Ship or Die 2025: Where AI Meets Web3: Reimagining Digital Infrastructure

The worlds of artificial intelligence and Web3 are colliding, ushering in a new era of digital infrastructure. In a groundbreaking discussion, industry leaders Chris Zhu and Tim Fan reveal how cutting-edge AI models and agentic crypto browsers are set to revolutionize the way we interact with the digital world, particularly in the realm of cryptocurrency trading.

Summary

In this enlightening conversation, Chris Zhu, CEO of Sonic SVM, and Tim Fan, a leading AI researcher, delve into the latest advancements in multi-modal AI models and their applications in the Web3 space. They discuss the recent release of Bago, an open-source multi-modal AI model capable of processing text, images, and videos, marking a significant milestone in AI development.

The conversation then shifts to the emerging field of agentic browsers, with a particular focus on their potential in the cryptocurrency sector. Tim Fan shares insights into the rapid development of agent models and the growing interest from major tech companies like OpenAI. The discussion culminates with the reveal of Donut, an innovative agentic crypto browser project that aims to transform how users interact with and trade cryptocurrencies.

This convergence of AI and Web3 technologies promises to address key challenges in the crypto space, such as wallet security and intelligent trading strategies. The experts highlight the potential of reinforcement learning (RL) in developing sophisticated trading models and discuss how these advancements could lead to more efficient and secure cryptocurrency transactions.

Key Points:

Bago: A Breakthrough in Multi-Modal AI

Tim Fan introduces Bago, a groundbreaking open-source multi-modal AI model. Unlike traditional AI models that require extensive human-annotated data, Bago was trained by "watching" large-scale internet videos, primarily from YouTube. This innovative approach allowed the model to develop emerging abilities over time, learning to understand the world through video content.

The training process for Bago was unique and somewhat serendipitous. Initially, the team thought they had made a mistake as the model showed no signs of learning. However, after running the training for a week, they began to observe emerging abilities. This discovery led them to continue training the model on GPUs for several months, resulting in the impressive capabilities seen today.

Bago's release has been met with significant enthusiasm from the developer community, garnering over a thousand GitHub stars and approximately a thousand adoptions within a short period. Its ability to process and understand text, images, and videos in a conversational manner opens up new possibilities for AI applications across various domains.

The Rise of Agentic Browsers

Tim Fan provides insights into the rapidly evolving field of agentic browsers, which has seen a surge of interest and development over the past year. He notes that while many initially believed general-purpose agents were a distant concept, recent releases from companies like OpenAI have accelerated progress in this area.

The conversation reveals that major players in the tech industry are investing heavily in this technology. OpenAI, for instance, has reportedly hired key members of Google's Chrome team to work on their browser model. This move signals the growing importance of agentic browsers in the future of digital interaction.

Fan emphasizes that the field has moved beyond debate and into active development, with dozens of open-source agent models now available for download and experimentation. This rapid progress indicates that agentic browsers are not just a passing trend but a significant shift in how we interact with the internet.

Donut: Revolutionizing Crypto Trading with AI

Chris Zhu and Tim Fan introduce Donut, an innovative project they're collaborating on that aims to create the first agentic crypto browser. This browser is designed specifically for cryptocurrency trading and financial execution, addressing key challenges in the intersection of AI and crypto.

The experts highlight two critical aspects that Donut aims to solve:

1. Security: Ensuring that financial credentials and trading history remain secure and are not exposed to potential misuse by AI training processes.

2. Domain-specific intelligence: Developing an AI model with specialized knowledge in cryptocurrency markets and trading strategies, going beyond the capabilities of general-purpose AI models.

Fan explains that while general AI models like those from OpenAI are powerful, they are not inherently capable of making sophisticated trading decisions. Donut aims to bridge this gap by creating a vertical domain knowledge base and integrating real-time financial market data.

Reinforcement Learning in Crypto Trading

Tim Fan elaborates on the potential of reinforcement learning (RL) in developing advanced trading models. He explains that the cryptocurrency market is particularly well-suited for RL applications due to the clear reward mechanisms inherent in trading (profit or loss).

