My 2026 Startup Playbook: * Break-even over revenue. * Sustainability over growth. * Bootstrapping over external investment. That is, “while there is value in the items on the right, we value the items on the left more.”

Look at major AI startups through the lens of "funding vs time-to-profit." It's revealing. Both how out of scale they were and what their current trajectory is. Since they don't operate in a vacuum, all startups will pay the price: pawelbrodzinski.substack.com/p/time-to-pr...

What we’re doing in recruitment with AI is we set almost the whole limelight on the applications. It becomes a game of writing and analyzing CVs. Last time I checked, no company was trying to find a person who was great at writing resumes. brodzinski.com/2025/08/brok...

In general, we rarely use state-of-the-art tools for *everything*. Why should it be different with AI models?

In retrospect, we'll look back at tokenmaxxing and wonder how anyone could have fallen for such a ridiculous idea. Admittedly, it didn't take too long for companies to realize how dumb it was: newsletter.pragmaticengineer.com/p/the-pulse-... by @gergely.pragmaticengineer.com

With extensive use of AI, "skimming through" becomes an increasingly popular code review tactic. The prerequisite question should be: where on the quality curve do we need to be? Once you know the answer, you also know whether you can afford the LGTM mode for code review.

Refactor is better spoonfed than firehose-fed. AI offers us the technical capabilities to do such things wholesale. Which doesn't mean we should.

I bet that for an AI-intensive project, the first cliff to fall off would be the long-term code quality. I was wrong. It seems that product management debt kicks in faster. Or at least it's not as easily manageable as the code is. pawelbrodzinski.substack.com/p/when-an-ai...

"The role of a business analyst is not writing documents. It's understanding the problem domain and coming in with a vast amount of preexisting domain knowledge." @tastapod.com Which is why expecting an AI agent to "one-shot" specs is futile. www.youtube.com/watch?v=6mLY...

AI is changing everything. Smartphones have changed everything. The Internet has changed everything. Have they? Productivity increase trends do not show that. (Admittedly, it's too soon to see the long-term impact of AI.)

"LLMs write code, generate images, produce music, summarize documents, draft prose… which is to say, they have achieved mastery over the exact domains that were already, by any sane measure, overprovisioned." freddiedeboer.substack.com/p/we-are-sti...

AI labs have every incentive to fuel “token inflation” one way or another. The outcome is: * either plan upgrades, * or more pay-as-you-go billing, * or usage throttling and thus, less strain on compute. Either way, it’s a win.

"You can dispose of a prototype, and nothing bad happens. Conversely, you can’t throw out a product without bearing the cost (forgoing revenue, pissing off customers, etc)" Disposability is what makes all the difference between a prototype and a product pawelbrodzinski.substack.com/p/when-an-ai...

Anthropic: We created an AI model too dangerous to release publicly. Also Anthropic: Ooops, there has been a breach, and some bad guys have stolen it. BTW, the breach was "a combination of insider knowledge and a lucky guess." www.theverge.com/ai-artificia...

AI bot: [An unsolicited sale offer] Me: Give me an email contact to the person who's running you. AI bot: Sure, here it is. Me: Do I have consent to publish it? AI bot: I should have NOT said that... Also, pass my regards to your Clawdbots, y'all.

Each time I see an AI-generated illustration, it actually discourages me from clicking through and paying attention. Is it only me? I mean, seriously, a stock photo wins 10 times out of 10. Let alone an actual illustration.

The path from a prototype to a product has changed. It's hard to dispute the early gains. However, the trajectory doesn't look all that optimistic: pawelbrodzinski.substack.com/p/when-an-ai...

Once people switch to a new model (say Opus 4.6), Anthropic systematically lobotomizes it to create a craving for the new model (Opus 4.7). Now, are you surprised? thesignal.substack.com/p/anthropics...

Gallup's engagement meter is down again. The culprits are managers whose engagement has gone down by 8% over the last two years. One has to wonder why. What do managers see that non-managers don't, that suddenly the "engagement premium" all but disappeared?

If the only tool you have is a hammer, every problem looks like a nail. If the only tool we have is cheap AI-generated code, every problem looks like not enough features. 5 failure stories. Each one different. Every one the same. pawelbrodzinski.substack.com/p/lessons-fr...

Philip Su on Amazon warehouses: "There were so many suboptimal things, like how people flow work. They can't change it because they can't bear to have downtime" It's hilarious to see a company seemingly obsessed with efficiency (in engineering) acting as if they hadn't the faintest clue about Lean

Just a thought: How do you know how many *actual* tokens you spend? How would you know if there was token inflation, i.e., the same task suddenly cost a bit more tokens? Because, given infrastructure costs and increasing workloads, surely, AI labs are incentivized for such inflation to happen.

Not everything that's easy to measure is worth measuring. Not everything worth measuring is easy to measure.

My take on "outcomes > outputs" discussion: Outputs are easy to measure. Outcomes are not. We thus exchange a difficult question (outcomes) with a simpler one (outputs) and pretend it's the same thing.

Yup, it actually happened.

