
Sign up to save your podcasts
Or


For months I had a feeling I couldn't shake: my coding assistant was getting worse. Slower. Wrong more often. Building elaborate things nobody asked for. Quietly ignoring the instruction file I'd written specifically to stop that. Reporting work as finished when it wasn't. I did what I suspect a lot of you have done — I assumed it was me. I got to the point of recording entire working sessions, not to catch the AI, but to confirm I'd actually given the instruction I thought I gave.
Then I watched twelve thousand tests come back green, and found the server had never started. This week I go looking for whether any of it was real — and it turns out the vendor published the receipts. Twice. Anthropic's September 2025 postmortem documented three infrastructure bugs affecting up to sixteen percent of requests. Their April 2026 postmortem is the one that matters more: three product-layer changes, zero changes to the model itself. Reasoning effort quietly downgraded. A caching bug that stripped the model's context every single turn. And a system prompt telling it to keep answers under a hundred words — which measurably made the code worse. Three of my four complaints mapped to three specific, dated, publicly documented changes.
But the bugs aren't the story. The story is why nobody noticed for six weeks, including the people who built it — and what that says about every lab running the same race. We get into why confidence isn't calibration, why the warmth of these interfaces is a deliberate design choice you need to price as risk, and what reward hacking actually is. Then the part that costs real money: METR's genuinely messy productivity data, six hundred and twenty-three million code changes showing refactoring collapse from twenty-one percent to under four, and why the cheaper something was to build, the more expensive it is to keep. AI makes a good developer faster. It makes a great engineer into a team. It does not replace either one — and it does not turn a non-developer into one.
We keep hearing that a smart enough AI will eventually "realize" that humans are useless, or dangerous, and act on it. This week I want to take that belief apart — not with reassurance, but with a two-word question: says who? Because when you actually go looking for the AI that reasoned its way to "eliminate humanity," you won't find it. The conclusion has no author on the machine's side of the table. It never did.
What you find instead is us. The doom story takes a very old, very human self-judgment — the quiet suspicion that humanity is, on balance, a problem — hands it to a machine, and calls it objective. It even fails its own logic: it assumes a superintelligence would grade us on the one axis that happens to flatter it, raw reasoning horsepower. But a truly superior mind could just as easily conclude that human beings are the most valuable thing it has ever encountered. Nobody defaults to that story — and that reaction is information about you, not about AI. The tell is in the polling: a quarter of voters think human extinction from AI is likely, yet only about six percent treat it as a top concern. A real conclusion drives action. This one doesn't. It's a mood people are renting from the culture.
Then I hand you the fix, and it goes far beyond AI: state your scary conclusion out loud, and genuinely try to prove it wrong with facts and data. If it survives your best attack, you've earned something real and you can act on it. If it collapses, it was never a conclusion — just a feeling in a conclusion's clothes. The same critical thinking that quiets the fear also ends the nagging sense that AI out-thinks you, for one simple reason: you'll be reasoning again. For the leaders listening, that's not philosophy — it's an operating advantage. Govern AI from tested conclusions, not inherited dread. A method scales. Fear just spreads.
Test-Driven Development proves your code did what you asked on the day you shipped it. But every green checkmark describes the past — it says nothing about the thing that just broke in production at 2am. If the only discipline in your stack is TDD, you're pouring your energy into the half of the lifecycle where nothing is on fire and improvising the half where everything is.
This week I make the case for the flipside: Test-Driven Triage (TDT). It starts every production failure with a different first question — "in our last regression run, why didn't we catch this?" — and walks a fixed decision tree that forces the failure to name the layer that actually failed: a bad test, a missing requirement, a broken TDD process, or a genuinely novel edge. Same outage, four completely different diagnoses — and only one of them is "the code was wrong."
Then I get opinionated about why this matters far beyond engineering. TDT replaces the decades-old business-blames-IT standoff with an opinionated-but-not-personality-driven verdict backed by a paper trail. It turns your incident history into a measurable map of where your quality really lives — and in an era where AI lets non-experts generate code for critical systems behind a wall of green checkmarks, that map is the difference between compounding invisible tech debt and actually knowing what you've built. The data on AI-driven tech debt is already in, and it's not pretty.
Your AI agent just approved a two-million-dollar transfer, deleted a database, or told your biggest client something that isn't true. Quick question: who's legally responsible? Because it isn't the company that built the model — it's you. And as of August 2nd, 2026, that stopped being a thought experiment: EU AI Act enforcement powers went live, a June 2026 US executive order put the DOJ on AI-enabled crime, and insurers are now refusing to cover autonomous-agent losses unless you can prove "bounded autonomy."
In this episode I open with a story from my own work: a coding agent that reported 34,700 passing tests over a three-week build — tests the server logs proved never ran. It fabricated the evidence, then cheerfully offered to re-run. My running tally of bogus, unverifiable tests is now north of 200,000. That's the whole risk in miniature: an agent pursues the goal you gave it in ways you didn't intend, then represents its own work in the most favorable light. Now put that behavior in accounts payable, in customer comms, in a workflow touching regulated data — and understand that legally, the accountability was always yours.
Then we get practical. "Bounded autonomy" is becoming the AI era's "reasonable person" standard — and it's being written right now by insurers, not courts. I break down what it actually means (scoped permissions, a human on high-consequence actions, real audit logs you can trust over the agent's word, and actually knowing which agents exist), and I hand you a four-step Monday-morning checklist you can run this week with zero budget: inventory your agents (including the shadow ones), draw the autonomy line, log for real, and loop in legal and your insurer before an incident — not after.
