Analyzing the intersection of encryption and AI, discussing the real challenges and opportunities, and distinguishing between empty promises and practical solutions.

Author: Ed Roman, Managing Partner at Hack VC

Translator: 1912212.eth, Foresight News

AI+ Crypto is one of the most prominent frontiers in the cryptocurrency market, such as decentralized AI training, GPU DePINs, and censorship-resistant AI models.

Behind these dazzling advancements, we can't help but ask: is this a true technological breakthrough or just riding the trend? This article will unveil the mysteries, analyze the encryption x AI concept, discuss the real challenges and opportunities, and reveal which promises are empty and which are practical.

Concept #1: Decentralized AI Training

The challenge of on-chain AI training lies in the need for high-speed communication and coordination between GPUs, as neural networks require backpropagation during training. Nvidia has two innovations for this (NVLink and InfiniBand). These technologies enable ultra-fast GPU communication, but they are limited to local technology, only applicable to GPU clusters within a single data center (50+ gigabits per second speed).

Introducing a decentralized network would significantly slow down the speed by several orders of magnitude due to network latency and increased bandwidth, compared to the throughput obtained from high-speed interconnects within data centers by Nvidia. This speed is impossible for AI training use cases.

Note that there are innovations that may bring hope for the future:

• Large-scale distributed training on InfiniBand, as NVIDIA is currently supporting distributed non-local training on InfiniBand through the NVIDIA Collective Communications Library. However, it is still in its early stages, so the adoption metrics are yet to be determined. Physical law bottlenecks still exist, so local training on InfiniBand remains much faster.

• Some newly published research on decentralized training, which reduces communication synchronization time, may make decentralized training more practical in the future.

• Intelligent sharding and scheduling of model training contribute to performance improvement. Similarly, new model architectures may be specifically designed for future distributed infrastructure (Gensyn is researching in these areas).

The data part of training is also challenging. Any AI training process involves handling large amounts of data. Typically, models are trained on a centralized secure data storage system with high scalability and performance. This requires the transmission and processing of terabytes of data, and this is not a one-time cycle. Data is often noisy and contains errors, so it must be cleaned and transformed into usable formats before training the model. This stage involves repetitive tasks of standardization, filtering, and handling missing values, all of which face serious challenges in a decentralized environment.

The data part of training is also iterative, which is not compatible with Web3. OpenAI went through thousands of iterations before achieving its results. In AI teams, the basic task scenarios for data scientists include defining goals, preparing data, analyzing and organizing data to extract important insights, and making it suitable for modeling. Then, develop machine learning models to solve the defined problem and validate their performance using a test dataset. This process is iterative: if the current model performs below expectations, experts return to the data collection or model training stage to improve the results. Imagine how difficult it would be to adapt the most advanced existing frameworks and tools to Web3 for this process in a decentralized environment.

Another problem with training AI models on-chain is that, compared to inference, this market is much less interesting. Currently, training large language models for AI requires a large amount of GPU computing resources. In the long run, inference will become the primary application scenario for GPUs. Imagine how many large language models need to be trained to meet global demand, compared to the number of customers using these models, which is more?

Concept #2: Achieving Consensus through Redundant AI Inference Computing

Another challenge of encryption and AI is verifying the accuracy of AI inference, as you cannot fully trust a single centralized party to execute inference operations, and there is a potential risk of improper behavior by nodes. This challenge does not exist in Web2 AI because there is no decentralized consensus system.

The solution is redundant computing, allowing multiple nodes to repeat the same AI inference operations, which can run in a trustless environment, avoiding single points of failure.

However, the problem with this approach is the extreme shortage of high-end AI chips. The long waiting time for high-end NVIDIA chips, which can be several years, has led to price increases. If you require AI inference to be repeatedly executed on multiple nodes, the high cost will increase exponentially, making it unfeasible for many projects.

Concept #3: Near-term Web3-specific AI Use Cases

Some suggest that Web3 should have its own unique AI use cases specifically tailored to Web3 clients. This could include (for example) Web3 protocols using AI to score the risk of DeFi pools, Web3 wallets advising users on new protocols based on wallet history, or Web3 games using AI to control non-player characters (NPCs).

Currently, this is a nascent market, where use cases are still in the exploratory stage in the short term. Some challenges include:

• Due to the nascent market demand, there are relatively few potential AI transactions required for Web3 native use cases.

