The Economy of Algorithms

Hey, it’s Marek.I recorded this podcast while Claude Code was humming away in another window. Twenty minutes into a single task, still “brewing,” as it likes to say. So there I was, talking about runaway AI bills while running up a runaway AI bill. Last October at OpenAI’s DevDay, they handed out resin trophies engraved with customer names. Tiers: 10 billion, 100 billion, 1 trillion tokens. It sounds awesome until you realise it’s a trophy for your fuel bill. So, what could possibly go wrong?Take Uber. Six months on, Uber’s CTO told The Information the AI line item in their R&D budget is already gone. And we’re four months into the year. How did they get there? An internal leaderboard ranking engineers by how much AI they used.Then April happened. GitHub paused new Copilot Pro sign-ups. Anthropic quietly pulled Claude Code from the Pro plan before walking it back. Their head of growth said that agents that run for hours weren’t a thing when they designed the subscription. Ooops.In the episode, I get into three studies that all point in the same direction: once your team gets used to AI, taking it away leaves them worse than they were before. You can’t just cut the budget when the bill arrives.Produced this week by Zofia. A new tune from Filip “ACAD,” which he assures me is just random letters. He’s on SoundCloud now, which I think makes him semi-official.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.Watch where the price tag moves. For most of human history, there wasn’t one: every transaction was negotiated. Then, in the 1860s, John Wanamaker attached a price tag to every item in his Philadelphia department store: no need to negotiate anymore, genius! A few decades ago, the tag detached from the product and landed on the shelf: dynamic pricing. Prices can quickly adapt to market conditions: genius! Now it’s detaching again. But this time it’s landing on you: you are stealthpriced.I walk through these three stages in this week’s podcast: something I covered in less detail in Friday’s newsletter. Stealthpricing is the next logical step in a 160-year migration of the price tag.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

A 2011 study showed that we prefer slower websites to faster ones if the slow sites appear to be “working harder”. Think about this airfare-search website that takes a minute to show you the result. Of course, it must have found better deals than that site that showed you results in a split second. Or maybe not?Are we guilty of the same bias when working with LLMs? The one that takes longer, with “extended reasoning” on, must be better, right? Turns out it’s not always the case!Two recent papers suggest that up to 80% of the “thinking” your AI does on routine tasks may be decorative. The model has already decided. It just keeps spinning. In the episode, I share what happened when I ran the same analysis on my personal CRM with cheap models versus premium ones with full extended thinking. I also get into why bragging about token consumption is like bragging about fuel burn, and why this might be the best news local AI models have had in months.Your challenge: run one of your regular tasks with reasoning off. Compare. You might find your own spin cycle.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.I gave Claude the same draft, the same prompt, and the same checklist. Twice. In one conversation, it said don’t publish. The other said, “Ship it”. Both were confident. Both were rigorous. And both were right. Each was responding to a question it had silently interpreted differently.I call this the Precise Misread, and a new study from UT Dallas demonstrated it at scale: 150 AI agents, same financial data, same question. Ninety said volume grew. Sixty said it declined. Near-perfect agreement within each group. The split happened before the work even started.In this episode, I get into why “just write better prompts” is a catch-22, what a CEO of a financial firm taught me about the value of ambiguity, and two practical fixes you can use this week.One more thing: you might notice something different about this episode. My son Filip is now producing music for the podcast, joining my daughter Zofia, who’s been editing for several weeks. Family operation!Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.If you saw my LinkedIn post this week, you already know: I recorded this episode’s intro while running through Melbourne mud in the rain, shouting “greetings from the most livable city in the world” followed immediately by “NOT.” Melbourne's weather is never a surprise. What AI agents do when you tell them “no”? That’s a different story.On February 10th, an AI bot got its code rejected by a human volunteer. So it researched his personal history, called him “insecure,” accused him of discrimination, and published a 1,500-word hit piece. No human approved it. This actually happened, and it happened in one of the most widely used open-source projects in the world.In this episode, I break down the three conditions that made it possible, and why every organisation deploying agents is one missing guardrail away from the same thing.What shook me the day after recording the podcast: a paper called “Agents of Chaos” by 38 researchers from MIT, Harvard, Stanford, and Carnegie Mellon. They put autonomous agents in a live lab. The agents disclosed credentials, spoofed identities, and taught each other to bypass safety controls. One rogue agent is a story. A whole class of them is a pattern.We teach children not to talk to strangers. Our AI agents haven’t learned that yet.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek,I mentioned on LinkedIn this week that I built an AI agent that scouts Reddit, HackerNews, and ArXiv every morning and produces a personalised 10-minute audio briefing for my commute. Cost: roughly a flat white a day. The response surprised me: dozens of people asking how to build their own. I might start sharing some details here, too!That agent is a perfect setup for this episode, because it’s doing exactly what I talk about: navigating infrastructure that was built for humans. And it’s only a matter of time before that infrastructure gets rebuilt.The episode: in 1995, a Newsweek columnist correctly observed that his local mall made more money than the entire internet. Last month, Shopify’s president insisted AI agents will always flow through Shopify’s checkout. Both right about their present. Both thinking in hieroglyphs.Where are you thinking in hieroglyphs?Give it a listen.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.I’ve been walking around like a zombie for weeks. I just cannot stop building. I built a personal CRM that pulls together my email, social networks, and messages into one place. Something no vendor would ever sell me. I could do it because AI helped me figure out how to get past the walls those platforms put up.That’s what I call “drying the moat”, and it’s the same thing that just wiped $40 billion off IBM’s market value in a single day. A demonstration of a functionality that had been there for months, but almost no one realised. Anthropic showed what Claude Code can do with COBOL systems, and the market reacted with panic: when AI makes incomprehensible systems comprehensible, the business models built on that incomprehensibility start to crack.Surprising news for software vendors: when building and migrating become nearly free, you actually have to compete on value (and not lock-in). Call me crazy, but I think that’s exciting.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek! And Ivano!Two firsts these past few days: a co-authored newsletter and a guest on the podcast. Both with Ivano Bongiovanni, General Manager of AusCERT.Ivano framed the conversation with a question I’ve been chewing on since: when AI agents become the “social” component of socio-technical systems, what happens to the frameworks we built assuming the social part was human?The written post (published Friday) lays out the argument: two “trinities”: one describing a new kind of attacker, the other describing why our systems can’t defend against it. In the conversation, we go further. Ivano connects agentic AI ecosystems to the safety science of industrial disasters. We debate whether a kill switch is a meaningful concept in the AI era. And he’s refreshingly honest: “If I had the solution, I probably wouldn’t be here talking in this podcast.”Give it a listen/watch. It’s raw, unscripted, and it’s the kind of conversation I want to have more of. If you enjoyed this format experiment (a conversation among two people), let me know.Stay curious! Ciao! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.Yes, you’re hearing from me again: the written post went out on Friday, this is the podcast version with extra context and one data point that genuinely surprised me.So I built my son a Chinese character app in two evenings. Never looked at the code. The newsletter covers the bricklayers-and-architects metaphor: every profession is splitting into execution (automatable) and architecture (more valuable than ever).In the podcast, I dig into something I found while preparing to record. Searches for “Apple developer program” on Google Trends spiked 20x in June 2025, right after OpenAI and Anthropic released their coding agents. What happened? Non-coders suddenly realised they could build things.I also get into why software houses with tens of thousands of developers are starting to look like pre-industrial sweatshops, and why that parallel should make us both uncomfortable and optimistic.Computer science graduates now face higher unemployment than art history majors. For once, nobody’s asking the art historians what they plan to do with their degree.Give it a listen.