Crypto VC Canonical founder: Will AI Agents create a dystopian future?

Translator's note: In this episode of the Chain Abstraction podcast, Altan, co-founder of the NAER ecosystem, and Jared, head of content at the NAER Foundation, had an in-depth conversation with Anand Iyer, founder of early-stage fund Canonical. Canonical has invested in well-known crypto AI projects such as Gensyn, Ritual, Sahara, Nuffle Labs, and Tensorplex Labs. In addition, Anand has one of the best X/twitter handles: https://x.com/ai.

Anand pointed out that in the future, AI agents will play a key role in assisting humans to complete various tasks, and blockchain provides a secure and reliable operating environment for these agents. In the era of rapid development of AI, user ownership and privacy protection of personal data are crucial. As an investor, he values projects with differentiation and defensive capabilities, and encourages entrepreneurs to have a long-term vision and promote innovation in technology and applications.

Application Scenarios of Chain Abstraction in Crypto AI

Jared: First, please briefly introduce your background.

Anand:I have a degree in computer engineering from Purdue University and have worked for tech giants such as Microsoft. I started programming at a very young age, using a programming language called BASIC. I was only eight years old at that time, which was already in the 80s. I spent a lot of time developing Microsoft's platform, and moved to Silicon Valley after graduating from college in 2001, witnessing the development and growth of Silicon Valley in the 90s and 2000s.

After that, I had the idea of starting a business. I worked in large companies for the first ten years, and I was a startup founder and CEO for the next ten years. Long story short, I came into contact with a group of Bitcoin miners by chance in 2012. It was through them that I entered the world of Bitcoin and cryptocurrency. The crypto community was still in its infancy at that time, like a small group of technology enthusiasts, and I felt that the atmosphere here was a good fit for me. My company experienced ups and downs and was eventually successfully acquired. Now I have decided to devote myself to the cryptocurrency field full-time, managing an early-stage fund called Canonical, focusing on investing in outstanding startups and founders. Recently, I have paid special attention to the intersection of artificial intelligence and cryptocurrency.

Altan: You are very interested in the intersection of cryptocurrency and AI. Can you talk about the importance of chain abstraction in this area?

Anand: I like to think in analogies. One constant learning is that history doesn't repeat itself, but it rhymes. Chain abstraction is a topic that has been discussed a lot recently, and I think it is necessary. If you look at the history of blockchain, you can understand why chain abstraction is so important today.

For me, the biggest breakthroughs in the Internet in the past were things like Kubernetes. It served as an abstraction layer to help deploy, manage, and scale applications across different cloud platforms. This is a bit similar to what is happening now. Developers and users need chain abstraction because they may want to use features on one chain and use accounts on another chain. In this way, abstraction becomes critical. I think there may be a few different entry points in AI: one is the developer's perspective and the other is the user's perspective.

For developers, it's whether developers need to learn the programming paradigm of a specific chain, such as WASM and Rust. What role can chain abstraction play in this regard? Or, can smart contracts be deployed more efficiently using LLM-driven tools? This is one way of thinking. Another is user interaction, which can also extend to agent interaction. I think the speed of user adoption of blockchain is increasing significantly. Although we all know that there is a certain degree of robot and farm operation, I think we should go with the flow instead of resisting this robotization. Because this is the inherent feature provided by blockchain, and this is where agent platforms come into play. They need to interact on behalf of users and ultimately take advantage of the functions provided by different chains for themselves. Therefore, I think chain abstraction becomes more important.

We are transitioning from interoperability to chain abstraction so that users, developers, users and agents can easily and concisely take advantage of the functions of various chains. In general, I think there are two directions of thinking. First, who is the user? Is it an agent? Is it an AI robot that needs to use these platforms? Second, how can we make these tools more easy to use and customizable? This is another need worth exploring.

Integration of AI Agent and Blockchain

Jared: I noticed that you also consider AI Agent as a user, which is very interesting. Blockchain is a good source of truth. Do this immutable record and the infinite creativity of AI need to work together and constrain each other in some way?

Anand:Yes and no. The word "constraint" has a strong meaning in the idea of blockchain. I think it makes sense to set some defenses and operating modes. But at the same time, it depends on our definition. We need to be able to audit and verify how these agents are built and operate, which is probably the most important thing. Today, we know very little about closed source systems and how they operate. So we should operate in a more open, transparent, verifiable, auditable way, which is the basis for building anything. The second is defense, what parameters do these agents need to operate under? This is in a way "constraint". I think we should not constrain the innovation part of it, but we should control the operability and auditability part. I think the two definitely go hand in hand.