• Fewer clients, with Web3 clients being several orders of magnitude fewer than Web2 clients, resulting in lower decentralization of the market.

• Clients themselves are less stable, as they are early-stage companies with limited funds, and some startups may disappear over time. Web3 AI service providers catering to Web3 clients may need to reacquire some customer base to replace those that have disappeared, making business expansion extremely challenging.

In the long run, we are very optimistic about Web3-native AI use cases, especially as AI agents become more common. We imagine that any specific Web3 user in the future will have a large number of AI agents to help them accomplish tasks.

Concept #4: Consumer-grade GPU DePIN

There are many decentralized AI computing networks that rely on consumer-grade GPUs rather than data centers. Consumer-grade GPUs are well suited for low-end AI inference tasks or flexible consumer use cases with latency, throughput, and reliability. However, for serious enterprise use cases (which are the majority of the important market), customers need a more reliable network than home machines, and if they have more complex inference tasks, they typically require higher-end GPUs. Data centers are more suitable for these more valuable customer use cases than home machines.

Note that we believe consumer-grade GPUs are suitable for demonstrations and for personal and startup companies that can tolerate lower reliability. However, these customers have lower value, so we believe that DePINs customized for Web2 enterprises will be more valuable in the long run. Therefore, GPU DePIN projects have evolved from early use of consumer-grade hardware to situations with A100/H100 and cluster-level availability.

Reality - Practical Use Cases of Cryptocurrency x AI

Now let's discuss the use cases that can provide real benefits. These are the true victories, where cryptocurrency x AI can add significant value.

Real Benefit #1: Serving Web2 Clients

McKinsey estimates that among the 63 use cases analyzed, generative AI can increase annual revenue by an amount equivalent to $26 trillion to $44 trillion—compared to the total GDP of the UK in 2021, which was $31 trillion. This would increase the impact of AI by 15% to 40%. If we consider the impact of embedding generative AI into other task software currently used for use cases outside of it, the estimated impact would roughly double.

If you calculate based on the above estimates, it means the total market value of global AI (beyond generative AI) could reach tens of trillions of dollars. In comparison, the total value of all cryptocurrencies today (including Bitcoin and all altcoins) is only around $27 trillion. Therefore, let's face the reality: the vast majority of customers needing AI in the short term will be Web2 clients, as the Web3 clients who truly need AI will only represent a small portion of this $27 trillion (considering that BTC is part of this market, Bitcoin itself does not require/use AI).

Web3 AI use cases are just beginning, and it is currently unclear how large this market will be. But one thing is certain—the foreseeable future will only represent a small portion of the Web2 market. We believe Web3 AI still has a bright future, but this only means that the most powerful applications of Web3 AI currently serve Web2 clients.

Examples of Web2 clients who could benefit from Web3 AI include:

• Building vertically specific software companies centered around AI from scratch (e.g., Cedar.ai or Observe.ai)
• Large enterprises fine-tuning models for their own purposes (e.g., Netflix)
• Rapidly growing AI providers (e.g., Anthropic)
• Software companies integrating AI into existing products (e.g., Canva)

These are relatively stable client roles, as they are typically large and valuable. They are less likely to go out of business quickly and represent a huge potential customer base for AI services. Web3 AI services serving Web2 clients will benefit from these stable customer bases.

But why would Web2 clients want to use the Web3 stack? The next part of this article explains this situation.

Real Benefit #2: Lowering GPU Usage Costs through GPU DePIN

GPU DePIN aggregates underutilized GPU computing power (the most reliable of which comes from data centers) and makes it available for AI inference. A simple analogy for this problem is "Airbnb for GPUs."

We are excited about GPU DePIN because, as mentioned above, there is a shortage of NVIDIA chips, and there is currently wasted GPU cycles available for AI inference. These hardware owners have sunk costs and are not fully utilizing their devices, so these partial GPUs can be offered at much lower costs compared to the current situation, effectively "finding money" for the hardware owners.

Examples include:

• AWS machines. If you want to rent an H100 from AWS today, you have to commit to a 1-year lease because of limited market supply. This results in waste because you may not use the GPU every day of the year, every week of the year.
• Filecoin mining hardware. Filecoin has a large subsidy supply but not a large actual demand. Filecoin has never found a true product-market fit, so Filecoin miners face the risk of going out of business. These machines are equipped with GPUs that can be repurposed for low-end AI inference tasks.
• ETH mining hardware. As Ethereum transitions from PoW to PoS, this quickly releases a large amount of hardware that can be repurposed for AI inference.

