The intersection of generative artificial intelligence (AI) and Web3 is one of the most fascinating trends in the digital assets space. While most people agree that generative AI is likely to play a role in the next generation of Web3 technologies, the specifics are far from trivial. After all, AI was never considered an important building block in Web3 architectures, and the different generations of L1s and L2s were not designed to run AI workloads.

The reality confronted by Web3 technologists when trying to envision adapting Web3 runtimes to generative AI technologies is an overwhelming mismatch in terms of data and computation requirements. Generative AI workloads are designed to be computationally intensive, running on highly parallelizable GPUs. Blockchain runtimes are quite limited in terms of their data and computation capabilities.

Jesus Rodriguez is the CEO of IntoTheBlock.

At the same time, Web3 desperately needs to incorporate generative AI capabilities in order to catch up with Web 2 alternatives. The obvious question then becomes: how would the integration of generative AI and Web3 materialize?

From all the trends in generative AI, there is one that seems to be ideal for the incorporation of blockchain capabilities and that, coincidentally, gathered mainstream attention through the recent announcements made by OpenAI in its Developer Days conference: have you heard of autonomous agents?

Here, I would like to explore two key points:

We regularly read overhyped publications about how blockchain and generative AI technologies are a match made in heaven. Those statements are good for making headlines, but they lack the technical rigor of understanding the current state of both technologies. Going deep into understanding the possible integration paths between Web3 and generative AI reveals a very challenging picture.

Following a first-principles approach to think about the potential integration between generative AI and Web3, we can consider two fundamental dimensions:

The first vector is quite puzzling. On one hand, we can clearly imagine a new generation of DeFi protocols or blockchain runtimes that will incorporate intelligent capabilities powered by generative AI. On the other hand, those use cases are very nascent or almost non-existent.

The second category offers a wider set of opportunities but is equally challenging. Psychologically, it is very tempting to try to adapt modern Web3 capabilities such as zero-knowledge computation to decentralized stacks for generative AI solutions. It shouldn't come as a surprise that we are seeing an increasing number of zero-knowledge (zk) or decentralized stacks for generative AI. While certainly interesting, those solutions seem quite mismatched with the current wave of generative AI stacks.

Read more: Jesus Rodriguez - AI Should Be Decentralized, but How?

For instance, zero-knowledge computation could unlock some use cases in terms of transparency in generative AI, but its computational costs make it extremely impractical to be applied at the scale of large transformer models. Similarly, decentralized GPU architectures are impractical for pretraining or fine-tuning generative AI models, given that they require an incredibly fast communication bus between GPUs.

Bridging the gap between generative AI and Web3 technologies requires more than a match of technological capabilities; it requires that these capabilities are a fit for the current use cases in both technology stacks. In that sense, there is a hot trend in generative AI that seems like a perfect candidate for Web3 stacks.

If you follow the generative AI space, you probably have come across projects such as AutoGPT, BabyAGI, or the just-announced OpenAI GPTs, which are based on semi-autonomous agent capabilities. The simplest version of semi-autonomous agents is intelligent models that can reason through abstract tasks, formulate, and execute plans in a given environment.

Consider an agent that is specialized in executing market research for a specific domain. This agent can start with an abstract goal such as "research a specific product" and formulate a plan to consult the appropriate information sources, summarize the information using a large language model, and distribute it to different parties. Semi-autonomous agents enhance foundation models with capabilities such as memory, tool integration, security guardrails, and many others.

In the last few months, semi-autonomous agents have gone from an obscure research topic to one of the hottest trends in generative AI. The semi-autonomous nature of this technology is precisely what makes it an ideal scenario for blockchain runtimes. In the current state of semi-autonomous agent technologies, there are four dominant use cases that are a good fit for blockchain runtimes. These use cases are not theoretical but rather real challenges facing semi-autonomous agent applications:

1) Transparency: One of the main value propositions of semi-autonomous agents is their ability to formulate and execute actions in a given environment. Blockchain runtimes can act as a system of record for those plans and decisions.

2) Decentralized Coordination: One of the ultimate goals of semi-autonomous agents is that they can collaborate to accomplish a specific goal. This level of collaboration requires decentralized coordination, which is a perfect fit for blockchains. You can imagine a scenario in which one agent can discover the capabilities of other agents via their smart contracts and interact with them.

3) Guardrails: Once agents start taking actions semi-autonomously, the idea of setting up guardrails around them to constrain the potential impact of those actions becomes increasingly relevant. Smart contracts represent an ideal mechanism to establish immutable guardrails around agents. Imagine an agent that can make financial decisions given a credit application, guard-railed by a smart contract that ensures that no sensitive information is produced as part of the outputs.

4) Economic Incentives: In the context of collaboration, crypto assets can play a significant role, serving as a key economic layer for semi-autonomous agents. Let's take the example of an agent that can generate marketing materials for a specific product using expensive image generation models. In that scenario, the agents can receive payments via crypto assets and execute the function autonomously.

Regardless of our passion for Web3 tech, we can clearly see that generative AI is evolving just fine without the help of blockchain runtimes. Finding a match between these two technology trends is not a very easy endeavor. The Web3 scenarios for generative AI are very nascent or non-existent.

The point then becomes finding the real challenges in generative AI that can be addressed with blockchain technologies. By "real," I mean not derived from a theoretical exercise, and keeping in mind that we are trying to match a technology with limited adoption in the real world with the fastest-growing technology trend of many generations.

Semi-autonomous agents seem to have the right combination of technological and timing fit for blockchain runtimes. The trend is becoming widely accepted as one of the next waves in generative AI, and there are real challenges in terms of transparency, coordination, and security that seem well-suited for blockchains. Whether that integration materializes or not will require some clever technical adaptations in both technical movements. But the need is real; semi-autonomous agents could be the one trend that bridges generative AI and blockchains.