Everyone says that Ethereum's Rollup-Centric strategy seems to have failed? And they deeply resent this L1-L2-L3 nesting doll game. But interestingly, the development of the AI track over the past year has also gone through a rapid evolution of L1—L2—L3. Comparing the two, where exactly does the problem lie?

1) The hierarchical logic of AI is that each layer addresses core issues that the upper layer cannot solve.

For example, L1's LLMs solve the foundational abilities of language understanding and generation, but logical reasoning and mathematical calculations are indeed hard shortcomings; thus, at L2, reasoning models specifically tackle this weakness. DeepSeek R1 can solve complex math problems and debug code, directly filling the cognitive blind spots of LLMs. After laying this groundwork, L3's AI Agent naturally integrates the capabilities of the first two layers, transforming AI from passive answering to active execution, allowing it to plan tasks, call tools, and handle complex workflows on its own.

You see, this kind of layering is “ability progression”: L1 lays the foundation, L2 addresses shortcomings, and L3 integrates. Each layer produces a qualitative leap based on the previous layer, and users can clearly feel that AI becomes smarter and more useful.

2) The hierarchical logic of Crypto is that each layer patches the problems of the previous layer, but unfortunately, this brings about brand new and larger problems.

For instance, if the performance of the L1 public chain is insufficient, it is natural to think of using layer 2's expansion solutions. However, after a wave of layer 2 Infra inflation, it seems that Gas has decreased, TPS has increased, but liquidity has become fragmented, and ecological applications continue to be scarce, making the excessive layer 2 infra a significant problem. Thus, they start creating layer 3 vertical application chains, but the application chains govern themselves and cannot enjoy the ecological synergies of the infra universal chain, resulting in an even more fragmented user experience.

As a result, this kind of layering has become “problem shifting”: L1 has bottlenecks, L2 patches, L3 is chaotic and dispersed. Each layer merely shifts the problem from one place to another, as if all solutions are merely aimed at “issuing tokens.”

At this point, everyone should understand what the crux of this paradox is: AI layering is driven by technological competition, with OpenAI, Anthropic, and DeepSeek all competing fiercely in model capabilities; Crypto layering is hijacked by Tokenomics, where the core KPI of each L2 is TVL and token price.

So, fundamentally one is solving technical challenges, while the other is packaging financial products? There may not be a clear answer as to which is right or wrong; it may vary by perspective.

Of course, this abstract analogy isn’t so absolute; it just seems that the comparison of the development paths of the two is very interesting, a little thought exercise for the weekend 💆.