Jared: I think the future will be me and my team of AI agents working together, and they will know me better than I know myself. Every morning I communicate with the AI to help me organize my work for the day, and this will only become more and more common. So one question is whether it is more important to have user-owned AI rather than company-owned AI, like NEAR is doing. What are your thoughts on this space? I feel like these AIs will know me so well that I want them to be user-owned, not the other way around.

Anand:In a way, it’s inevitable. Now we’re starting to realize the potential of AI, which for a long time we only discussed theoretically. Like Neal Stephenson’s The Diamond Age, where the characters grow up with AI. It’s going to become natural and that’s what we want. We want agents to help us with tasks like last night someone sent me a message on Telegram and I want to know the right response. Machines are designed to augment human behavior, but we also need to think about how much responsibility we give them to actually work for us. These are all things that need to be considered. User-owned AI is the inevitable direction of things because you want to have control over how these things work. The concept of digital twins is becoming more and more popular today, and in order to do that, they need to know you and they need to operate in a privacy-centric way. All of this sounds exactly like what we’ve been preaching in the blockchain world for more than a decade, and I feel like these two technologies are on a collision course.

Altan: Suppose in a world where everything is done by Agents, they have their own autonomy to operate. If an Agent is performing a task, where does the security come from? How to convert the security of this off-chain operation into on-chain proof? For Agents, is security driven by cryptoeconomic value, or is it achieved through cryptography through zero-knowledge proofs?

Anand:In this case, the use case dictates what the flow chart looks like. The reality is that there are many cases where agents can run autonomously using cryptoeconomics because that is more natural and local. Any time you need to use an off-chain or external environment to operate on-chain transactions, it only adds complexity. We are getting better and better at using oracles, provers, and off-chain computation, but these still add additional complexity. So, I think cryptoeconomics is a better place to start, but there may also be use cases where you have to consider off-chain processes.

We have users and agents, user-assisted agents and their modes of operation, which all need to be encoded and verified, and this is where blockchain comes in. Then they need to transact, how do they transact with each other? Just like humans have bank accounts, these agents don't know what a bank is. For them, wallets and cryptocurrencies will be the first choice. There are cases where agents need to operate with insufficient information and therefore need to rely on humans to fill in the missing information. This is where off-chain computation becomes really important. I remember at a side event we hosted in September 2023, someone put it very succinctly: We will be working for AI for the foreseeable future. I remember him predicting that this would happen this year, and while we’re not sure it will happen, it’s already happening. Agents will post bounties for humans to fetch information for them, and they’ll pay them for it. It’s going to be a very interesting vision of the future.

It feels very sci-fi, and even a little dystopian in places, but the pace of technological progress is unprecedented. I don’t want to sound like a futurist, but I must say that I took a programming language called LISP in 1999, and it was called an “AI course.” We thought robots were going to take over the world, and we were going to live in a future with flying cars. The internet was just emerging, and we thought things would get there very quickly. However, we found that the development and progress of AI was missing. There were two very different eras before and after the Transformer model, and the lack of data limited the model from being fully trained.

I mention this because while this has always felt like science fiction, it is now closer to reality than ever before. The things we are discussing may soon become true. As a pure technology enthusiast, I find this exciting, although perhaps a little naive. I feel like we will become super efficient and focus on the things we really care about because everything else will be done by these smart user-owned models and agents.

Data Privacy Challenges Facing AI Agents

Altan: Another problem is that we have a lot of centralized companies. Data collection and processing is mainly done by Scale, model training is dominated by OpenAI, Google, etc., and GPUs are almost completely monopolized by Nvidia. How can we win in the face of such huge capital and power? How to decentralize?

Anand:The past few decades of Internet life have taught us a lot. Take Google as an example. It is the cornerstone of obtaining Internet information, but its business model is interesting. Although you search on it, it will redirect you to the content page containing that information. But the level of intelligence we are seeing now is trying to allow you to access this information without leaving the current experience, such as ChatGPT. Last week, several journalists began to discuss Perplexity because it is summarizing news articles in real time and rewriting the content of the articles. This is not just retroactive information or Wikipedia-like, but real real-time information. Perplexity or Twitter (now called X)'s Grok can give you real-time information based on existing data patterns.

