Date: 2026-04-04
In a new Lenny Rachitsky's podcast episode titled "An AI state of the union: We've passed the inflection point, dark factories are coming, and automation timelines".
Highlights: https://simonwillison.net/2026/Apr/2/lennys-podcast/
Transcript: https://www.lennysnewsletter.com/p/an-ai-state-of-the-union
Simon's entire worldview rests on a paradox: AI has made the act of creation nearly free, which makes the acts of judgment, verification, and lived experience infinitely more valuable. The people who thrive won't be the fastest builders -- they'll be the ones who know what's worth building, can verify it works, and have the discipline to resist the intoxication of speed. The biggest risk isn't that AI replaces engineers; it's that engineers stop doing the hard, slow, human parts of engineering because the fast parts feel so good.
- Using AI Well Is Exhausting, Not Easy
Running four agents in parallel wipes him out by 11 AM. This directly contradicts the popular narrative that AI makes work effortless. The real skill is cognitive orchestration -- managing attention, verifying outputs, maintaining architectural coherence -- and it demands more experience, not less.
- Speed Has Broken Trust -- Including Self-Trust
Simon built software faster than he could use it, and admits he doesn't trust his own AI-built tools because he hasn't lived with them. His estimation skills are "broken." This reveals a profound tension: velocity without inhabitation produces hollow software. The thing users care about most -- that the maker has used their own tool for months -- is exactly what AI-speed development undermines.
- The "Middle Career" Squeeze Is the Workforce Story Nobody's Telling
Per ThoughtWorks' finding: AI is great for seniors (amplifies deep expertise) and great for juniors (flattens the onboarding curve). But mid-career engineers are in the most danger -- not senior enough to direct agents masterfully, not junior enough to benefit from the leveling effect. This maps to a broader K-shaped disruption across knowledge work.
- Prototyping as a Superpower Is Now Commoditized
Simon candidly admits his career-long competitive advantage -- rapid prototyping -- is gone. Anyone can vibe-code a convincing UI. The new differentiator isn't building prototypes, it's knowing which prototype to bet on (product judgment, usability testing, taste).
- 97% Effectiveness Is a Failing Grade
This connects to the "normalization of deviance" concept -- people gradually accept declining standards because AI output looks plausible. A 3% error rate in agentic code, compounded across thousands of decisions, is catastrophic. The danger isn't spectacular failure; it's the slow erosion of quality that nobody notices until it's systemic.
- The "Lethal Trifecta" Is the Hidden Danger No One Talks About Enough
Simon identifies three converging threats -- prompt injection, over-trust in AI outputs, and the security nightmare of tools like OpenClaw -- that together create catastrophic risk. The key nuance: people want AI assistants so badly they will ignore security concerns entirely. OpenClaw went from first line of code to a Super Bowl ad in ~2 months, despite being a security minefield.
- November 2025 Was the Real Inflection Point — Not the ChatGPT Launch
The subtle insight here is that the chatbot era (2022-2024) was a warmup act. The real revolution started when AI agents could run code, test it, iterate autonomously, and operate in loops. Software engineers are "bellwethers" -- what's happening to them now will ripple outward into law, journalism, product management, and all knowledge work.
- Journalists May Be Better Positioned Than Engineers
A wonderfully counterintuitive point: journalists are trained to work with unreliable sources and triangulate truth. That's exactly the skill you need with AI. Professionals who already distrust their inputs have an unexpected advantage.
- The Bottleneck Has Shifted from Writing Code to Verifying Code
This is perhaps the deepest structural insight. Building is now cheap; knowing whether what you built is correct is the expensive part. Simon advocates a Red/Green TDD pattern -- write tests first, let the agent write code to pass them. The act of testing IS the engineering now.