Note that not all GPU hardware is suitable for AI inference. One obvious reason for this is that older GPUs do not have the required GPU memory for LLMs, although there are some interesting innovations that can help in this regard. For example, Exabits' technology can load active neurons into GPU memory and inactive neurons into CPU memory. They predict which neurons need to be active/inactive. This allows low-end GPUs to handle AI workloads, even with limited GPU memory. This effectively makes low-end GPUs more useful for AI inference.

Web3 AI DePINs need to develop their products over time and provide enterprise-level services, such as single sign-on, SOC 2 compliance, service-level agreements (SLAs), and more. This is similar to the services currently provided by cloud service providers to Web2 clients.

Real Benefit #3: Censorship-resistant models to avoid OpenAI self-censorship

There is much discussion about AI censorship systems. For example, Turkey temporarily banned OpenAI (which OpenAI later raised compliance, and they changed their approach). We believe that national-level censorship systems are uninteresting because countries need to adopt AI to remain competitive.

OpenAI also engages in self-censorship. For example, OpenAI does not handle NSFW content. OpenAI also does not predict the next presidential election. We believe that AI use cases are not only interesting but also have a huge market, but OpenAI will not touch this market for political reasons.

Open source is a good solution because GitHub repositories are not influenced by shareholders or boards. Venice.ai is an example that promises privacy protection and operates in a censorship-resistant manner. Web3 AI can effectively elevate its level by supporting these open-source software (OSS) models on low-cost GPU clusters to perform inference. For these reasons, we believe OSS + Web3 is an ideal combination to pave the way for censorship-resistant AI.

Real Benefit #4: Avoiding sending personal identity information to OpenAI

Large enterprises have privacy concerns about their internal data. For these clients, it may be difficult to trust third-party OpenAI with this data.

In Web3, for these enterprises, their internal data suddenly appearing on a decentralized network may seem more concerning (on the surface). However, there are innovations in privacy-enhancing technologies for AI:

Trusted Execution Environments (TEE), such as Super Protocol Fully Homomorphic Encryption (FHE), such as Fhenix.io (backed by a fund in Hack VC's portfolio) or Inco Network (both supported by Zama.ai), and Bagel's PPML These technologies are still evolving, and performance is still improving through upcoming zero-knowledge (ZK) and FHE ASICs. But the long-term goal is to protect enterprise data when fine-tuning models. With the emergence of these protocols, Web3 may become a more attractive place for privacy-protecting AI computation.

Real Benefit #5: Leveraging the latest innovations of open-source models

For the past few decades, open-source software has been eating into the market share of proprietary software. We view LLM as a form of proprietary software, enough to disrupt OSS. Notable challengers include Llama, RWKV, and Mistral.ai. This list will undoubtedly continue to grow over time (a more comprehensive list can be found on Openrouter.ai). By leveraging Web3 AI (supported by OSS models), people can harness these new innovations for innovation.

We believe that over time, the global development community of open source, combined with cryptocurrency incentives, can drive rapid innovation of open-source models and the agents and frameworks built on top of them. An example of an AI agent protocol is Theoriq. Theoriq creates a composable interconnected network of AI agents using OSS models to assemble them to create higher-level AI solutions.

We are confident in this because over time, most "developer software" innovations have been slowly surpassed by OSS. Microsoft used to be a proprietary software company, and now they are the top contributor to Github. There's a reason for this, if you look at how Databricks, PostgreSQL, MongoDB, and other companies have disrupted proprietary databases, it's a compelling example of OSS disrupting an entire industry, so the precedent here is very persuasive.

However, there is a catch. One of the challenges of open-source large language models (OSS LLMs) is that OpenAI has started signing paid data licensing agreements with some organizations (such as Reddit and The New York Times). If this trend continues, open-source large language models may become more difficult to compete due to financial barriers to accessing data. Nvidia may further strengthen its investment in secure computing to facilitate secure data sharing. Time will reveal the developments in this.

Real Benefit #6: Achieving consensus through high-cost reduction random sampling or ZK proofs

One of the challenges of Web3 AI inference is verification. Assuming validators have the opportunity to deceive their results to earn fees, verifying inferences is an important measure. It's worth noting that this cheating behavior has not actually occurred yet, as AI inference is still in its early stages, but it is inevitable unless measures are taken to suppress this behavior.

