Lately I have been thinking about whether AI can have good ideas.

Good ideas are different from writing error-free sentences, making useful suggestions, or producing PowerPoint packs. Good ideas are not just new. 99.9% of start-ups in the graveyard had novel-sounding ideas that mattered to exactly no one. Good ideas are not merely useful either, because useful ideas are often incremental in nature.

Good ideas are the sparkles we experience that lead to some of the proudest moments in personal and human history. Yang Chen-Ning and Lee Tsung-Dao’s moment of questioning whether conservation of parity may not hold. Watson and Crick’s guess that DNA may not be single-stranded or triple-stranded, but double-stranded.

These ideas are all that matters by scale.

So the question is not whether AI can generate ideas. Of course it can. The real question is: can it generate good ideas?

Before this question can be answered, we need to know where good ideas come from.

I would say that a good idea usually comes from one of the two things. It is either applying mental models from one field to another field, or it is a low-consensus, high-conviction bet about the future shaped by execution quality.

Charlie Munger is probably best known for taking a multidisciplinary approach and extracting key mental models from a wide range of domains. For example, he incorporated psychology in investing before behavioural finance was established as a field.

Low-consensus, high-conviction bets are the defining feature of generational companies and technologies. Interestingly, they often look obvious after the fact. These ideas ranges from allowing strangers to stay in your home to making rockets reusable. This is still happening today. For example, we don’t know what the next form factor for computing will be.

I can’t help but notice that these two sources of good ideas still rely heavily on humans. The former relies on human intelligence, and the latter relies on human execution. AI can produce many plausible and sometimes useful ideas, but those are still far away from truly good ideas. If ideas have a quality spectrum, the left and the middle parts of the spectrum are being hollowed out by AI, no doubt. But the ideas on the far right are still firmly gatekept by humans.

Some may challenge that it is a matter of time before AI reaches the stage where it can generate good ideas that no longer require human intelligence. I agree that day will eventually come, but it is still helpful to understand the structural limitations that AI, especially LLM-based AI, needs to overcome first.

The first problem LLMs face is computational complexity. Forming a bridge between different fields is computationally intensive for computers, because it is an O(n²) problem. Each field has millions of theories, and there are thousands of fields, so finding a meaningful bridge could easily mean trillions of computations.

How do humans do it then? The answer lies in abstraction.

This type of abstraction is sometimes dependent on a person’s unique background and depth of understanding. Some of the great takeaways may be vastly different from the consensus, and the abstraction can look nothing like the underlying texts. LLM tokenisation, on the other hand, focuses heavily on text compression and probabilistic prediction. As a result, it can over-represent low-value mainstream clichés.

If we agree that the bottleneck preventing AI from generating good ideas is abstraction, then it also means AI’s ideas can become good once the right abstraction is supplied. It is as if AI will dig 99% of the areas with no gold by default, and we get no gold. However, if we ask AI to dig in the 1% area where we know there could be gold, AI can do it effectively and in great depth.

An example is relying on AI to give someone product ideas. The result is often general ideas with little value. However, if we are specific about a product being a challenger to the mainstream product, and that counter-positioning is a weapon to use, AI can give specific advice around product design, channel choice, and milestones related to counter-positioning.

The human abstraction part becomes increasingly important when the information aggregation part is commoditised by LLMs. This is where it starts to resemble what good advice looks like for start-ups.

An example is the advice received by ReciMe founder Christine Nguyen in 2022 around solving the hard, supply side of a two-sided marketplace.

ReciMe has since become one of the best recipe apps in the world.

Eventually, AI should be able to form better abstractions and better bridges between fields. It may become much better at identifying which mental model applies to which problem. It may even become better at finding the 1% area where there could be gold.

When that happens, the value chain will increasingly move from the strategy layer to the execution layer.

This may sound depressing at first. If AI can eventually help with abstraction, strategy, and even taste, what is left for humans? But I actually find this future quite encouraging, because execution is not just the mechanical implementation of an idea; it is also where many of the best ideas are born.

A lot of valuable insight does not come before action. Before execution, many ideas are just stories. They may sound plausible, elegant, or even profound. But reality has the final say, as users do not behave the way we expect, distribution channels do not work the way they look from the outside, a product feature that feels minor can become the core wedge, and a detail that looks operational can become the strategic unlock.

This is the kind of insights AI cannot fully produce because it can’t experience the friction of building or anticipate market and customer behaviours ahead of time.

In fact, a world where execution becomes the main source of good ideas with decent rewards is a world worth looking forward to, because it is a world that belongs to the builders.