Solution for the Problem
However, the challenges surrounding benchmark testing for models, as well as the apparent trade-offs between cost, throughput, and quality, make positive competition challenging. BitTensor is one of the largest cryptocurrencies focused on artificial intelligence in this category, aiming to address this issue, although it still faces some technical challenges that may hinder its widespread application (see Appendix 1).
Additionally, trustless model inference (i.e., proving that model outputs are actually generated by the claimed model) is another area of positive research in Crypto x AI. However, we believe that as the size of open-source models decreases, these solutions may face challenges in terms of demand.
In a world where models can be downloaded and run locally, and content integrity can be verified through established file hashing/verification methods, the importance of trustless inference is less clear. Indeed, many people still cannot train and execute LLMs on lightweight devices such as mobile phones, but powerful desktop computers (such as those used for high-end gaming) can already be used to execute many high-performance models.
Data Sources and Identity
As the output of generative AI becomes increasingly difficult to distinguish from human output, the importance of tracking AI-generated content has become a focus of concern. GPT-4 is three times faster than GPT-3.5 in passing the Turing Test, and we can almost certainly say that in the not-too-distant future, we will not be able to differentiate between online personalities created by machines and those created by real humans. In such a world, determining the humanity of online users and adding watermarks to AI-generated content will become crucial functions.
Decentralized identifiers and personality verification mechanisms like Worldcoin aim to solve the former problem of identifying humans on-chain. Similarly, releasing data hashes to the blockchain can help with verifying the age and source of content, thus empowering data sources. However, as with the previous section, we believe the feasibility of Crypto-based solutions must be weighed against centralized alternatives.
Additional research is also underway regarding AI watermarks to embed hidden signals in text and image outputs to allow algorithms to detect if content is generated by AI. Many leading AI companies, including Microsoft, Anthropic, and Amazon, have publicly committed to adding such watermarks to their generated content.
Furthermore, many existing content providers have strict records of metadata retention for compliance reasons. Therefore, users generally trust metadata associated with the release of social media (even if not their screenshots) despite being stored in a centralized manner.
It is worth noting that any Crypto-based data source and identity solution needs to be integrated with user platforms to be widely effective. Therefore, while Crypto-based solutions are technically feasible in proving identity and data sources, we also believe that their adoption is not a given and will ultimately depend on business, compliance, and regulatory requirements.
Trading AI Narratives
Despite the aforementioned issues, many AI tokens have outperformed Bitcoin, Ethereum, and major AI stocks like Nvidia and Microsoft since Q4 2023. We believe this is due to AI tokens benefiting from the broader Crypto market and the associated AI hype (see Appendix 2).
Therefore, even during Bitcoin price declines, AI-focused tokens experience price volatility that may be positive during Bitcoin downturns. Figure 5 shows the performance of AI tokens during days of Bitcoin trading decline.
Overall, we still believe that the AI narrative trading lacks many short-term sustainable demand drivers. The absence of clear adoption predictions and indicators has led to various meme-like speculative sentiments dominating the market, which in our view may not be sustainable in the long run.
Ultimately, price and utility will converge, and the unresolved question is how long it will take and whether utility will rise to match the price, or vice versa. That is, we do believe a sustainable constructive Crypto market and superior performance to the AI industry may maintain a strong Crypto AI narrative for a period of time.
Conclusion
The role of Crypto in AI does not exist in a vacuum, as any decentralized platform competes with existing centralized alternatives and must be analyzed against broader business and regulatory requirements. Therefore, we believe that simply replacing centralized providers with decentralized ones for the sake of “decentralization” is not enough to drive meaningful market adoption. Generative AI models have existed for several years and have maintained a degree of decentralization due to market competition and open-source software.
One recurring theme in this report is that while Crypto-based solutions are often technically feasible, they still require significant work to achieve functional parity with more centralized platforms, assuming these platforms do not stand still during this period. In fact, due to consensus mechanisms, centralized development often outpaces decentralization, which can pose challenges for a rapidly evolving field like AI.
Given this, we believe that the overlap between AI and Crypto is still in its early stages and may change rapidly in the coming years with the development of the broader AI landscape. The decentralized AI future envisioned by many Crypto insiders is currently uncertain, and the future of the AI industry itself is still largely undetermined. Therefore, we believe a cautious approach is wise in such a market and a deeper exploration of how Crypto-based solutions can truly provide meaningful and better alternatives, or at least an understanding of the underlying narrative, is necessary.
Appendix 1: BitTensor
BitTensor incentivizes different information markets within its 32 subnets. This aims to address some of the issues with benchmark testing by allowing subnet owners to create game-like constraints to extract information from information providers, thereby allowing the wisdom of the market participants to establish models in various markets. For example, its flagship subnet 1 is centered around text prompts and incentivizes miners who can generate the best textual response based on prompts sent by validators in that subnet. In other words, it rewards miners who can generate the best textual response to a given prompt, as judged by other validators in that subnet. This allows network participants to attempt to establish the wisdom of the market in various markets.