Power is gathering in centralized entities. On the one hand, we are moving towards a centralized world with a lack of transparency; on the other hand, people are questioning the degree of privacy of these centralized services. Someone posted on Reddit that he connected a Nest Cam to his apartment, collected video clips and gave it to GPT-4 to process, and GPT-4 could tell him the location of items such as keys in real time. I think it's cool, but also a little crazy. Using a centralized server to handle what happens at home means that you are only one step away from data leakage. We have heard of many data breaches, such as the recent DocuSign breach. Some companies' only responsibility is to protect user information from being leaked, and once a leak occurs, the consequences are disastrous. Imagine that a hacker hacked into the video camera in the home and then transmitted the data to a centralized service, and privacy protection was turned off by default. You may have unknowingly agreed to provide information to optimize these models. These are existing problems, so we need to move away from centralized services to user-owned or collaborative management of these models and operations.

What do you want the world to look like in 20 or 30 years? Jared said it well, there are a group of agents working for me and handling a lot of things for me. How many of these agents would you rather have controlled by a large centralized entity trying to monetize your data, rather than you having ownership and control over it yourself, while enabling all the technological advancements? That’s the core of the issue, it comes down to ownership, data privacy, and autonomy.

Altan: Data privacy is not new. In 2017, Google tried federated learning, claiming that it would not extract user data, but just train data on the user's device and send it back to a central server so that they could not see the user's data. But the problem is, my grandmother or mother didn't care about this. So I wonder, are we going to have a split? Some people don't care, and people like us who are tech-savvy, like those who work in the crypto field, care a lot. Is this what the future looks like?

Anand: The average user does not care about this, 100% agree. Even the people around us, even those in the tech community, today there are a lot of people who don't care enough. But we are going through a technological revolution and renaissance, and as developers, it is our responsibility to think about these issues. Because we are leading the way here. Do you trust these centralized entities and let them monopolize our data for their own interests and uses, or do you have some kind of ownership at the end? A more forward-looking example is data DAOs, which is an increasingly popular concept. For example, there's a company called Vana that's thinking about data DAOs.

I would feel more comfortable if I had some control over this federated learning that you mentioned. Fred Wilson of Union Square Ventures talked about these primitives about a year and a half ago, like a wallet that can be used to sign transactions or perform certain actions on my behalf. What if we could use this similar model to handle my data? This would open up a process like data attribution, data contribution, data encryption or data stripping for training without any personally identifiable information (PII). These are all things we can just do from first principles.

It's well known that OpenAI is not really open, and I think other models like Llama have the same problem. The question is, although these models disclose some of their weights, what is the source of the data? How is this data trained? Are the people who contribute the data compensated? Are they participating voluntarily? What is the current situation? We should be proud of using a system that has transparency and user autonomy. But it doesn't start with the user because the user generally doesn't care about these issues and I don't think they will in the foreseeable future. That said, I think the current generation probably cares more about these issues than the older generations did. My daughter and the younger generation right now care more about issues like climate because they affect them and their future. So I think the current generation cares about these issues, the previous generation probably didn't. But let's build on these principles and build a better future.

Jared: Privacy data may not be the key to attracting people to participate, but the value you mentioned is the key. What blockchain is really good at is creating value or making value visible, capturing value in a way that centralization cannot do. In the past, we have been in the Internet environment of Web 2.0, where platforms like Myspace profited by farming people's data, and this was the business model we all accepted. But now, we are experiencing a revolution and renaissance, as you said, we can give users more choices about their data and behavior. They can decide whether to share data, participate in AI training, and get actual rewards for their contributions. The low-cost and efficient value transfer mechanism of blockchain provides the basis for realizing this vision.

Anand: Yes, we all do it and are used to it, and we are more aware of it now than ever before. I don't use the GPT model now (this is just my personal choice, please refer to it as appropriate), and all reasoning is done on the local machine. I have a Mac with an M2 chip, which can handle some complex reasoning tasks.

Maybe this is a bit paranoid, but my mentor told me in the 2000s that as a developer, you need to maintain a healthy paranoia. So I always work with this mentality. I hope to see more people have autonomy and be able to operate the data more on their own. There is a company called EXO, and they are a good example. They have achieved reasoning on edge devices, or let users better manage and operate their own data. Of course, we have also heard that Apple likes to label everything as Apple, so they now call it "Apple Intelligence". A lot of things will be done on edge devices, and if the device can't handle it, ChatGPT or other models will be used. User data is only one step away from being shared with centralized services.

I think the existing centralized and decentralized models will coexist, and they are not alternative relationships. This is like the debate between open source and closed source, which is a cliché example, like Android versus iOS. We need these comparisons to meet the various needs arising from intelligence.