The experts envision a closed control environment where an agent model can learn and improve its trading strategies through RL. This approach would allow the model to refine its decision-making process, rewarding successful trades and penalizing poor choices.

By combining powerful reasoning capabilities with RL in a controlled environment, Donut aims to create a sophisticated AI model capable of making informed trading decisions in the volatile cryptocurrency market.

Facts + Figures

• Bago, the open-source multi-modal AI model, gained over 1,000 GitHub stars and approximately 1,000 adoptions shortly after its release.
• The development of Bago involved training the model on large-scale internet videos for several months.
• OpenAI reportedly hired key members of Google's Chrome team around November or December of the previous year to work on their browser model.
• The collaboration between Chris Zhu and Tim Fan on the Donut project began around March.
• There are currently dozens of open-source agent models available for download and experimentation.
• The convergence of AI and crypto trading is seen as a recent trend, with major developments occurring within the past year.
• Donut is set to be announced to the world as the first crypto agentic browser in the week following this discussion.
• The development of agentic browsers has seen contributions from researchers and developers from institutions like Berkeley, Stanford, MIT, and Matab.

Top quotes

1. "We don't want to do the traditional training, which you require a lot of human annotated data. We wanted to just learn the intelligence by itself." - Tim Fan on Bago's training process

2. "Unless the moment I feel this is actually not a hype, but a real moment of wave." - Tim Fan on the development of agentic browsers

3. "You cannot expose your financial credential, your API, or your tracking history to OpenAI. They will definitely use that for training for purpose." - Tim Fan on the security concerns of using general AI for crypto trading

4. "RL is perfectly designed for this task. As long as you can have abundant reward value and reward function, then you can basically, in a closed control environment, the agent to mask or the task." - Tim Fan on using reinforcement learning for crypto trading

5. "I feel Donuts have a clear killing advantage. So basically the team is funded by people with perfect sense on finance, on security and also on agent." - Tim Fan on the potential of Donut

Questions Answered

What is Bago and how does it differ from other AI models?

Bago is an open-source multi-modal AI model that can process and understand text, images, and videos in a conversational manner. Unlike traditional AI models that rely on human-annotated data for training, Bago was trained by "watching" large-scale internet videos, primarily from YouTube. This unique approach allowed the model to develop emerging abilities over time, learning to understand the world through video content without the need for extensive human labeling.

What are agentic browsers and why are they important?

Agentic browsers are a new type of web browser that incorporates AI agents to enhance user interaction and automate tasks. They are important because they represent a significant shift in how we interact with the internet, potentially automating complex tasks, improving search capabilities, and providing more personalized experiences. The development of agentic browsers has gained significant traction in recent years, with major tech companies like OpenAI investing heavily in this technology.

What is Donut and how does it aim to revolutionize crypto trading?

Donut is an innovative project aimed at creating the first agentic crypto browser specifically designed for cryptocurrency trading and financial execution. It aims to address two critical challenges in the intersection of AI and crypto: security of financial credentials and domain-specific intelligence for trading. By combining powerful AI reasoning capabilities with reinforcement learning in a controlled environment, Donut seeks to create a sophisticated model capable of making informed trading decisions in the volatile cryptocurrency market.

How does reinforcement learning (RL) apply to cryptocurrency trading?

Reinforcement learning is particularly well-suited for cryptocurrency trading due to the clear reward mechanisms inherent in trading (profit or loss). In the context of crypto trading, RL can be used to create a closed control environment where an AI agent can learn and improve its trading strategies over time. The model can be rewarded for successful trades and penalized for poor choices, allowing it to refine its decision-making process and potentially develop sophisticated trading strategies.

What are the main challenges in using AI for cryptocurrency trading?

The two main challenges in using AI for cryptocurrency trading are security and domain-specific intelligence. Security is crucial because financial credentials and trading history must be protected from potential misuse, including by AI training processes. Domain-specific intelligence is necessary because general-purpose AI models, while powerful, are not inherently capable of making sophisticated trading decisions in the crypto market. Developing an AI model with specialized knowledge in cryptocurrency markets and trading strategies is essential for effective AI-driven crypto trading.