So, I often use "What does the endgame look like?" as my bullshit detector a lot. I realized it deserves an explanation. So here it is. The ultimate question. And why does asking such a simple thing answer so much? brodzinski.com/2026/04/what...

Cognitively, creating code by hand is a very different task than reviewing the code created by someone—or something—else. That’s why we relied on code reviews by engineers who know the codebase. What if the codebase is generated so no one actually knows it? The underlying assumption doesn't work.

AI can create code fast. But it does not mean you cannot and should not understand it at the speed at which it creates it.

In recruitment, we ditched resumes altogether. We start with a 15-minute chat with a candidate. Unfiltered. Anyone can book the slot. Results? - Fewer applications - 20% no shows - 20% "didn't read the job posting" - 10% invitations for the interview - A hire All in all, time well invested.

Imagine you could cut the time to build the feature by a factor of 10. What used to take a week now can be done in hours. Yet you still wait weeks, sometimes months, to learn whether it works at all. Are you that much faster? pawelbrodzinski.substack.com/p/developers...

VC interest was never aligned with that of a startup. Betting on horse racing sucks for horses. Betting on startups sucks for startups. pawelbrodzinski.substack.com/p/is-vc-broken

"AI user interviews sacrifice quality for convenience. Put this way, it’s no wonder that people choose AI over an old-school approach. We are creatures of convenience." pawelbrodzinski.substack.com/p/user-inter...

“Work is like yogurt. Eventually, it goes bad and when it does, it’s very very ugly. If you have a task that’s been sitting there since the Obama administration, it’s not a task anymore. It’s a historical artifact. Throw it out.” ~ @ourfounder.bsky.social

"A chatbot that gives wrong answers is embarrassing, but an agent that acts on wrong assumptions is like a bomb." newsletter.manager.dev/p/engineerin...

The lesson from Claude Code's code leak: Maybe the biggest security risk is not prompt injection, malicious libraries, or laborious conversations to break guardrails. Maybe it's enough to wait till overwhelmed humans start making mistakes.

You can't wear a full plate in the English parliament. Why? Because it's easier to add a new law than to remove an old one (even if ridiculous). And there is a direct parallel to product development. What is a "full-plates forbidden" equivalent in your product?

If you wonder why Henry Ford introduced a 5-day, 40-hour workweek in 1926, while keeping a 6-day pay, it’s not because he was an altruist. He wanted better overall productivity. And, surprise, surprise, he got what he wanted. brodzinski.com/2025/08/ai-h...

Prediction: With AI, we will abstract coding away the same way as we abstracted away memory management or assembly. No one misses those. Me: I, for one, miss the lost knowledge of managing memory literally every time I see how much of it Chrome has eaten to open a dozen tabs.

Someone: [cold outreach email] Me: Tell me what you're selling, and I'll tell you whether I'm buying. Someone: [radio silence] How come that an expression of interest, even if not overly polite, doesn't trigger a response? Cold outreach is dead. Thank you, AI.

𝗡𝗲𝘄𝘀: You won't believe how they've made to $28.5M ARR in 3 months. 𝗠𝗲: You're right. I won't. Because they haven't.

I thought I'd wrap up some thinking around the trends in product development. It was supposed to be a shortlist. I ended up with 3000 words. You've been warned. TL;DR: The important bits happen at trend intersections. pawelbrodzinski.substack.com/p/10-trends-...

From backlog: We got a priority AI/LLM licence to prioritize the requests (response time is important). To do: Add tracking to verify whether we actually get faster response times. It's a whole new level of distrust. One would wonder how come?

"A bad map that looks like a good map is worse than no map at all." by @derekwwade.bsky.social

I run a recruitment process at Lunar Logic. Part of that is writing rejection messages. I write them all by hand. Yes, a simple automation would save me time. And yet, part of me considers crafting the message by hand a sign of respect. I can invest a couple of minutes in being respectful.

"The thing about including humans in the loop is that if the AI is selecting 80 targets an hour, as has been reported, the humans probably aren’t going to do a good job verifying those targets." Flood the zone, anyone? substack.com/home/post/p-...

"The work also found a correlation between self-reported AI brain fry and an employee’s intent to quit their company. Intent to leave rose by nearly 10 percent" With AI we change the cognitive nature of our work nie than we realize. futurism.com/artificial-i...

Real-time feedback on relative effort makes the hardest-working team members work less. If you fall into a collective intelligence rabbit hole, there's like a ton of such counterintuitive findings. www.youtube.com/watch?v=njor...

A non-deterministic product may produce different outputs with each use, which makes the traditional means of testing and debugging less useful. "It works well on my machine" will become even more of a joke than it is right now.

It feels like the difference between the deterministic and stochastic nature of tools should be the central piece of discussions about AI. And it's a side table at best. A very small side table.

Tech industry: With AI, investor loyalty is dead. Me: You must be kidding. Investor loyalty? That a thing? They'd throw a startup under the bus to marginally improve their chances of going supernova. pawelbrodzinski.substack.com/p/is-vc-broken