Investment Value and Future Outlook of Crypto AI

Altan: There are many people who enter the field of crypto AI just because of the hype of AI, and they just want to extract value from it. As an investor, how do you view this? What is the value of crypto AI in the short and long term?

Anand:In the short history of crypto, there have been very few winner-take-all situations. We have a lot of great L1 networks, a lot of great L2 networks, and a lot of L3 networks that are emerging. It is not uncommon in the blockchain and crypto space to use token incentives to solve the same problem in a differentiated way. The idea originated from the need for access to compute resources, especially bare metal compute resources. Take Akash, for example, they started earlier than most projects and are one of the most reliable systems for accessing bare metal compute resources. These compute resources can be used for model training, inference, and other purposes.

But as we develop, people start to move up the technology stack and move to more software-centric models, such as API-centric approaches. This is a bit like the difference between Heroku and AWS: Heroku is more focused on simplifying deployment and fast online, while AWS provides lower-level services. People want to be able to deploy applications quickly and use them immediately, so we have to continue to evolve in this direction. At present, we are seeing a lot of low-hanging fruit being achieved one by one, and the most prominent one is the access to compute resources. Let's get through this quickly, and there may be some winners along the way. The market structure is not going to be "stubby," it's going to be a "head" and a very long "tail," where a few big winners and a lot of small players coexist.

I'm very interested in differentiation and defensibility. That's the most important thing today. I would also emphasize that if you spend more time in AI than in crypto and find highly defensible and differentiated applications, bring them into the crypto ecosystem, or use cryptoeconomics or crypto primitives to build decentralized AI, that's very valuable. So I don't think there's a specific short-term or long-term call to action here.

For me, the most important thing is talent, bringing people with solid backgrounds into this space. We like to work with smart and insightful people who are building something that's very hard to do so that we can build differentiated products and go to market.

Jared: In your day-to-day work in the industry, what area do you think should get more attention that you feel a lot of people don't pay attention to? It could be AI or crypto.

Anand:For me it's data. Data is like oil in some ways, it's necessary for large-scale systems to function, but we tend to take it for granted. This will continue to be a problem, and the existing centralized systems already have large models today. It will be difficult for me to catch up with their level because they have 30 years of data accumulation and mature systems to scrape data from the Internet and train models. So data is something I think about a lot. It has several branches, including user-contributed data and the source of data. For example, what does Jared's data mean? We can look at this issue from both the user and developer perspectives.

From a developer's perspective, let me give you an example that may not be appropriate: suppose I use a trained model for inference, and I want the model to tell me which different data sources contributed to this inference. This is like running in debug mode, I want to figure out what the inference is based on and why the model gives me such an answer. Because the model gives more than just a website link, but the result of intelligent processing. I don't understand the specific source of this intelligent information, because a lot of complex work needs to be done at the bottom to achieve this. But I think that figuring this out is a very important step in understanding the decision-making process of the model and improving its interpretability.

Energy is another aspect that needs to be paid attention to now and in the future. In Zuckerberg's interview with Dwarkesh, a topic that was repeatedly mentioned was energy demand and how we take energy consumption for granted. Centralized services are consuming energy unscrupulously, using billions of watts of electricity to train models, and all these GPUs are running at full speed. And that's just the training phase, and there's inference later. I'm not sure at what point this problem can be connected to blockchain infrastructure, but I think this is the starting point to start solving this problem. I'm throwing this question out there for others to take or run from.

Final Words of Advice

Altan: If anyone is listening to this podcast, what should they take away from it?

Anand:I'd say two things. One, build for the future. Have a long enough time horizon, because I think things are going to change a lot in the next few years. The smartest and best founders put themselves into the future and think about what the world is going to need in the short term. That time frame could be six months, it could be two years. Because those who live in the future have extraordinary insights into what to build and how to get there.

The second point, and I think Marc Andreessen put it more succinctly: The best founders are often the ones who make their perception a reality, and your job is to try to convince others. For example, I see a world that needs a decentralized social network, similar to Farcast's relationship to Twitter. Dan has been relentless in pushing this over the past few years. I think visionary founders like these are the ones who can see the future, turn their perception into reality, and spread this idea to others. If you are such a person, I look forward to talking with you, understanding your ideas, and applauding you for being one step ahead and creating the future.

Original link

欢迎加入律动 BlockBeats 官方社群：

Telegram 订阅群： https://t.me/theblockbeats

Telegram 交流群： https://t.me/BlockBeats_App

Twitter 官方账号： https://twitter.com/BlockBeatsAsia