I have been thinking a lot lately about how wild the last two years have been. Back in 2024, everyone in tech was acting like a wizard. People looked at the explosion of ChatGPT and generative AI and thought they could predict the exact future of humanity. Founders created massive five-year roadmaps assuming that AI would just keep getting bigger, smarter, and crazier every single week without ever slowing down.
But if the shift into 2026 has taught us anything, it is that thinking in a straight line is a total trap.
To build software that actually lasts, we have to stop trying to guess the future and start preparing our systems for reality. I keep coming back to a favourite quote from Howard Marks: “You Can’t Predict. You Can Prepare.” Technology and human habits do not just go up forever — they move in loops called cycles.
When you mix this cycle idea with Andrew Chen’s concepts from The Cold Start Problem, you get a perfect roadmap for building AI products today.
The AI Playground Swing
Trends in tech behave a lot like a swing on a playground. They move back and forth between two extremes because there is no single, perfect design. When a trend swings too far to one side, its problems start to show, and it naturally swings all the way back.
Look at what happened with AI between 2024 and 2026:
The App Split vs. The Big Bundle: In 2024, there were thousands of tiny “AI wrapper” apps — one for writing emails, one for making avatars, one for editing PDFs. It got messy fast. By 2025, the swing rushed backward. Big tech companies started bundling AI directly into systems we already use every day, like Apple Intelligence or Microsoft Copilot.
The Cloud vs. Your Phone: We started by running massive AI models on giant cloud servers because devices were too weak. Now the swing is moving toward on-device AI — small, fast models running locally on our actual phones and laptops without needing the internet.
Andrew Chen calls the hardest part of launching any new tool the Cold Start Problem. When an AI tool is new, it is useless if it doesn’t know anything about you or have any data loops. The biggest mistake a product team can make is panicking and launching a generic AI feature to millions of people at once. Instead, you have to build an Atomic Network — the smallest, tightest group of users or data points that makes the tool actually useful. Make the AI loop work perfectly for a tiny, core group first.
The Easy Money Hangover
Remember 2024, when investors threw millions of dollars at literally any company with “.ai” in its name? Risk completely disappeared. Marks puts it perfectly: “The worst loans are made at the best of times.” In the software world: the most useless features are built when companies have too much cash.
When money is easy to get, founders build without discipline. They bribe users with heavy discounts or flashy ads just to get downloads. We saw this with expensive AI hardware gadgets — pins and pocket devices that nobody actually needed. Cool toys. Bad products.
Chen warns that every network has a Hard Side — the small group of people who do the heavy lifting (creators, super-users, domain experts). If you use cheap investor money to ignore the hard side of your user base, your app will collapse the second the money stops. Real value comes from solving real problems, not buying fake hype.
The AI Ceiling
When a product gets popular, it hits Escape Velocity. Growth speeds up, users invite their friends, and it is easy to get giddy and assume exponential growth lasts forever.
But trees don’t grow to the sky.
AI has been hitting a massive structural ceiling lately:
The Data Wall: AI models are running out of high-quality human text to train on.
The Spam Loop: The internet is getting flooded with low-quality AI-generated synthetic text, which poisons the data pool for future models.
Marks highlights a study showing that out of 150 giant companies, almost none could grow continuously at high speeds for decades. The single company that won didn’t rely on a magical new model — it succeeded through solid, basic blocking and tackling in areas of stable consumer demand.
The secret to winning the AI race isn’t having a magical new model. It is doing the basics right — making things easy to use, highly reliable, and genuinely helpful.
The Action Plan
We can never predict exactly when a consumer trend will shift, but we can prepare.
Cold Start → First Stage: People are bored of basic chatbots; observers think AI is a passing fad. Stop building generic bots. Focus on solving one specific, hard workflow for a small group of users.
Escape Velocity → Third Stage: Everyone shouts “this time is different!” Money and copycat features fly everywhere. Resist the urge to add AI to every button. Fix the Hard Side of your user base so they stay forever.
Hitting the Ceiling → Inevitable Reversal: Models hit performance walls; users get tired of AI notifications. Pivot to fundamentals — data privacy, speed, and real user retention over hype.
The Drowning Man Warning
Marks tells a story about a six-foot-tall man who tragically drowned while trying to cross a river that was only five feet deep on average. The average hid the one critical hole where the riverbed dropped to eight feet.
In software, relying on averages and macro AI hype will get you ruined. Your total user sign-ups might look amazing on average, but if your core users are churning because the AI keeps hallucinating, your product is quietly sinking underneath the surface.
What the wise person does in the beginning, the fool does in the end. The builders who win this cycle will be the ones who stop chasing the loud, expensive fads, prepare for the inevitable slowdowns, and focus on real human utility.