However, this validation and incentive mechanism is still in its early stages and is susceptible to adversarial attacks, especially if models use other biased models for evaluation (although progress has been made in this regard, using new synthetic data for evaluating certain subnets). This is particularly true for “fuzzy” outputs like language and art, where evaluation metrics may be subjective, leading to the emergence of multiple performance benchmarks for models.
For example, the validation mechanism for BitTensor’s subnet 1 currently requires:
Validators to generate one or more reference answers, and all miners’ responses are compared. Those most similar to the reference answers receive the highest rewards and ultimately the maximum incentives.
The current similarity algorithm uses a combination of string matching and semantic matching as the basis for rewards, but it is difficult to capture different style preferences through a limited set of reference answers.
It is currently unclear whether the models generated by BitTensor’s incentive structure can ultimately outperform centralized models (or if the best-performing models will shift to BitTensor), or how they can accommodate other trade-offs such as model size and underlying computational costs. A market that allows users to freely choose models that suit their preferences may be able to achieve a similar resource allocation through the “invisible hand.” In other words, BitTensor does attempt to address a challenging problem in an expanding problem space.
Appendix 2: WorldCoin
Perhaps the most prominent example of AI tokens following the AI market hype is Worldcoin. It released the World ID 2.0 upgrade on December 13, 2023, which went largely unnoticed, but it experienced a 50% surge after Sam Altman promoted Worldcoin on December 15.
Speculation about the future of Worldcoin remains speculative, partially due to Sam Altman being a co-founder of Tools for Humanity, the developer behind Worldcoin. Similarly, the release of Sora by OpenAI on February 15, 2024, led to nearly a three-fold increase in price, despite no related announcements on Worldcoin’s Twitter or blog (see Figure 6). As of the time of writing, Worldcoin has a valuation of $80 billion, which is close to OpenAI’s $86 billion valuation on February 16 (a company with annual revenue of $2 billion).
Extended Reading:
Worldcoin surged 150% in a week! Daily active users “exceed one million,” with a 10-fold growth in three months.
Some countries, such as China, have linked online personalities to government-controlled databases. While most regions in the world are not as centralized, a consortium of KYC providers can also offer identity verification solutions independent of blockchain technology (potentially in a similar way to trusted certificate authorities that constitute the foundation of today’s internet security).
Research is also underway regarding AI watermarks to embed hidden signals in text and image outputs to allow algorithms to detect if content is generated by AI. Many leading AI companies, including Microsoft, Anthropic, and Amazon, have publicly committed to adding such watermarks to their generated content.
Furthermore, many existing content providers have strict records of metadata retention for compliance reasons. Therefore, users generally trust metadata associated with the release of social media (even if not their screenshots) despite being stored in a centralized manner.
It should be noted that any Crypto-based data source and identity solution needs to be integrated with user platforms to be widely effective. Therefore, while Crypto-based solutions are technically feasible in proving identity and data sources, we also believe that their adoption is not a given and will ultimately depend on business, compliance, and regulatory requirements.
Trading AI Narratives
Despite the aforementioned issues, many AI tokens have outperformed Bitcoin, Ethereum, and major AI stocks like Nvidia and Microsoft since Q4 2023. We believe this is due to AI tokens benefiting from the broader Crypto market and the associated AI hype.
Therefore, even during Bitcoin price declines, AI-focused tokens experience price volatility that may be positive during Bitcoin downturns. Figure 5 shows the performance of AI tokens during days of Bitcoin trading decline.
Overall, we still believe that the AI narrative trading lacks many short-term sustainable demand drivers. The absence of clear adoption predictions and indicators has led to various meme-like speculative sentiments dominating the market, which in our view may not be sustainable in the long run.
Ultimately, price and utility will converge, and the unresolved question is how long it will take and whether utility will rise to match the price, or vice versa. That is, we do believe a sustainable constructive Crypto market and superior performance to the AI industry may maintain a strong Crypto AI narrative for a period of time.
Conclusion
The role of Crypto in AI does not exist in a vacuum, as any decentralized platform competes with existing centralized alternatives and must be analyzed against broader business and regulatory requirements. Therefore, we believe that simply replacing centralized providers with decentralized ones for the sake of “decentralization” is not enough to drive meaningful market adoption. Generative AI models have existed for several years and have maintained a degree of decentralization due to market competition and open-source software.
One recurring theme in this report is that while Crypto-based solutions are often technically feasible, they still require significant work to achieve functional parity with more centralized platforms, assuming these platforms do not stand still during this period. In fact, due to consensus mechanisms, centralized development often outpaces decentralization, which can pose challenges for a rapidly evolving field like AI.
Given this, we believe that the overlap between AI and Crypto is still in its early stages and may change rapidly in the coming years with the development of the broader AI landscape. The decentralized AI future envisioned by many Crypto insiders is currently uncertain, and the future of the AI industry itself is still largely undetermined. Therefore, we believe a cautious approach is wise in such a market and a deeper exploration of how Crypto-based solutions can truly provide meaningful and better alternatives, or at least an understanding of the underlying narrative, is necessary.