What is real and what is hype

The Web3 AI space is one of the hottest in the crypto world, combining great promise with significant hype. It seems almost heretical to point out the number of Web3 AI projects with multi-billion dollar market caps but no practical use cases, driven purely by proxy narratives from the traditional AI market. Meanwhile, the gap in AI capabilities between Web2 and Web3 continues to widen at an alarming rate. However, Web3 AI is not all hype. Recent developments in the generative AI market underscore the value of more decentralized approaches.

When we consider all these factors, we find ourselves in an overvalued and overfunded market that is disconnected from the state of the generative AI industry, but can still unlock tremendous value for the next wave of generative AI. It’s understandable to feel confused. When we distance ourselves from the hype and analyze the Web3 AI space through the lens of current needs, clear areas emerge where Web3 can deliver significant value. However, to do so requires fighting through a dense reality distortion field.

As crypto natives, we tend to see the value of decentralization in everything. However, AI has become an increasingly centralized force when it comes to data and computation, so the value proposition of decentralized AI must start by counteracting this natural centralizing force.

When it comes to AI, there is a growing disconnect between the value we create in Web3 and the needs of the AI ​​market. The worrying reality is that the gap between Web2 and Web3 AI is getting wider rather than narrower. This is largely due to three key factors:

Limited talent in AI research

The number of AI researchers working in Web3 is in the low single digits, which is hardly encouraging for those who claim that Web3 is the future of AI.

Limited infrastructure

We haven’t yet managed to get web apps to work properly with Web3 backends, so thinking about AI is a long shot, to say the least. Web3 infrastructure brings computational limitations that are impractical for the lifecycle of generative AI solutions.

Limited models, data and computing resources

Generative AI relies on three things: models, data, and compute power. None of the major pioneering models are designed to run on Web3 infrastructure; there is no foundation for large training datasets; and there is a huge quality gap between Web3 GPU clusters and those required for pre-training and fine-tuning basic models.

The harsh reality is that Web3 has developed a “poor man’s” version of AI that essentially tries to match the capabilities of Web2 AI but creates inferior versions. This reality is in stark contrast to the enormous value proposition of decentralization in several areas of AI.

To avoid this analysis becoming an abstract thesis, let us rather look at various decentralized AI trends and evaluate them in terms of their AI market potential.

The reality distortion in Web3 AI has led to the first wave of innovation and funding being focused on projects whose value propositions seem to have nothing to do with the realities of the AI ​​market. At the same time, there are other emerging areas in Web3 AI that hold enormous potential.

Decentralized GPU infrastructure for training and fine-tuning

In recent years, we have seen an explosion of decentralized GPU infrastructures designed to democratize pre-training and fine-tuning of base models. The idea is to create an alternative to GPU monopolization by established AI labs. The reality is that pre-training and fine-tuning large base models requires large GPU clusters connected by super-fast communication buses. A pre-training cycle of a 50-100B base model in a decentralized AI infrastructure could take over a year, if it works at all.

The idea of ​​combining zero-knowledge computation (ZK) and AI has given rise to interesting concepts to enable privacy mechanisms in base models. Given the importance of ZK infrastructure in Web3, several frameworks promise to embed ZK computation in base models. Although theoretically attractive, when applied to large models, ZK-AI models quickly encounter the challenge of being prohibitively expensive from a computational perspective. In addition, ZK will limit aspects such as interpretability, which is one of the most promising areas of generative AI.

Crypto is about cryptographic proofs, and sometimes they are attached to things that don’t need them. In the Web3 AI space, we see examples of frameworks that issue cryptographic proofs for certain model outputs. The challenges in these scenarios are not technological, but market-related. Essentially, proof-of-inference is something like a solution looking for a problem, and real use cases are lacking today.

Agent workflows are one of the most interesting trends in generative AI and hold significant potential for crypto. By agents, we mean AI programs that can not only passively answer questions based on inputs, but also perform actions in a given environment. While most autonomous agents are built for isolated use cases, we are witnessing the rapid emergence of multi-agent environments and collaboration.

This is an area where crypto can unlock tremendous value. For example, imagine a scenario where an agent needs to hire other agents to complete a task or stake value to vouch for the quality of their results. Providing agents with financial primitives in the form of crypto rails opens up many use cases for agent collaboration.

One of the most well-known secrets of generative AI is that the open-source AI space is suffering from a huge funding crisis. Most open-source AI labs can no longer afford to work on large models and are instead focusing on other areas that do not require massive amounts of computational access and data. Crypto is extremely efficient at capital formation with mechanisms such as airdrops, incentives, or even points. The concept of crypto funding rails for open-source generative AI is one of the most promising areas at the intersection of these two trends.

Last year, Microsoft coined the term “Small Language Model” (SLM) after releasing its Phi model, which, with less than 2 billion parameters, was able to outperform much larger LLMs in computer science and math tasks. Small base models – think 1 to 5 billion parameters – are a key enabler for the viability of decentralized AI and open up promising scenarios for on-device AI. Decentralizing models with several hundred billion parameters is nearly impossible today and will remain so for a while. However, small base models should be able to run on many of today’s Web3 infrastructures. Promoting the SLM agenda is critical to creating real value with Web3 and AI.

Generation of synthetic data

Data scarcity is one of the biggest challenges with this latest generation of base models, so research is increasingly focused on mechanisms to generate synthetic data using base models that can complement real-world datasets. The mechanisms of crypto networks and token incentives can ideally coordinate a large number of parties to collaborate in creating new synthetic datasets.

There are several other interesting Web3 AI trends with significant potential. Given the challenges of AI-generated content, proof-of-human output is becoming increasingly relevant. Evaluation and benchmarking is one AI segment where Web3’s trust and transparency features can shine. Human-centered fine-tuning, such as reinforcement learning with human feedback (RLHF), is also an interesting scenario for Web3 networks. As generative AI advances and Web3 AI features evolve, more scenarios are likely to emerge.

The need for more decentralized AI capabilities is very real. While the Web3 industry may not yet be able to compete with the value created by AI megamodels, it can unlock real value for the generative AI space. The biggest challenge for Web3 AI development may be overcoming its own reality distortion field. Web3 AI offers tons of value; we just need to focus on building real things.

Note: The views expressed in this column are those of the author and do not necessarily reflect the views of CoinDesk, Inc. or its owners and affiliates.