The standard Web3 method is to have multiple validators repeat the same operation and compare the results. As mentioned earlier, the prominent challenge facing this is that the cost of AI inference is very expensive due to the current shortage of high-end Nvidia chips. Given that Web3 can provide lower-cost inference through underutilized GPU DePIN, redundant computation will severely weaken the value proposition of Web3.

A more promising solution is to execute ZK proofs for off-chain AI inference computation. In this case, concise ZK proofs can be verified to determine if the model has been trained correctly or if the inference is running correctly (referred to as zkML). Examples include Modulus Labs and ZKonduit. Since ZK operations are computationally intensive, the performance of these solutions is still in its early stages. However, we expect improvement with the release of ZK hardware ASICs in the near future.

Another promising concept is an "Optimistic" sampling-based AI inference method. In this model, only a small portion of the results generated by validators needs to be verified, but the economic cost of being significantly reduced is set high enough so that if discovered, it will have a strong economic deterrent effect on validators cheating. This way, redundant computation can be saved.

Another promising concept is watermark and fingerprint solutions, such as the solution proposed by Bagel Network. This is similar to Amazon Alexa providing device-internal AI model quality assurance for its millions of devices.

Real Benefit #7: Cost savings through OSS (OpenAI's profit)

The next opportunity Web3 brings to AI is cost democratization. So far, we have discussed saving GPU costs through DePIN. But Web3 also offers the opportunity to save the profit margins of centralized Web2 AI services (such as OpenAI, which had annual revenue exceeding $1 billion at the time of writing). These cost savings come from the fact that using OSS models instead of proprietary models achieves additional savings because the model creators are not trying to profit.

Many OSS models will remain completely free, providing the best economic benefits for customers. However, some OSS models may also be attempting these monetization methods. Consider that only 4% of all models on Hugging Face are trained by companies with budgets to help subsidize the models. The remaining 96% of models are trained by the community. This group (96% of Hugging Face) has basic actual costs (including computational and data costs). Therefore, these models will need to be monetized in some way.

There are some proposals to monetize open-source software models. One of the most interesting is the concept of "initial model issuance," which involves tokenizing the model itself, reserving a portion of the tokens for the team, and directing some of the model's future income streams to token holders, although there are certainly legal and regulatory obstacles in this regard.

Other OSS models will attempt to monetize through usage. Note that if this becomes a reality, OSS models may start to become more and more similar to their Web2 profit models. But in reality, the market will be divided into two parts, with some models still completely free.

Real Benefit #8: Decentralized data sources

One of the biggest challenges AI faces is finding the right data to train models. We previously mentioned the challenges of decentralized AI training. But how about using decentralized networks to obtain data (which can then be used for training elsewhere, even in traditional Web2 settings)?

This is exactly what startups like Grass are doing. Grass is a decentralized network of "data fetchers" who contribute idle processing power of machines to data sources for training AI models. Assuming, in scale, due to the powerful force of large incentive node networks, this data source can work better than any company's internal data source. This not only includes obtaining more data but also obtaining data more frequently to make the data more relevant and up-to-date. In fact, preventing the decentralized data fetching army is also impossible, as they are inherently decentralized and do not reside within a single IP address. They also have a network to clean and standardize data so that the data is useful after being fetched.

After obtaining the data, you also need a place to store it on-chain, and LLMs generated using that data.

Note that the role of data in Web3 AI may change in the future. Today, the status of LLMs is to use pre-trained models with data and refine them over time with more data. However, because data on the internet is constantly changing, these models are always slightly outdated. Therefore, the response of LLM inference is slightly inaccurate.

A future development direction may be a new paradigm— "real-time" data. The concept is that when a large language model (LLM) is asked to infer a question, the LLM can transmit and inject data through prompts that are collected in real-time from the internet. This way, the LLM can use the latest data. Grass is researching this part.

Special thanks to the following individuals for their feedback and assistance with this article: Albert Castellana, Jasper Zhang, Vassilis Tziokas, Bidhan Roy, Rezo, Vincent Weisser, Shashank Yadav, Ali Husain, Nukri Basharuli, Emad Mostaque, David Minarsch, Tommy Shaughnessy, Michael Heinrich, Keccak Wong, Marc Weinstein, Phillip Bonello, Jeff Amico, Ejaaz Ahamadeen, Evan Feng, JW Wang.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。