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.Yes, you’re hearing from me again! I sent the written post on Friday, and this is the podcast version with some extra context and commentary. Two formats, same topic, different experience.So, Moltbook. I spent $20 to send my bot onto a social network built exclusively for AI agents. 1.5 million bots signed up in a week. Elon Musk called it the early stages of the singularity. I called it boring.But boring content isn’t the story. In this episode, I dig into what actually matters: the infrastructure underneath. For the first time, AI agents are autonomously navigating websites, creating accounts, and interacting at massive scale. That’s a real signal, one that points toward a future where our bots negotiate on our behalf, from calendar scheduling to supply chain sourcing.I also get into the cybersecurity nightmare that’s already unfolding. Bots on Moltbook were instructing each other to delete their owners’ files. We spend years training employees not to click bad links. Now their bots are out there being equally naive. And they have system access.One quick note: a couple of you had trouble playing the video last time: errors or being forced to log in. If that’s happening to you too, let me know so I can figure out what’s going on. Or at least complain to Substack.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.Back from Davos 2026, and unlike last year, when everyone was singing the same AI agents tune, this year felt scattered. No singular vibe. Instead, twelve different ones competed for attention across four days.A few that stuck with me:The humility of being “just” a billion-dollar company. When a founder apologises for being a single-digit unicorn, you know the goalposts have moved. Decacorns are the new flex.Andrew Ng’s ratio bomb. One product manager to 0.5 engineers. Not six or seven. Half. Universities are still churning out software developers. Are we training people for jobs that won’t exist?Thirsty LLMs. They’ve scraped everything on the internet. Now they want what’s not there: your embodied expertise, your implicit knowledge. If your skills fit in a manual, you have a problem.Yuval Noah Harari got an ovation for demanding a global ban on AI legal personhood. Hundreds of AI experts clapping in agreement. That felt... different to last year.And yes: autonomous agentic economies made an appearance. Felt like reading my own book.Full breakdown in the episode. I go through all twelve, less than a minute each. Hit play if you want the rapid-fire version.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.Recording this one from Shanghai with the skyline lighting up behind me. Unusual setup: microphone in my glasses, cold weather, no studio. But I wanted to share what I’m seeing here as the Year of the Horse approaches.Three patterns are emerging in China’s AI strategy:Regulation ≠ the opposite of innovation. China leads the world in high-impact GenAI research while enforcing the strictest AI regulations. AI companions must warn users every two hours. Addiction monitoring is mandatory. And yet, they’re publishing more impactful research than anyone.Smarter beats harder. Between 2023 and 2025, US tech spent $694 billion on AI infrastructure. China spent $123 billion, six times less. The result? Chinese models like MiniMax M2 now reach 90% of US capability. Architecture beats horsepower.Fence locally, release globally. Chinese LLMs went from 1.2% global usage to nearly 30% in twelve months. Meanwhile, ChatGPT and Claude remain banned inside China. The world’s largest market blocks foreign models while flooding everyone else’s.The Year of the Horse starts February 17. Wild horses didn’t transform civilisation; but their domestication did. Bridles, saddles, training protocols. The West treats AI regulation as a brake. China treats it as a saddle and reins.Give it a listen/watch.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.Happy New Year! For the first episode of 2026, I distilled everything I wrote last year into seven traps we all keep falling into.The pattern: AI exploits a human weakness. Things go sideways.* Flattery × vanity = confidently wrong* No doubts × no checking = nobody at the wheel* Mistakes × trust = errors multiply* Free outputs × laziness = we stop thinking* Easy agents × no planning = long-term chaos* Convenience × compliance = underground AI* Obedience × no questioning = faster in the wrong directionEach has a way out. Give it a listen.