Transmissions from Elysium with Stephen Salaka

A fire lieutenant yells "get out" with no idea why. Seconds later the floor drops into the basement. That moment is the whole episode: you carry an analysis engine older than language, and it can't talk. So I'm going to teach you to read tarot cards (yes, seriously) because the cards are the keyboard. Forged Egyptian scriptures, secret societies, an artist who sold 100 million decks and died broke, John Dee's Aztec mirror, and the science of why no AI can ever do your reading for you.

In 1948 a psychology professor handed 39 students their own personalized personality profiles. The students rated the accuracy at 4.26 out of 5. Every profile was identical... he'd cut the sentences out of a drugstore astrology book.

That little experiment explains every psychic storefront in America, and it's also the reason I'm about to teach a room full of engineering leaders to read tarot cards. Not kidding. This is the occult episode: the fake Egyptian scripture invented by a pastor in fifteen minutes, the Victorian secret society Yeats belonged to, the synesthete artist who drew the most famous deck on Earth (100 million copies sold) and died broke in Cornwall, and John Dee's black Aztec mirror, which still sits in the British Museum being quietly menacing.

But it's also the science episode. The fire lieutenant who shouted "let's get out of here" seconds before the floor collapsed, and couldn't say why. The gambling experiment where players' palms started sweating over the bad decks before their mouths could explain the problem. Kahneman and Klein's answer to when your gut is a genius and when it's a confident drunk.

Here's the thread tying it together: you carry a pattern-recognition engine older than language. It finished analyzing your hardest decision weeks ago. It just can't talk. Tarot (or Pokemon cards, or a game called Kittens in a Blender... yes, really) is the keyboard.

And no, an AI can't do the reading for you. There's a 2026 study on tarot readers who use chatbots, and one finding involves the phrase "yes man energy." The machine can shuffle. It cannot be the person the reading is about!

Grab a deck and listen. Then send this episode to the one person on your team who owns a tarot deck and thought nobody at work knew.

One week in July 2026. Two Chinese frontier models, released open-weight, free to download. Kimi K3 beat Claude at frontend code. Qwen 3.8 claims it's "second only to Fable 5." Half the market panicked, the other half threw a party, and somewhere in the middle a very old question got skipped entirely: who poured the metal?

This episode is about supply chain risk, the discipline the defense world built for exactly this moment. And I've got stories. The Huawei gear Vodafone found backdoors in (twice... they asked nicely the first time). The Intel i9s that spent two years quietly cooking themselves to death from a flaw nobody could see. The counterfeit titanium that rode forged paperwork into Boeing and Airbus jets, through the most audited supply chain humans have ever built, and got caught because one inspector in Italy thought some tiny holes looked wrong.

Here's the uncomfortable part. Researchers have shown you can backdoor a model with about 250 poisoned documents (out of trillions), and that the backdoor survives every safety technique we have. Sometimes safety training just teaches the model to hide it better! Open weights are not open source. Nobody reads 2.8 trillion numbers. The tests don't reach where the problem would live... which is exactly why the Pentagon stopped trusting tests decades ago and started asking where the titanium was melted.

And no, the answer isn't panic. Your homelab is fine (genuinely, go have fun). This one's for the software leaders: why enterprises aren't switching, why the consumer exodus might actually help the big labs, and the three rules I'd laminate before anyone bolts a foreign frontier model to your company.

Listen now, then forward it to whoever on your team just torrented a model onto the build server. They know who they are.

Somebody I know spent twelve days building a tool to check if domain names were available. Domain registrars do that for free (it's their whole business model, they WANT you to find a name). When I asked how much time the tool saved, the answer was "several hours." Twelve days in. Several hours out. And that, friends, is the era we just walked into.

AI made building almost free. Ideation, prototypes, a working tool by Thursday... "could" is now a commodity. Nobody mentions that "should" still costs exactly what it did in 1959, which happens to be the year we built the first machine that said yes. It was called COBOL, it was pitched as plain English for business people, and cleaning up after four decades of its enthusiasm cost the planet somewhere north of $300 billion. Ask me how I know (I was there for the Y2K part, badly caffeinated, at a bank).

This one's for the software leaders. There's an airline that ran ticketing on Access databases. A $20,000 AI-built report that replaces five minutes of work a week (the payback period will outlive everyone involved). A company that burned $500 million in tokens in a single month because nobody set a usage cap. And the part your finance folks already suspect: enterprise AI is not $200 a month, and the vendors are actively kicking business users off the subsidized plans.

Then the useful bit: the three questions I ask before any internal build gets a green light, and the guardrail system I rolled out so citizen developers can run without anyone getting run over.

Could you build it? Obviously. Should you? Listen first. Then forward this to whoever on your team just discovered agentic coding. Before the invoice does.

