AI feels like the new kid in town: unusually gifted, loved by those at the top and watched by everyone else. The rules are often bent to help it succeed. New teams are created around it. Old processes are bypassed. Products are launched because they contain AI, not necessarily because anyone asked for them.

But the whispers have started. Perhaps the new kid has received more attention than it has earned. This raises a broader question. Should we redesign the environment around AI, or should AI first prove that it can work within the environment we already have?

In other words, should AI products be built top-down or bottom-up?

The discussion around AI has an almost unchallenged focus on creating more intelligence, based on the assumption that more intelligence means more value. This may be true on average, but it is not necessarily true for an individual user of AI. It reminds me of the old taunt often directed at scientists:

It is not just wealth, but also power. Today’s world is not ruled by its most intelligent people. If intelligence does not rule today’s world, why would we expect it to rule tomorrow’s?

As AI advances, fewer people can meaningfully feel the difference between frontier models and the models they already use. The improvements may be significant on benchmarks, but far less obvious in everyday life.

Interestingly, this means LLM development is becoming increasingly specialised. A smaller group of power users can detect the difference in reasoning quality, tool use or context handling across model releases, while most users cannot.

This is almost the opposite of how products usually evolve. Typical products begin in a niche and gradually expand into the mainstream. Frontier AI may be moving in the other direction: increasingly impressive to specialists, but increasingly difficult for ordinary users to distinguish. This is creating a mismatch between AI’s eyewatering capex growth and where its values accrue. To make AI investments sustainable, the application layer needs to deliver more values.

In the commercial world, AI arguably still has only a handful of proven use cases: coding assistance, document processing, customer support and workflow automation.

The bottleneck to commercial value is no longer simply the intelligence level of AI. It is the lack of sizeable and valuable applications that are only possible because of AI.

If someone is not good at their job and does not know what good looks like, throwing AI into the mix does not change much. The key still lies in finding a valuable problem and making something people want.

Suppose we have decided to build AI applications around the problems in front of us. The next question is whether to go top-down or bottom-up. The difference is scale. Should we first decide that an entire system needs to be replaced by AI, or should we build one useful AI product at a time?

Paul Graham recently summed this up well:

Startups rarely have the resources to pursue a grand goal from day one. Starting small is necessary as they have to earn the right to do the next thing.

As clichéd as it sounds, large corporates benefit enormously from a meritocracy of ideas as they have the opposite advantage. Scale, funding and distribution are assumed, and this makes top-down transformation seem more achievable. If anything, the top-down approach should work inside a large corporate. Once a good internal product is found, it can often reach millions of users quickly and translate into scalable impact almost immediately.

But big-bang launches fail in large organisations too. We have all seen or heard about major product launches that arrived with a blast and disappeared quietly over time. I would attribute this to some combination of complacency, a solution-in-search-of-a-problem mindset and low talent density: not enough people with the vision and taste required to see what the future product should become.

In addition, starting bottom-up also protects the organisation’s trust wallet. Every large AI launch asks users, executives and customers to believe that the organisation knows what it is doing. That trust is finite and can be overdrawn through repeated public missteps, as companies such as Meta and Baidu have shown.

Are we ruling out the top-down approach, then? Almost.

There are stories in the technology world that we never stop talking about because they sounded crazy at the time but proved essential in hindsight.

Amazon was already a low-margin business when Jeff Bezos began building out its own logistics capabilities, as if low margins were not challenging enough. Building the infrastructure layer of anything is extremely costly. Think about how much it costs to build a highway to the space. Yep, Bezos is on it again, with his company Blue Origin.

But in hindsight, without its supply chain excellence, Amazon would never have reached its current scale.

Similar decisions were made by Steve Jobs in controlling both the operating system and hardware for Apple devices, Elon Musk in bringing much of rocket development and manufacturing in-house, and Cursor in building its own IDE experience rather than remaining a plug-in for VS Code.

Few of these decisions required anything less than an all-in bet.

What made them even harder was that there was little meaningful derisking or experimentation available. You cannot properly test half an operating system, half a rocket company or half a new computing interface.

As random as it sounds, decisions like these tend to be made with intuition.

This is not to say that it is all guesswork or luck. I believe good taste is developed through deep work in a domain and ongoing, independent reflection on what good looks like.

Mathematically, most people can see the first-order impact of a decision. Far fewer can see the second- and third-order impacts. The people who can are often not just right, but early. It is almost as if they have a time machine that allows them to see the world two years from now.

Anyone can make a bet when they feel like doing so. But the ability to make great bets is not evenly distributed. It takes some combination of gift, deliberation and independence.

This is why grand attempts to rebuild entire systems so often fail. Crypto promised to rebuild modern banking from the ground up, but never truly succeeded in replacing it. A grand vision alone is not enough. The question is surprisingly relevant for AI today.

A truly AI-native product may require the entire system to be redesigned from first principles. It may not be possible to reach it through a series of small improvements to the existing workflow. But this approach only works when the founder/builder has unusually deep product intuition: enough domain knowledge, taste and independent thinking to see a future that the market cannot yet articulate.

What appears to be a top-down bet may actually be years of bottom-up learning compressed inside one person’s head. For everyone else, contact with users is still a better substitute for clairvoyance.

In the end, both top-down and bottom-up product building are still about making something people want. One comes from seeing the future. The other comes from iterating your way into it.