5.6 million) set the agenda for Wall Street and the labs. Caveat: SemiAnalysis sells consulting to the firms it covers.

Ben Thompson — founder of Stratechery. Business strategy first, technical detail second. If you want to understand market dynamics rather than model weights, Thompson is the reference.

Gergely Orosz — author of The Pragmatic Engineer. Conducts the most consistently high-quality long-form interviews with practitioners — useful as a primary-source archive of how senior engineers think about AI when they’re being honest.

How to use this list

Don’t follow everyone. Pick based on what you need. Understanding how LLMs work: Karpathy, Raschka. Staying grounded when hype gets loud: LeCun, Chollet, Hightower. Engineering practice: Fowler, Beck, Cherny, Husain. Founders and knowledge workers: Shipper, Thompson. Infrastructure economics: Patel, Thompson.

The honest experts disagree with each other, sometimes loudly. LeCun thinks LLMs are a dead end. Karpathy thinks they’re the foundation. That disagreement is the feature. The alternative is a single curated narrative — exactly what the hype machine offers. You don’t need consensus. You need calibrated uncertainty from people who know what they’re talking about.

My routine: Willison’s blog daily. Raschka and Lambert weekly. Karpathy and Thompson whenever they publish. About 30 minutes a day over breakfast. Enough to stay current without drowning.

Follow the signal. Ignore the noise. Build things.

This is the free web edition of Chapter 22. The full text — with complete follow lists, platform-specific links, podcast recommendations, and curated reading paths by role — is available in 42: The AI Builder’s Stack, coming Q3 2026 on Amazon in hardcover, paperback, and digital.