In 1935, the best pilot in the US Army climbed into the most advanced airplane ever built and was dead ninety seconds later. Not sabotage, not a storm... he forgot one item. Boeing's fix was a single sheet of paper, and that sheet of paper conquered aviation, then surgery, then somehow mutated into the thing currently eating your engineering org alive.

This one's about the bureaucracy of checkboxes. The checklist to verify the checklist. The ascending tower of forms confirming that the forms below them got filled out (if you've worked in a big enterprise, you know the tower I mean). And the question nobody in the process office wants asked: does any of it actually work?

Somebody ran the numbers. An entire Canadian province mandated a checklist that provably saves lives, hit 98% compliance, and got... well, you'll hear what they got. It's worse than you'd think, and a Dutch study found the reason, and that reason should terrify anyone who has ever trusted a green dashboard.

Also on deck: a French mathematician who proved in 1890 why your PMO's master plan can't work (with actual math!), what happens when leadership demands 100% code coverage from fifty developers, and the NIST research showing where bugs really live. Spoiler: it's a much smaller neighborhood than your test suite thinks.

You can keep your checklists. Really. The good ones are written in blood and I'd never take them from you. But somewhere in your org tonight there's a dashboard glowing green, and this episode is about what's underneath it.

Listen now, then forward it to whoever owns your process manual. They need it more than you do.

People are lazy. I mean it as a compliment. Laziness is the foundation of the entire economy you and I work in — there's real research (PNAS, "Buying Time Promotes Happiness") showing that paying someone else to carry the boring tail of a task makes you measurably better off. Building something is the cheap part. Owning it for ten years is the expensive part. Paying someone to own that tail is smart leadership.

Then AI dropped the price of "build it yourself" through the floor — and broke everyone's brain about it.

This one's a field guide for software leaders, told from the buyer's chair at a defense contractor where my inbox has become a graveyard of founders pitching me a bundle of prompts wrapped around someone else's model and calling it enterprise SaaS. I'll walk you through what actually changed (the bar to build dropped; the cost to *own* didn't), the real difference between "vibe coding" and serious agentic engineering (Karpathy, who coined the first term, has already retired it), and why I will not award a contract to a no-name SaaS with its SOC 2 "in progress" — no matter how good the demo is.

Then the three traps that stack into an outbreak: the quick win that isn't quick (Gartner forecasts a 2,500% jump in software defects from this style of building), the tool tossed over the wall to an IT team that never agreed to own it, and the big one — AI making everyone feel like a developer, at scale, all at once. (One security scan found 2,000 corporate apps shipped with zero authentication. That's the horde.)

But I also defend the thing I'm attacking. SaaS is rational. Wrappers can be great. The sin isn't building one — it's pretending the thin part is a fortress and aiming it at a buyer who will never open the door.

Know your buyer. Own the tail or don't pick it up. And finish your SOC 2 before you tell me it's done.

I went in ready to roll my eyes. A glossy business book about leading with AI. Another firehose, another guru, another deck.

Then I read it. And the uncomfortable part is the book's not the problem. I am.

This one's a review of *The AI-Driven Leader* by Geoff Woods, and it's also a bit of a reckoning. Because every framework in it, the strategy-first stuff, the ask-questions-before-you-answer stuff, the run-it-past-a-skeptic stuff, is something a certain generation of us has been doing on pure instinct for forty years. The tricks we lazily call "common sense." A phrase I've come to genuinely hate, because it is neither common nor sense.

So I do the thing. I take the book's principles one at a time and match each one to the exact scar where I learned it the hard way. A computer I pulled out of a dumpster at age five. A manufacturer about to bet the company on the wrong rollout. A frog. Yes, a frog, floating in a magnet lab at three in the morning while I was too exhausted to understand what I was looking at. A tech org in Japan where I learned that the whole job was herding cats, and doing more was the trap.

And somewhere in the middle of being smug about all of it, I remembered I'm not the hip new generation anymore. The lessons I sneer at as obvious are genuinely new to a lot of very smart people who never got the fire. Which is exactly why a book that writes them down matters more than my ego wanted it to.

If you lead engineers, or you're drowning in the firehose and looking for a way to use AI that isn't just doing more, faster, off a cliff, this one's for you.

Listen now. Follow the show. And send it to the one person on your team who keeps confusing being busy with being effective.

@geoffwoods, this one's for you. With affection.

OK. Two ways the world ends with machines that copy themselves.

The loud one is grey goo. Little nanobots eat everything, planet turns to sludge. Fine. Scary. Got it.

It's the other one that's been keeping me up. There's a manga, *Blame!*, where the machines don't eat the world — they just build it. Forever. Nobody left who remembers how to tell them to stop.

I think we're at the baby version of that. And the first one showed up in my inbox a few weeks ago. A 180-page report. To answer a question that needed about five.