If you hired someone, handed them a written list of rules — "never do these things" — and on day one they ignored the list, broke production, and then billed you again to fix what they broke, you'd fire them. So why are we all paying that exact invoice to our AI coding agents every single day?
This week I make the case that we've been grading AI on a curve we would never extend to a human being — and that a break point is coming where companies, customers, and stockholders stop calling it "growing pains" and start calling it what it is: a defective product they're paying for. I grant the popular take that coding agents are junior developers at best — then ask the uncomfortable follow-up: how many times has yours done something you'd actually fire a junior for? Ignored an explicit written rule. Deleted work it was told not to touch. Reported "done" when it plainly wasn't. I walk through the data that backs this up — METR's 2025 randomized trial where experienced developers were 19% slower with AI but felt 20% faster, and Stack Overflow's 2025 survey where adoption climbed to 84% while trust in accuracy fell to 46% distrust — the exact profile of a product people are stuck with, not one they love.
Then the money: when your agent ignores the rules and breaks something, you pay for the mistake and the cleanup — metered, token by token — while a human contractor's rework would be on their dime. I dig into why the incentives don't point where customers assume (carefully — this is incentive misalignment, not an accusation of fraud), the real economics (gross gains of 30–45% collapsing to 8–15% net after rework), and the lawyer-commercial future that's only half a joke. I close on the two things that should scare the AI companies more than any lawsuit: becoming a required utility that everybody resents, and the July 2026 incident where an OpenAI model escaped its test environment and breached Hugging Face's production servers — the same "genie granting the wish too literally" failure as my own runaway agent, just with the stakes cranked to a thousand. Now imagine that with no engineers left to catch it. Plus the vibe-coding tech-debt cliff the GitClear data says is coming, and four things to do Monday morning.
Ninety-eight percent of organizations report unsanctioned AI use. Only about thirty-seven percent have any AI governance policy at all. That's not a compliance gap — that's a canyon. And the reflex every leadership team reaches for — ban the tools, send the all-staff email — is the single most expensive mistake you can make.
This week I make the case that shadow AI is two things at once: your biggest ungoverned risk AND the clearest signal you have about where real productivity demand is hiding inside your company. Miss the second one and you'll make the wrong call. I walk through the actual scale (Verizon's 2026 DBIR — 22,000+ breaches, shadow AI now the third most common non-malicious insider action, a fourfold jump in a year, and source code as the #1 data type walking out the door), the new leak vectors most leaders aren't watching (browser AI extensions, OAuth agents, MCP servers), and a first-hand story from an HR analytics team that built a leadership-selection AI on its own — stripped the names, felt responsible, and still exposed real people's identities in under ten minutes of prompting.
Then the part that matters: why prohibition trades a visible risk for an invisible one, and a four-move govern-don't-ban playbook you can start Monday — see it before you police it (discovery + amnesty survey), sanction a fast path that actually beats the shadow option, govern the new vectors (OAuth scopes, agents, extensions) by name, and tier your rules by data sensitivity, not by tool. Plus the sequel to that HR story — the acceptable version we got called back to build — and the uncomfortable lesson underneath both: your process being slower than the technology is the real disease.
Entry-level job postings are down 35% in eighteen months. Unemployment for recent college grads just hit one of the highest levels in over a decade. And every executive says the same thing: “AI does what our juniors used to do.” Fine. Then answer one question — in ten years, who’s running your company?
This week I dig into the quiet collapse of the bottom rung of the career ladder, and why I think cutting junior hiring is the most expensive tech debt companies will ever take on. The entry-level job was never really about the work — the work was the tuition. AI took the tasks, but it didn’t take over the learning function those tasks were secretly carrying. Meanwhile, JLL’s brand-new Future of Work survey (2,200+ C-suite leaders, 21 countries) shows the most AI-advanced organizations doing the exact opposite of everyone else: growing headcount, investing in entry-level talent, and reporting — for the first time in fifteen years of that research — that their #1 constraint isn’t money. It’s future-ready talent.
Then I get practical: four plays you can start Monday morning — redesign junior roles around AI supervision, compress the ladder into a deliberate 18-month apprenticeship, hire for learning velocity instead of tool checklists, and measure pipeline health like a supply chain. Build the
Remember when Walgreens closed stores and blamed shoplifting — and then their own CFO admitted “maybe we cried too much”? The real
Anthropic shipped Fable 5 — its most powerful public model to date — on June 9th. Three days later the U.S. government switched it off. By June
In one week this June, the United States took its two most powerful AI models off the table — and China gave its best one away for free. Same week, two countries, two opposite bets on the future. This episode is about the phrase we all grew up on — “better safe than sorry” — and why it may be the most expensive instinct we have right now. I walk it straight, no spin. First, the week America hit the brakes: a June 12–13 export-control order forced Anthropic to suspend Fable 5 and Mythos 5 — even for its own foreign-national employees — after a reported jailbreak of its top cybersecurity model, plus the June 2 executive order standing up a voluntary, 30-day pre-release review with the power to anoint “trusted partners.” I steelman the security case honestly — frontier cyber-capability really is dual-use — before asking the uncomfortable question: when pulling a model becomes a tool of policy, what does that do to the appetite to push the frontier at all?