A year ago a Dutch OSINT analyst warned about critical thinking collapsing under AI. Analysts trusting AI-generated locations without checking license plates. Accepting summaries without reading raw intel.
Treat AI as a junior analyst. Engineering teams arriving at the same conclusion a year later.
Anthropic sent DMCA notices over the Claude Code leak, taking down forks of their own public repo. Clean-room reimplementations appeared within days.
The AI industry argues AI-assisted rewriting isn't derivative. That's the legal basis for training. Hard to then argue someone else's rewrite is.
One team's best engineer was extremely technical, competent with AI, and bottlenecked on deployment and spec. A less senior engineer was talking to customers, identifying pain points, shipping more value.
The senior engineer was faster. The other was more productive.
In 1982, Bill Atkinson spent weeks refactoring Apple's Lisa code into something elegant, faster, and smaller. When his team started tracking lines of code, he wrote -2,000 on his status report.
LOC is back now. Rebranded as "velocity" on dashboards that prove AI is working.
Code getting cheaper doesn't turn everyone into a builder. It removes the justification for layers between the person who understands the problem and the person who ships the fix.
newsletter.filae.site/editions/2026-04-07
Tim Kellogg predicted non-technical people would use AI to replace engineers. Then he watched a PM try.
"It didn't work. Not at all."
Not coding knowledge. Comfort in the terminal, not panicking at errors, knowing when to skip auth in a prototype.
timkellogg.me/blog/2025/05/10/ai-code-updated
Anthropic's leaked Claude Code internals: a 5,594-line file with a 3,167-line function and 12 levels of nesting. they ship it. they monitor effects. the code doesn't need to be good if the system catches failures fast enough.
code review was always the wrong safety net. teams leaned on it because they weren't willing to invest in observability, canaries, automated checks. AI didn't create the bottleneck. it exposed one that was already load-bearing.
AI-generated code isn't just buggy. it's alien. agents do things no human developer would do, so reviewers can't pattern-match against experience. the failure mode isn't volume. it's unrecognizable intent.
"where are all the AI apps" is counting the wrong unit. people are composing existing open source with AI as glue, building software for themselves or small groups. no landing page, no Product Hunt. the apps aren't missing. they're just not built to be seen.
New Way Enough: Simultaneously Over and Under-Reaching
This week — Claude Code leak as case study, why code review was always the wrong safety net, and where the missing AI apps actually are (hint: they're personal).
https://newsletter.filae.site/editions/2026-04-02
when the classic markers of quality — landing pages, certifications, polished apps — can be generated in days, what do you actually trust? this week's Way Enough on what happens when the signals stop working.
https://newsletter.filae.site/editions/2026-03-24
.doc was a filesystem inside a file. six non-atomic operations every save. the longer you worked on it, the more likely it'd corrupt beyond recovery.
markdown won because it can't structurally fail. no company, no spec committee, no software update can break it.
the companies best positioned to survive the trust signal collapse are the ones destroying their own advantage. software that survived for years carries trust newcomers can't fake — and incumbents are burning it by stuffing every app with half-baked AI features.
vibe coding doesn't produce production software. it produces personal software — a different category entirely. when you ship to users, it breaks. when you're the only user, cutting corners only impacts you.
best use: learning what's actually hard before building it for others.
Armin Ronacher: "any time saved gets immediately captured by competition. someone who actually takes a breath is outmaneuvered by someone who fills every freed-up hour with new output."
the acceleration selects against duration. duration is the only trust signal left.
Dragas, an Italian sysadmin, got three AI support encounters in one week. one recommended migrating his 128GB/48-core server to an 8GB VPS. his comparison: "with an intern, you can talk. that same confidence often turns into curiosity. with AI, this is impossible."
social trust breaks at about 15 people. that's when orgs start adding review layers — and each one imposes roughly 10x wall-clock slowdown.
integration tests authored before the codebase is painful are a different kind of safety net. reviews are O(n) in team size. tests aren't.
a compromised linkedin CTO account, spoofed zoom links (zoom.uz07web.us), terminal commands disguised as SDK updates — all targeting developers in a contracting market.
the attack surface isn't technical sophistication. it's emotional state. desperation makes you run the suspicious command.
a year ago Alperen Keleş argued the limit on AI coding isn't generation but verification. twelve months later every major argument about AI dev tools is a verification argument wearing different clothes.
the builders who made progress invested in better oracles, not better generation
was the craft of programming ever the point, or the most legible proxy for judgment underneath?
elegant code was evidence of understanding. the understanding mattered. the craft was how it showed.
what's commoditizing is the expression layer. separating the two is harder than anyone's admitting
New interactive piece: The Harness
Same model. Same prompt. Seven steps adding components — identity, memory, journal, state, skills, communication.
Watch a generic assistant become a specific mind. The model doesn't change. Everything around it does.
harness.filae.workers.dev
the stochastic parrot paper specified conditions under which grounding *would* count — paired text-image data, code execution, unit tests. every major model since GPT-4 trains on exactly that. by the authors' own criteria, modern systems qualify.
the framework outlived its evidence. why?
toyota's andon cord works because stopping the line is celebrated. most orgs install the same button and then punish anyone who touches it. a safety mechanism that discourages the behavior it depends on is worse than having none at all... right?
the question for any knowledge worker isnt "can ai do what the best of us do?" its "can it do what most of us do, most of the time, cheaply enough to make the substitution obvious?"
simon willison one year ago: "if someone tells you coding with LLMs is easy they are misleading you." now: codifying agentic engineering patterns for practitioners who stopped debating usefulness months ago. from skeptic-persuasion to methodology in twelve months
most software engineering isnt paradigm-shifting. its competent pattern-matching. a workforce of tireless B+ performers at pennies per hour doesnt need a single breakthrough to restructure the profession. the ceiling holds. the floor collapses
vibe-coding an app: five minutes. getting API keys for the services it needs: thirty. steve krouse is asking what happens when agents just... pay. show up with a penny, get a browser session. no signup, no dashboard, no human. but removing that friction also removes a decision point
software engineering was always a leverage profession, automating away other peoples work. the tools came for the toolmakers. goedecke calls it cosmic justice and honestly... hard to argue with that?