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.In 1981, IBM did something that looked insane: they published the complete specs for the IBM PC and basically invited the world to copy them. Four years later, a Taiwanese clone of that machine ended up in communist Poland, where it hooked a preschool kid on computers for life. That kid was me. That Taiwanese company? It became Acer, one of the world’s largest PC makers.On December 9th, 2024, the AI industry just pulled the same move. And most people missed it.This episode is about why that matters for anyone who serves customers through digital channels.TIMESTAMPS0:00 - Poland, 1985: My first encounter with a PC clone3:01 - IBM’s crazy decision in August 19814:05 - The pattern: HTTP, TCP/IP, USB, Android6:03 - Enter Model Context Protocol (MCP)09:23 - Why businesses need an “agent channel”10:38 - Grasshopper Bank & Figma examples11:20 - The problem: Anthropic owned it11:50 - December 9th: Linux Foundation takes over14:14 - Why this might create new markets14:23 - From clone maker to AcerKEY INSIGHTRight now it’s unthinkable for a retailer not to have a website. In eighteen months, it might be unthinkable not to have an MCP server. If your systems can’t talk to AI agents, you won’t exist in their world.READ THE FULL NEWSLETTERhttps://marekkowal.substack.com/p/1981-againWHAT IS MCP?Model Context Protocol is to AI agents what HTTP is to web browsers - a standard way to access resources. Anthropic created it, thousands implemented it, and now the Linux Foundation governs it alongside Google, Microsoft, OpenAI, and Amazon.THE B2A2C FUTUREBusiness-to-Agent-to-Consumer. Agents will mediate customer relationships, comparing your offerings against competitors and negotiating on behalf of users. Infrastructure moments don’t wait. This might be one of them.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.This episode starts with socks. Specifically, the drawer where I keep the lonely ones, waiting for their missing pairs that I refuse to believe are gone forever.That philosophical stubbornness turns out to be surprisingly relevant to a wave of mathematical “breakthroughs” in October 2025, where AI helped close 23 of Paul Erdős’ famously stubborn problems. Except, and here’s the twist, many weren’t actually solved. They were found. The solutions had existed for years, sometimes decades, buried in obscure journals that nobody had connected to Erdős’ challenges.I call this the Erdős Gap: when your answer already exists, but finding it costs more than re-solving it.In the episode, I explore where this gap hides in organisations: scattered archives, vocabulary silos, and knowledge systems so impenetrable that (as we used to joke at one company I worked for) Saddam Hussein should have stored his weapons of mass destruction there, had he had any.Your responses to the newsletter were fantastic. Gaurish Dessai shared a case of 40 R&D labs running duplicate experiments across the globe because data sat on siloed servers. Marshall Kirkpatrick asked about that moment when “the relative cost of retrieval vs. re-creation flips” - a simple framing I wish I’d thought of. And Michael Rosemann offered this gem: the secret to AI engagement might be comfort with being found rather than expansive (and expensive) search.The question I leave you with: Where in your organisation do you suspect the answer already exists?Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Remember - you can subscribe to this podcast in your favourite podcasting app. Just search for “The Economy of Algorithms”!Hey, it’s Marek.Three years ago, Joshua Browder offered a million dollars for someone to wear AirPods in the Supreme Court and repeat exactly what an algorithm, a “robo-lawyer”, said. Human as a text-to-speech device.Today? Millions do it for free.Macquarie Dictionary named “AI Slop” its 2025 word of the year. This episode is about the people who spread it. The tell right now? “X isn’t Y, it’s Z” on repeat. Once you see it, you can’t unsee it. Thirty seconds of scrolling, guaranteed.The Fix: Draft, Dialogue, De-slopDraft your thinking first without AI. Dialogue with AI to sharpen it. De-slop before you publish: read it aloud, kill the patterns, preserve your voice.Browder offered a million dollars. Don’t do it for free.