The Decentralized Future of Private AI with Illia Polosukhin - #749
🎯 Summary
Podcast Summary: The Decentralized Future of Private AI with Illia Polosukhin - #749
This episode of the Twombo AI podcast, hosted by Sam Charrington, features Illia Polosukhin, co-founder of Near AI and co-author of the seminal “Attention Is All You Need” paper, discussing his current focus on building a decentralized and private future for Artificial Intelligence. The conversation bridges Polosukhin’s foundational work on Transformers with his current efforts at Near Protocol and Nira AI to address centralization risks in the AI landscape.
1. Focus Area
The primary focus is the Decentralized Future of Private AI. The discussion centers on mitigating the risks of centralized AI monopolies by leveraging blockchain principles (user ownership, self-sovereignty) combined with cutting-edge hardware capabilities (Confidential Computing) to enable private, scalable cloud-based AI services. Secondary themes include the evolution of model training data (shifting from raw user feedback to curated, verifiable supervision) and the regulatory burdens (like GDPR) making user data a liability for developers.
2. Key Technical Insights
- Confidential Computing as the Enabler: The convergence of recent hardware advancements (Intel’s 5th Gen Xeons and specific Nvidia modes) enabling Confidential Computing (secure enclaves) is the critical technical breakthrough. This allows computation to occur in an encrypted state, inaccessible even to the hardware owner, providing cloud-level intelligence with local-level privacy guarantees.
- Decentralized Confidential Machine Learning (DCML): This framework combines user data privacy (data remains encrypted) with model developer IP protection (model weights are encrypted and only decrypted within the secure enclave during inference). This solves the dual trust problem in the cloud.
- Shift in Model Refinement: The industry is moving away from relying solely on noisy, large-scale user feedback (like clicks) for fine-tuning towards more curated, verifiable supervision, synthetic data, and expert human labeling (e.g., using specialized CS students for code-related models).
3. Business/Investment Angle
- Data as a Liability: Regulatory pressures (GDPR, data taxes in China) are transforming consumer data from a “gold mine” into a significant legal and financial liability for centralized developers.
- Decentralized Compute Market: The DCML approach opens up the possibility of leveraging distributed, residential cloud hardware (GPUs/CPUs) for AI workloads without the model developer needing to manage data compliance or the user needing to trust a single cloud provider.
- Model Leakage Risk: Centralized GPU cloud providers pose an existential risk to model developers whose proprietary weights can leak (e.g., the mentioned Mistral weight leak via Hugging Face), driving the need for encrypted model deployment solutions.
4. Notable Companies/People
- Illia Polosukhin: Co-author of the Transformer paper, co-founder of Nira AI, and founder of Near Protocol (a major blockchain network with 50M MAUs, initially born out of solving global micropayment issues for data labeling).
- Google Research: Where Polosukhin worked on question answering and machine translation, leading to the Transformer architecture.
- Intel & Nvidia: Their recent hardware updates enabling secure enclave functionality were crucial for realizing the DCML vision.
- OpenAI/Mistral: Mentioned as examples of centralized entities whose data practices or security vulnerabilities highlight the need for decentralized alternatives.
5. Future Implications
The industry is heading toward a model where intelligence is accessible via the cloud but remains fundamentally user-owned and private. This decentralized confidential cloud aims to offer the scale and intelligence of centralized services while eliminating the “1984” risk associated with a few profit-driven entities controlling fundamental decision-making technology. It suggests a future where developers push software, not data, to users, while still enabling powerful background processing.
6. Target Audience
This episode is highly valuable for AI/ML Engineers, Blockchain Developers, Product Managers in Tech, and Venture Capitalists focused on infrastructure, privacy-preserving computation (PETs), and the long-term governance of foundational AI models.
🏢 Companies Mentioned
💬 Key Insights
"And then we also have this concept called multi-party computation. So the Near blockchain itself kind of right now, part of our nodes, form this multi-party computation network which allows inside the secure enclave effectively have its own private key to decrypt things."
"There is, I mean, we've seen this with OpenAI, right? There's news that the fact that scanning all the chat logs and then the ones that are flagged are sent to human for evaluation and then to police, right?"
"The overhead on the computation side is 1 to 5%."
"This multi-party computation network which allows inside the secure enclave effectively have its own private key to decrypt things."
"what's happening is, you know, you checkpoint your, you know, model weights on-chain, so you know the hash, the hash of the model weights, and encrypted hash as well. You know, the encrypted data is uploaded kind of decentralized storage."
"Or you can just say, "Actually, I'm going to build my app and deploy it into this cloud where it runs on it on your side, right?" You know, and saves context there as well in your data store. And so now your data store becomes even more useful because it has all the notes as well there, and your AI can now read over those notes."