A Crypto-Incentive Approach for Training Data

Key Concepts:

• Commoditized AI models
• Data bottleneck in AI development
• Multi-modal models (robotics, video, audio)
• Crypto incentives for data collection and labeling
• IP safety and data privacy
• Decentralized data collection and licensing
• Immutable ledger for data tracking and attribution
• Data dividends
• Deep in app
• Physical infrastructure app

1. The Data Bottleneck in AI Development

• AI models have become commoditized, and compute costs have decreased.
• The primary bottleneck is now the availability of high-quality, relevant data, especially for multi-modal models used in robotics, video, and audio applications.
• Scraping data from the internet is often insufficient and raises IP and safety concerns.
• "The bottleneck, as you see it in many others, here is the data."

2. Crypto Incentives for Data Collection and Coordination

• The solution involves using crypto incentives to coordinate the collection, labeling, and curation of data.
• Crypto enables the coordination of a massive labeling effort.
• Crypto is good at coordinating incentives at a global scale and paying contributors instantaneously with stablecoins or other crypto tokens.
• "We are using actually crypto incentives to actually coordinate that massive label coordinate."

3. Advantages of Using Crypto and Blockchain

• Global Coordination: Crypto facilitates the coordination of incentives on a global scale, enabling instant payments via stablecoins.
• Immutable Ledger: Blockchain technology allows for putting data and IP on an immutable ledger, enabling tracking, attribution, and licensing through smart contracts.
• These advantages are compelling even for AI specialists.

4. The Vacuum Left by Meta's Acquisition of Scaling

• Meta's acquisition of Scaling has created a vacuum in the data licensing and collection space.
• Scaling relied on shell companies in developing countries to coordinate data collection, which can be problematic.
• Crypto offers a more efficient and transparent alternative to these "digital sweatshops."
• "40 days ago actually Metta Semi acquired actually as you know a scaling on right as a that left a very huge vacuum in the space to actually for competitors and new players to come in."

5. Real-World Application: Training Humanoid Robots

• The concept is analogous to Axie Infinity, where people in developing countries earned tokens by playing the game.
• In this case, people are incentivized to video themselves in real-world scenarios to help train AI models, such as Tesla's Optimus robot.
• Training a humanoid robot requires a vast amount of egocentric video footage from various settings.
• Instead of hiring shell companies, crypto incentives can coordinate this data collection at a mass scale.
• "You're going to use people, incentivize people to sort of video themselves in the real world. That then ultimately helps Tesla build Optimus in the real world."

6. IP Safety and Decentralized Data Licensing

• Data contributors set their licensing usage terms on the blockchain, ensuring IP safety.
• The platform is transitioning to a decentralized deep in app and physical infrastructure app for data collection.
• Currently, the company acts as an agent to secure contracts.
• "The reason why it's IP safe is that everyone who are actually contributing data, anyone can contribute this data anywhere on the world, right? And then actually set their licensing usage terms on the blockchain."

7. Demand and Competition

• Many organizations are interested in this solution, with some dropping contracts from Scale.
• The company has no plans to be acquired, emphasizing the advantage of an open system that can work with any IP.
• "Actually many, many organizations that I was really surprised to see stage company. They were willing to work with us because a lot of them were telling us that they were dropping contracts from scale."

8. Technical Terms Explained

• Multi-modal models: AI models that process multiple types of data, such as video, audio, and text.
• IP safe: Ensuring that data used for training AI models respects intellectual property rights.
• Immutable ledger: A blockchain-based record of data and transactions that cannot be altered.
• Deep in app: Decentralized application.
• Physical infrastructure app: Application that collects data from the physical world.

9. Synthesis/Conclusion

The core argument is that while AI models and compute power are becoming more accessible, the availability of high-quality, IP-safe training data remains a significant bottleneck. The proposed solution leverages crypto incentives and blockchain technology to coordinate global data collection, ensure IP rights, and provide a more efficient and transparent alternative to traditional data sourcing methods. This approach aims to fill the vacuum left by Meta's acquisition of Scaling and cater to the growing demand for AI training data, particularly in robotics and multi-modal applications.