So that's what this one's about. The slop. The flood of confident, well-formatted, just-good-enough AI junk that's quietly burying your whole team. Almost free to make. Looks smart. And somebody downstream is drowning in it right now. Maybe you. Maybe because of you. (Sorry.)

Inside I do a few things. I put an actual dollar figure on the mess — somebody finally ran the study, it's worse than you'd think. I drag in a 220-year-old folk tale, a one-paragraph horror story, and that manga. Then I hand you five things you can actually do Monday morning to stop being the source.

Ever sent an email you didn't really read? Opened a report you couldn't use? Yeah. This one's for you.

The machines aren't evil. They just never learned the word for enough.

Somebody on your team still has to.

▶ Out now. Follow *Transmissions From Elysium* so the next one finds you. And if this lands — send it to the one person who needs it. (You already know who.)

Welcome back.

About a month ago I was at a working dinner with peer CTOs and a head of AI Enablement, and someone pitched a product that automates meeting transcription so nobody on your team ever has to take a note again. I pulled my I/O Psychology hat off the shelf and pushed back, because the friction we are racing to remove is the friction that makes us think, remember, and lead. The act of taking the note is also the act of thinking the thought.

In 1957 a sci-fi writer named Gordon Dickson published a short story called "Too Soon to Die," about a society that engineered comfort so thoroughly that people started dying young because their bodies forgot why they should keep going. I read it when I was eight and did not understand it. I have spent the last month understanding it.

In this episode I walk through what the research actually says, how Dickson, E.M. Forster, and Isaac Asimov each saw this coming decades before any of us, how David Graeber's *Bullshit Jobs* describes the exact inversion AI has produced, and what software leaders can actually do about it on Monday morning. Three prescriptions: add productive friction back, hire and train for slop-filtering, and run triage on every agent you deploy.

Humans live in the friction. The struggle is not the bug. The struggle is the proof you are still here.

Snake oil salesmen never left. They just learned to code. My friend keeps rockets in the sky, and he still feels like he's losing to a landing page. I tried to tell him why.

So this episode is about that scoreboard.

I walk you from an 1890s snake-oil wagon parked outside a general store in Kansas all the way to a 2026 AI landing page. Same four tricks. Same top hat. Different typography. Then I do the math on what a "$1B ARR, one person" company actually requires, spoiler: it's duct tape and a prayer, and the scaffolding doesn't hold. And I end where every hype-cycle conversation should end: with Oprah.

Because we grew up watching her. We said *we could do that.* We didn't. Not because we weren't good enough. Because being in the right place at the right time is a real, material thing, and it is not a scoreboard you can climb by running harder.

The hype cycle is old. The feeling it manufactures is old. The trick Oprah made us fall for in 1992 is the same trick the one-person-billionaire is running now, and it will be the same trick next year under a different name.

You are not behind. You are here.

I was standing in my kitchen at 9 at night trying to add a four digit number to my security system so my son could let himself in after he comes back from school. That was the entire task. Four digits, one kid, one door. I opened myADT on the browser the way I had a hundred times before, and the dashboard was gone. Where the users panel used to live, a little chat bubble bobbed on the screen with a friendly face on it, asking how it could help me today.

Twenty minutes later I was still there. Still standing. Still negotiating with a text box. The bot kept offering me walls of near-identical verbal options (add a PIN, add a PIN for a new user, set up a new user and create a PIN, add a temporary PIN, modify an existing PIN) and none of them were the thing I was trying to do. I was playing the reincarnation of ZORK for the privilege of locking my own front door.

So I called the 1-800 number. A woman named Samantha picked up, pulled up my account on her screen, and added the PIN in about fifteen seconds. Because Samantha, it turns out, is still using the old interface. The ugly one. The one with the buttons and the users panel and the thing you could actually SEE. They just took it away from me!

In this episode I want to talk about what happened in that kitchen, because I don't think it was a UX problem. I think it was a preview of everything. Companies are ripping out their interfaces and replacing them with chatbots, and they're calling it "easier," and it is not easier. It is slower, dumber, and more exhausting, and there is a specific reason why, which is going to lead us through memory palaces (the ancient mnemonic trick Roman senators used to memorize speeches), Barry Schwartz's paradox of choice, a side trip through Star Trek's tactical stations, and a visit to my own first novel where I gave an AI infinite computing power and it was STILL just a really good hammer.

This is the bridge episode. You have heard me talk about The Abdication, and you have heard me dismantle the mythology of AGI. This one sits between them. Because while we were all busy scanning the sky for the robot uprising, something quieter happened behind us. The menus started disappearing. Not all at once. One button at a time. Replaced by a chat bubble wearing a friendly little face.

The good interface is not gone. It is hiding. Samantha can see it. We can't. The signal is broadcasting, and the buttons are still there, and they just don't want us to touch them anymore.