.svg)
Top insights from the expert panel on AI and crypto
We’ve just held a Twitter Space on AI and crypto with our friends from Alethea AI, Panther Protocol, and Swap GPT, and it was one of the best expert panels we’ve ever had! We gathered the best insights in this recap, but you should definitely listen to the full recording here.
Panel speakers
- Host: Alejo Pinto, Chief Growth Officer at Pontem Network
- Dr. Anish Mohammed, co-founder, CTO & Chief Scientist at Panther Protocol, a ZK-powered cross-protocol layer that enables data privacy and DeFi compliance
- Shradhaa Agarwal, co-Founder and CTO at SwapGPT, provider of AI tools for liquidity management and decentralized trading
- Keshav Saraogi, co-Founder and CEO at SwapGPT
- Brent Homesley, Head of Partnerships at Alethea AI, a research & development studio focused on generative AI and known for CharacterGPT, an algorithm that lets you generate characters from text prompts
Can crypto combat AI centralization?
Alejo Pinto started by explaining how the emergence of LLMs like ChatGPT could be a fundamental turning point in the history of technology:
“We see fiction become reality. LLMs are a tipping point comparable to the splitting of the atom, or the invention of the internal combustion engine, or the printing press, or cryptocurrency itself. It’s both exciting and scary to know that AI may soon be everywhere.”
AI is the biggest tech narrative right now, alongside Web3. Many call them a perfect match and imagine various ways in which artificial intelligence will “disrupt” blockchain. But the two technologies have opposite opposite relationships to centralization, which could make marrying them extremely challenging.
Alejo Pinto explained that AI is a massive centralizing force. Most of the majorLLMs belong to large corporations, especially since Microsoft purchased OpenAI. Google, Meta, Microsoft, etc. accumulate data from users, as well as the algorithms themselves, which in turn helps them extract even more data, feed that back to the AI, and so on. It’s an endless cycle.
Brent Homesley from Alethea AI added that in the future, as LLMs accumulate even more information, AI may even come to know more about you than you know yourself. Artificial intelligence will suggest products before you even realize you need them.
Crypto, by contrast, is geared toward decentralization. It’s a force that works against the control of powerful corporations and for data sovereignty. Can crypto help counterbalance AI’s centralizing impulses?Brent Homesley believes the answer is yes:
“Think of what happened with TikTok and Instagram. Creators on Instagram realized they were essentially making money for the centralized company. So they started leaving and going over to TikTok where they could monetize their content better.
It’s the same with ChatGPT – when you use it, you are training the model for them. AI is getting smarter off the labor of the masses. But just as it happened with TikTok, some people may leave centralized AI models, and we will have AI platforms that are co-owned”.
Brent also stressed another way in which AI and crypto are different:
“AI is an abundance engine. It can create ideas and even businesses, and it will do so as many times a day as you ask it to. Web3 is a scarcity engine, because on the blockchain, you can prove ownership of a scarce resource. For example, you can ask AI to generate 100 images, then right-click and save them all – that’s abundance. But you can also mint an NFT of such an image; it will be recorded on the blockchain and you will own it.”
Centralized AI is like “junk food”
But do people really care about the privacy of the data that they share with AI? Indeed, most happily use ChatGPT without worrying that it will save the chat history, be used for targeted advertising, or worse. Plus, it’s thanks to centralization that AI products are so cheap for end users: corporations don’t charge you because they can monetize all that information.
However, Alejo Pinto said that privacy is starting to become a consumer trend. He predicts a dual ecosystem, where some continue to use cheap or free centralized AI and others opt for more expensive decentralized products. He offered an interesting comparison with food:
“In the 70’s, 80’s, and 90’s, people mostly ate processed food and nobody was concerned about healthy eating. But in the past 20 years, organic food has become a huge trend, though it’s more expensive than regular food. In the future, centralized AI will be like junk food – cheap, fast, calorie-dense, but not ‘healthy’. Many users won’t care and will still consume it. Decentralized, ‘organic’ AI will be more expensive and better for your privacy health, so some users will opt for that. In any case, decentralized AI needs to have equally good UX and UI.”
AI for DEX trading and DAOs
When discussing the best use cases for AI in Web3, the speakers agreed that we should focus on areas where artificial intelligence can assist human decision-making, because Ii’s too early to replace humans with AI. For example, LLMs often “hallucinate”: they generate confident-sounding responses that are totally false, and mix them up with true statements in such a way that it can be difficult to spot.
Alejo even compared it to self-driving cars: AI can assist you in driving, but it’s a long way until you can take your hands off the wheel and completely entrust your safety to an autonomous vehicle.