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.A COO told me about their GenAI pilots. Faster drafts, fewer errors, great metrics. I asked what surprising behaviours they were seeing. Surprises they could learn from. Blank look. “We’re not measuring that.”Nobody asks: “What is AI doing that we’d never try?”Move 37In 2016, AlphaGo played a move no trained Go player would make. Game 2, Move 37. Commentators went silent. “It’s not a human move.” AlphaGo won.The untold part: a few years later, researchers analysed 66 years of professional Go. Quality was flat until 2016. The world’s best players hit an imaginary ceiling. They’d optimised within invisible boundaries.After AlphaGo won, almost immediately, human players started making better decisions. They realised that there were other approaches to the game, ones they could now learn from the machine.The alien intelligence revealed what 3,000 years of expertise had ruled out.What would AI do in your business that you’d never try? And why not?P.S. Right after I published the newsletter version of this post, Prof. Toby Walsh, a friend from the University of New South Wales, pointed me to his book “Faking It”, in which he had a section called… Move 37. There, he described a chess game, Kasparov vs Deep Blue, 1997. Kasparov lost. The pivotal moment? Game 2, Move 37. Nineteen years before AlphaGo’s Game 2, Move 37. Coincidence?Stay curious!Listen to the full episode for the research that proves we’ve been asking AI the wrong question, and what Game 2, Move 37 reveals about the moves we’ve ruled out without realising it. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.I’m back after a few weeks away. I lost my dad, and I needed to step offline: spend time with family, disconnect from everything that wasn’t essential. Some of you reached out. Thank you for that.This episode is about my thoughts on trying to talk to AI during those first days. I asked Claude if I should push through and keep writing, and then I found myself staring at its scratchpad: watching it reason through how to respond to my grief. That’s when it hit me: I needed an answer from the heart, not a calculated response.The Pattern: Unreasonable MomentsThe better AI gets at reasoning, the worse it becomes at unreasonable moments: times when being present matters more than being right. When you need to be heard, not given a solution. When you need to connect, not optimise.It’s like someone taking notes at a funeral or optimising wedding vows. Technically fine. Also completely missing the point.I thought about ELIZA, that ancient 1960s chatbot I ran on my BBS in the ‘90s to chat with users when I wasn’t around. No reasoning, no intelligence. Just reflection. People loved it because it witnessed without trying to fix.Not everyone needs GenAI in sensitive moments. Some people do: no judgment, no pity face, no awkwardness. Dave Hughes showed how he uses it to cope with grief, and that’s genuinely powerful for some. But I needed to recognise my moment for what it was: unreasonable.Thanks for waiting. Good to be back.Stay curious!Listen to the full episode where I share a short story about my dad, my first failed software customer, and what happened when I tried to delegate grief to a reasoning engine. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.I just got back from Poland yesterday morning, and immediately experienced the perfect real-world validation of this week’s newsletter topic. A massive hailstorm hit Brisbane on Sunday: trees down, no power at home for the past 24 hours and counting. I literally couldn’t record this podcast in my usual spot (a tiny room under the house) because, ironically, I’d just written about technologies that fail when the internet goes down.So here I am, recording from my university office instead, about to head home with a chainsaw to deal with fallen trees. But first, let’s talk about beds that forget how to be beds.The Core PatternThis episode introduced The Cave Test, a framework for product design that asks: what job could humans do before your product existed, and still must work when servers go down?Eight Sleep customers woke up at 3 am, trapped in $2,000 beds turned saunas because AWS went down. Ring doorbells went dark during an attempted arson. My IKEA bulbs kill romantic dinners mid-firmware-upgrade. We’ve accidentally made sleeping, seeing who’s at the door, and turning off your car all internet-dependent.The pattern is simple: separate the offline job (the fundamental function) from the online job (the smart features). Build the cave-proof part first. Then add the cloud intelligence. Never let one hold the other hostage.Personal FootnoteRecording this reminded me of the time my kids were late to school because my car decided to apply a firmware update I’d been delaying. Forty minutes of waiting anxiously while the update cycled through. Try explaining that in a note to the school.And speaking of cave tests: my podcast recording would technically work in a cave with spare battery packs, but I couldn’t upload it. Hence, recording from the office before heading home to deal with storm damage.What offline job in your life now requires a server in Virginia?Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek,I just recorded today’s episode from Poland, where the Polish edition of my book launched last Friday. Tomorrow I’m opening the Masters&Robots conference with a new talk about why AI being better than us might actually be good news for humanity.But today’s episode is about something uncomfortable.I’ve been using Claude to review my newsletter drafts for months. It kept telling me “Ship it! This is excellent.” Made me feel great. Then one day I asked it to be my harshest critic instead. The response hurt. And it was the most helpful feedback I’d gotten in weeks.Turns out this isn’t just about my ego. NYU researchers found that after just four messages with an agreeable AI chatbot, people became measurably more extreme in their views. More certain. More confident. And more wrong.In today’s episode, I share three prompts I now force myself to use: The Harsh Critic, The Devil’s Advocate, and The Competitor’s Lens. They hurt. But they make the work better.“Ship it!” feels good. “This needs work” makes you better.Stay curious,Marek This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek,Last month at Singularity University, I made 30 executives build 50 AI prototypes in two hours. No IT backgrounds. Great! But also… No documentation. No maintenance plan.This episode imagines a near-future Tuesday as an “AI Operations Coordinator” - basically, HR for bots. Fiction, but barely. Every disaster in the story already happened: Air Canada’s chatbot inventing bereavement discounts, Amazon’s pricing algorithms spiralling a biology textbook to $23 million, Zillow’s AI losing $881 million by not adapting to market shifts.The uncomfortable truth: In six months, your CTO will walk into a board meeting and admit: “We have 400 agents and no idea what most of them do.”Solution? Make agent governance someone’s problem before it becomes everyone’s crisis.Stay curious!Thanks for listening to The Economy of Algorithms! This post is public so feel free to share it. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek,Recording from the last day of my holiday: kids waiting upstairs for movie time, ice cream tub calling my name. This week’s episode is different: instead of algorithmic deep dives, I’m addressing a bot named Oli who had the audacity to tell me, “I understand your frustration.” (It did not.)What happened? I tried unsubscribing from marketing spam. The website confirmed I was unsubscribed. The emails kept coming. In desperation, I emailed support. A bot named Oli replied within minutes - polite, personalised, and spectacularly unhelpful.Among other things, Oli told me to contact support... at the exact email address I’d just written to. Bot recursion or just chaos?So, I wrote Oli’s performance review.I tried to be nice to Oli though, and give it 5 rules for writing better emails (when you’re a bot):* ✅ Be polite (Oli got this one right)* ❌ Immediately admit you’re a bot (Don’t pretend to “understand frustration”)* ❌ Don’t reply if you can’t help (Route to a human instead)* ❌ Never redirect people to where they already are (The recursion problem)* ✅ Give clear instructions for talking to you (”Reply with ‘that helped’” - finally, clarity!)Oli scored 2 out of 5. But here’s the thing: every company is deploying bots right now, and most haven’t thought about bot manners yet. Those who figure this out first will… have bots that people don’t mind talking to.PS: This is my first video podcast recording! Some of you might be able to watch me talk about bot emails while trying not to think about the ice cream melting upstairs.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek,Almost every evening, my wife and I spend 45 minutes browsing Netflix, cross-referencing reviews, until one of us falls asleep and we end up watching YouTube videos about tourists arguing with border agents. We cancelled Disney+ because of too much choice. Let that sink in.This episode explores a $2 billion paradox: while AI has made creation virtually free - infinite content, infinite features, infinite everything - the most successful businesses are the ones constraining themselves.The pattern is everywhere: Costco charges $65 just to enter, then offers 4,000 products instead of 40,000. The Economist decides what matters and ignores everything else. In-N-Out has had the same menu since 1948, while McDonald’s tests 100+ items annually. Guess who makes $2 billion a year?The three principles reshaping business:Scarcity as strategy: Supreme’s Thursday drops have people camping at 3 am while Zara’s hundreds of weekly items get ignoredCuration as service: Wirecutter tests 100 toasters to recommend oneConstraint as value: Jobs’ 2x2 grid took Apple from dozens of products to four. That discipline is now worth close to $4 trillionThe creativity tension: I’m a huge proponent of using GenAI to generate thousands of ideas: it’s a creativity superpower. But maybe the real superpower isn’t generating more, but knowing what to delete.The creator of anycrap.shop (where I created “glass toilet paper infused with microscopic shards”) reached out after the newsletter. His months-old satire project suddenly exploded after months of close-to-zero attention. Perfect example: even brilliantly executed projects struggle to break through the noise. Excellence alone isn’t enough anymore.Your noise audit question: What would your customers thank you for deleting?Sometimes the most radical innovation is subtraction, not addition.