Author: David & Goliath
Compiled by: Shenchao TechFlow
Currently, the computing and training aspects of the AI industry are primarily dominated by centralized Web2 giants. These companies hold a dominant position due to their strong capital, cutting-edge hardware, and vast data resources. While this situation may persist in developing the most powerful general machine learning (ML) models, for mid-tier or customized models, the Web3 network may gradually become a more economical and accessible source of computing resources.
Similarly, when inference demands exceed the capabilities of personal edge devices, some consumers may choose the Web3 network to obtain less censored and more diverse outputs. Instead of attempting to completely disrupt the entire AI technology stack, participants in Web3 should focus on these niche scenarios and fully leverage their unique advantages in censorship resistance, transparency, and social verifiability.
The hardware resources needed to train the next-generation foundational models (such as GPT or BERT) are scarce and expensive, with demand for the highest performance chips continuing to outstrip supply. This scarcity of resources leads to hardware being concentrated in the hands of a few well-funded leading companies, which use this hardware to train and commercialize the most performant and complex foundational models.
However, the pace of hardware updates is extremely fast. So, how will outdated mid-tier or low-performance hardware be utilized?
These hardware resources are likely to be used for training simpler or more targeted models. By matching different categories of models with hardware of varying performance, optimal resource allocation can be achieved. In this case, Web3 protocols can play a key role by coordinating access to diverse and low-cost computing resources. For instance, consumers can use simple mid-tier models trained on personal datasets, and only opt for high-end models trained and hosted by centralized companies when handling more complex tasks, while ensuring user identities are hidden and prompt data is encrypted.
Aside from efficiency issues, concerns about bias and potential censorship in centralized models are also growing. The Web3 environment is known for its transparency and verifiability, capable of providing training support for models that have been overlooked or deemed too sensitive by Web2. Although these models may not be competitive in performance and innovation, they still hold significant value for certain groups in society. Therefore, Web3 protocols can carve out a unique market in this area by offering more open, trustworthy, and censorship-resistant model training services.
Initially, both centralized and decentralized approaches could coexist, each serving different use cases. However, as Web3 continually improves developer experience and platform compatibility, and as the network effects of open-source AI gradually become evident, Web3 may ultimately compete in core areas dominated by centralized companies. Especially as consumers become increasingly aware of the limitations of centralized models, the advantages of Web3 will become more pronounced.
In addition to training mid-tier or niche models, participants in Web3 also have the advantage of providing more transparent and flexible inference solutions. Decentralized inference services can bring various benefits, such as zero downtime, modular model combinations, public model performance evaluations, and more diverse, uncensored outputs. These services can also effectively avoid the 'vendor lock-in' problem faced by consumers who rely on a few centralized providers. Similar to model training, the competitive advantage of decentralized inference layers does not lie in raw computational power itself, but in addressing long-standing issues, such as the transparency of closed-source fine-tuning parameters, lack of verifiability, and high costs.
Dan Olshansky proposed a promising vision of creating more opportunities for AI researchers and engineers through POKT's AI inference routing network, allowing them to realize their research outcomes and earn additional income through customized machine learning (ML) or artificial intelligence (AI) models. More importantly, this network can facilitate fairer competition in the inference service market by integrating inference results from various sources, including decentralized and centralized providers.
Although optimistic predictions suggest that the entire AI technology stack may fully migrate on-chain in the future, this goal still faces significant challenges from the current concentration of data and computing resources, as these resources provide substantial competitive advantages to existing giants. However, decentralized coordination and computing networks demonstrate unique value in providing more personalized, economical, and open competitive AI services that are resistant to censorship. By focusing on these critically valuable niche markets, Web3 can establish its own competitive barriers, ensuring that the most influential technologies of this era can evolve collaboratively in multiple directions, benefiting a broader range of stakeholders rather than being monopolized by a few traditional giants.
Finally, I would like to especially thank all members of the Placeholder Investment team, as well as Kyle Samani from Multicoin Capital, Anand Iyer from Canonical VC, Keccak Wong from Nectar AI, Alpin Yukseloglu from Osmosis Labs, and Cameron Dennis from NEAR Foundation, who provided reviews and valuable feedback during the writing of this article.