But even before we let AI inform our decision-making, we have to make sure that it uses reliable data and that the model is verified. Keshav Saraogi and Shraddha Agarwal from SwapGPT explained:
“AI predictions are only as reliable as the data. You need to use vetted, cleaned, and structured data – something that is easy to get in centralized finance but can be difficult in DeFi. At SwapGPT, we use many sources of on-chain data, such as blockchain explorers and graphs, but also social media sentiment.
Once you’ve cleaned all of it, you have to validate the model – the task is to avoid hallucinations from Day 1. DeFi is a rapidly changing market, and it can be difficult for machine learning models to react and adapt quickly enough.”
Suppose we build a validated model using clean data. What can it be used for in Web3? The SwapGPT team believes trading and liquidity management are the best use cases. Indeed, AI can process huge loads of data from DEXes, and even find correlations with social media trends.
Such a model can then provide suggestions for managing liquidity and trading. It could help avoid emotional trading, improve trade execution, and avoid losses. You can even use natural language prompts based on GPT to make communication with the AI easier.
Keshav Saraogi stressed that an AI liquidity manager is not a self-driving vehicle: it won’t do all of the trading work for you.
“The human trader will have to provide inputs for the model: how to execute the trade, how much risk to take, etc. Once done, the AI will be able to react to market events faster than a human trader, and you won’t have to monitor the model actively.”
Another interesting use case for AI in Web3 is DAO decision-making. Here, AI can generate proposals for the DAO to vote on, or evaluate if a proposal is good. Alejo Pinto cited the famous credit protocol Maker, which recently published a roadmap with lots of AI features.
Brent from Alethea AI even thinks that DAOs could emerge around AI models themselves:
“Imagine a group of people with a vested interest in an LLM. They could form a DAO and vote on which points they want or don’t want the model to be trained on. Or, if it’s a trading AI, how they want it to trade. It’s like crowdsourcing information from a DAO”.
Is AI privacy too expensive?
Perhaps the most thought-provoking part of the discussion revolved around the costs and economic incentives for decentralized AI.
As for the costs, Alejo and the guests agreed that there is no practical way yet to run full-scale ML models on-chain. Computation and storage would be too expensive. Perhaps in the future the fastest, most scalable chains like Aptos will be able to support AI apps.But computation and training data will have to remain off-chain for now.
Here, Anish Mohammed from Panther Protocol was the best expert to ask. First of all, Anish explained that the biggest issue isn’t even the cost of computation, but storage. Computation costs can go down exponentially, making ML models much cheaper to run. However, storage normally happens in a closed network, and the network is a constraining resource that expands linearly.
You can try to conquer this by building decentralized storage systems, and in fact, this is the approach of federated learning, or distributed AI. However, it tends to have lower accuracy than centralized AI.
Keshav and Shraddha from SwapGPT agreed that off-chain computation is a good compromise, as long as transactions themselves are on-chain. However, with the spread of AI in Web3 it will be important to verify that off-chain computations are correct.
The best way to do this might be zero-knowledge proofs. However, according to Dr. Mohammed, building a decentralized version of any ML app with ZK proofs will always cost more than a centralized model. That extra cost will have to be covered, perhaps by user fees. But will users pay for these features?
Read our article about ZK technology
Anish Mohammed has doubts:
“Current AI pricing mechanisms don’t account for privacy and compliance. If you take these factors into consideration, there may be valid economic incentives to do this – to use ZK to build AI products. But I haven’t seen anything driven by privacy alone so far. If it was, Google and Facebook would already be on it.
Privacy is like a smoothie at McDonald’s. I won’t go to McDonald’s just to order a smoothie – but if I go to get a burger, I may get a smoothie, too.”
Hopefully, we will see incentives align for AI apps to use decentralization and ZK-based privacy mechanisms. Fast blockchains like Aptos should help reduce the costs.
There were too many great insights in this panel to cover here, so make sure to listen to the entire AI + Crypto panel here.
About Pontem Network
Pontem Network is a blockchain product studio building for Aptos, the L1 blockchain known for its sub-second finality and security. Our products include:
- Pontem Wallet: the only triple-audited wallet for Aptos with 300,000 installs
- Liquidswap DEX: one of the most popular DEX on Aptos
- Move Code Playground: the first browser code editor for the Move language.
Pontem is currently preparing to release PontemAI , a crypto-native chatbot powered by best-available market data.. Pontem also sponsored the AI track at Aptos’s recent Hack Holland event.