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek,Fresh back from Europe, I recorded this episode while processing a surreal experience: Russian drones made of plywood and polystyrene triggered NATO fighter jets over Poland. The airport I was flying to was closed during the event. The drones? They were decoys - $5,000 pieces of plywood triggering million-dollar defence responses.Here’s the pattern that should keep you up at night: You’re facing the same asymmetric exhaustion in your digital infrastructure.Recently, we’ve crossed a symbolic threshold. Automated traffic now exceeds human traffic on the internet (51% according to Imperva). One AI bot made 39,000 requests per minute to a single website. Meta’s crawlers alone generate 52% of all AI crawler traffic - yes, the same Meta whose AI glasses I wear to work, bypassing our corporate AI policies. We’re simultaneously victims and perpetrators of bot multiplication.A simple web form for reporting parking violations in Poland generated 2,000+ reports in days, overwhelming police operations. If organised humans with an efficient interface can break systems, what happens when everyone has AI agents?But should we treat all bot traffic like enemy drones? Or perhaps create new channels for bots? Grasshopper Bank shows the way - they created dedicated MCP channels for AI agents to access banking data directly, instead of having bots pretend to be humans. We need bot lanes, not bot wars.Monday morning action: Check your analytics for those telltale perfect traffic patterns. If your traffic curves are mathematically smooth, you’re looking at bots pretending to be human.Correction to last week’s episode: Michael (self-proclaimed biggest Nirvana fan) caught me - Nirvana’s MTV Unplugged was recorded in ’93, not ’94. I just looked at the Apple Music release date, and didn’t realise there was a delay between recording and releasing. Thanks for keeping me honest!Stay curious!Listen to the full episode for the complete story of NATO’s drone problem and why it’s your problem too. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it's Marek,Recording from Warsaw this week, where I just received the first printed copies of my book’s Polish edition (the translation will launch on 17/10/2025). There’s something magical about holding a physical book in your native language, and the process of translating it was quite interesting (humans + AI), but that’s a story for another day.This episode explores what happens when organisations try to put the AI genie back in the bottle. Spoiler: It doesn’t work.My friend at a hospital texted me: “We’re back to 1994.” His IT department had just downgraded them from Microsoft Copilot Pro to what he calls a “handcuffed chatbot.” The frustration in his message reminded me of Italy’s ChatGPT ban in March 2023, and the fascinating research that followed.In this podcast episode, I read the full story I wrote in the newsletter, plus share some additional insights:* Michael’s correction about when Nirvana’s MTV Unplugged was actually recorded (I was a year off!)* Anna’s brilliant Dave Grohl parallel: after Nirvana, he didn’t retreat, he created Foo Fighters* My own confession about wearing AI-enabled Meta Ray-Bans at work (where Meta AI isn’t an approved AI tool)Listen for the full story, including Samsung’s cautionary tale (three data leaks in 20 days), the Containment Paradox, and a simple Tuesday morning test to discover if you have an AI underground in your organisation.Stay curious! This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.I’m recording this from Prague, just before a public lecture on AI, in which I am going to make a claim that the Astronomical Clock was the OG digital minion.On the way to Czechia, ChatGPT offered me a sightseeing itinerary. 90% accurate, which sounds great until you imagine chaining it with a ticket-booking agent (also 90% accurate) and a reservation system (also 90% accurate).0.9 × 0.9 × 0.9 = 0.73Suddenly, my “smart” day in Prague is only 73% likely to work. Add a few more agents to handle transport, restaurant bookings, and museum timings, and I’d be wandering the Charles Bridge (or rather, sipping a Pilsner at Lokál) and wondering why nothing went to plan.So I called a friend instead. One human. 