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!

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 ignored

Curation as service: Wirecutter tests 100 toasters to recommend one

Constraint as value: Jobs’ 2x2 grid took Apple from dozens of products to four. That discipline is now worth close to $4 trillion

The 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!

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.

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!

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 = 0.73

Suddenly, 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,Marek

P.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.

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 reader

That 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.

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,Marek

P.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.