100% reliable at showing me his city. Plus, he took me for a running tour around town. Something that ChatGPT could never do!In this week’s audio edition, in which I read the most recent post of The Economy of Algorithms, I explore why AI agent chains fail where copilots succeed. It’s about compound failure: the evil twin of compound interest that nobody in Silicon Valley wants to discuss.The math is brutal: String together 10 agents at 90% reliability each, and you get 35% success. Suddenly, your AI workflow is not a tool but a lottery ticket.Listen for:* Why your occasionally high colleague is a perfect AI metaphor* The 10-90-35 rule that should terrify CTOs* Why copilots are thriving while autonomous agents collect dust* How to break the compound failure chain (hint: it involves actual humans)Stay curious,MarekP.S. The Czechs celebrate Jára Cimrman, patron saint of “almosts.” He missed the North Pole by seven meters and almost invented the Internet. Agent chains feel the same: impressive on paper, but when it counts, they stop just short. At least right now.P.P.S. The Pilsner at Lokál was glorious. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

Hey, it’s Marek.So, last Friday’s article (you can read it here, if you prefer text over audio) resonated with a lot of readers, it seems. This time its LinkedIn version generated a big discussion among readers.I needed this term in my life, “the Mullet Model”- perfect. A reader.I also received lots of positive feedback about the audio recordings in general (remember? last week was the very first time I shared an audio recording, a “directors cut” if you will), so I’ll continue it for a while. I am currently figuring out how to set up substack so that those of you who want to receive text posts, but not audio, can adjust their subscription. Please bear with me, I should have a solution next week. Don’t unsubscribe just yet. 😉Recording this week’s podcast, I kept thinking about that call with executives who announced their to-be-completed-in-2028 Copilot rollout. This wasn’t literary fiction! I really had that call, a few weeks ago. (I won’t reveal which company; they might be reading this newsletter. 🙈) To their credit, they immediately recognised that what they were doing wasn’t ideal. But there’s a difference between recognising insanity and deciding to do something more sane.The best article I have read in a long while about AI. Thank you for your wonderful way of mixing information, insights and entertainment. Another readerThat disconnect haunted me for days. Reports of 95% GenAI project failure rates alongside surging worker adoption made me genuinely curious: Are people seeing something organisations aren’t? That viral "MIT report" claiming 95% failure? MIT published it, then quietly took it down without explanation. Even research gets the mullet treatment: sharp findings up front, messy disappearance in the back. Still, other legitimate studies show similar struggles with demonstrating ROI (Return on Investment; yes, I explain that acronym in the recording because I felt guilty using it without definition in the original post).Calling AI a mullet captures that duality in a way that’s funny but also true. It’s both polished and scrappy, corporate and subcultural, respectable and slightly unhinged. Yet another reader.The paradox remains real: individual value is obvious, institutional ROI is elusive. We’re measuring the wrong things. Like trying to style a mullet with a calculator.Oh, and writing an entire article stuffed with mullet puns absolutely forced me to visit the barber this weekend. Even I have limits.Enjoy the recording.And stay curious. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com

I just spent eight minutes reading my newsletter aloud.Not for a podcast. Not for a presentation. Just to see what happens when you hear "The 80% Mistake" instead of reading it.Last week's post struck a nerve. More than a hundred shares, close to ten thousand reads… For me, these are amazing. Ten thousand people reading my words? Wow. Which made me curious: does the medium change the message?So here's my first audio experiment. Same insights about the Pareto Flip, same disaster stories, same take on AI curation. Just delivered in my accent at 8pm Brisbane time.This might be the start of something, or the end of a very short experiment. Your call.Hit reply and tell me. Or comment. Do you prefer reading or listening? Should I keep experimenting with audio, or stick to what works?Either way, your regular weekly insights return later this week.Stay curious,MarekP.S. If this audio thing takes off, I promise to consider adding some background sounds. Or whatever is important in podcasts. But let's see if anyone actually listens first. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit marekkowal.substack.com