Episode

The Decentralized Future of Private AI with Illia Polosukhin - #749

Podcast: The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)
Published: Sep 30, 2025
Duration seconds: 3903
Processing state: processed
Canonical source: https://twimlai.com/podcast/twimlai/the-decentralized-future-of-private-ai/
Audio: https://pscrb.fm/rss/p/traffic.megaphone.fm/MLN2189764781.mp3?updated=1762292711
JSON: /v1/public/podcasts/twiml-ai-podcast/episodes/the-decentralized-future-of-private-ai-with-illia-polosukhin-749
Markdown: /podcast/twiml-ai-podcast/the-decentralized-future-of-private-ai-with-illia-polosukhin-749.md

Actions

POST https://stenobird.com/v1/public/podcasts/twiml-ai-podcast/episodes/the-decentralized-future-of-private-ai-with-illia-polosukhin-749/transcription-requests
Idempotently request low-priority transcript generation for this episode.
GET https://stenobird.com/podcast/twiml-ai-podcast/the-decentralized-future-of-private-ai-with-illia-polosukhin-749.md
Read the agent-friendly Markdown representation of this episode resource.

Summary

Illia Polosukhin, co-author of the Transformer paper, explains how decentralized infrastructure can solve the growing privacy and liability risks of centralized AI. He details a vision for a decentralized cloud using secure enclaves to enable private model training and verifiable inference.

Topics

Transformer Architecture
Decentralized AI
Confidential Computing
Data Privacy
Blockchain
Near AI
Machine Learning
Formal Verification
AI Agents

Highlights

Main idea: Centralized data collection is transitioning from a gold mine to a legal liability due to global regulations like GDPR
Practical takeaway: Developers can use decentralized clouds and secure enclaves to run software on user data without ever actually seeing or possessing that data
Failure mode: The lack of transparency in 'open weights' models leads to wasted research resources as scientists struggle to reproduce results without knowing the full training process
Main idea: A tokenized incentive model can reward website owners for contributing data to training sets while maintaining strict privacy through encryption
Practical takeaway: Trust in AI agents requires formal verification at the invocation layer to ensure composable guarantees as different AI systems call one another

Chapters

1:00 The Transformer Legacy and Near AI: Illia discusses his transition from developing the Transformer architecture at Google to founding Near AI to tackle the challenges of scalable, private intelligence.
6:00 The Shift from Data Assets to Liabilities: An exploration of how increasing global privacy regulations make holding user data a significant risk for developers, necessitating a shift toward pushing software to the user.
10:40 Confidential Computing and Secure Enclaves: How new hardware generations and NVIDIA-enabled secure modes allow for running Docker containers in environments that protect both user data and model weights.
20:30 The Decentralized Cloud Architecture: A look at the middle layer where users contribute data for processing and model providers provide intelligence, all while maintaining a zero-knowledge relationship with the cloud provider.
25:35 Tokenized Incentives for Data and Training: How a token-based system can facilitate the crawling of the web and reward content owners for contributing data to model training without compromising privacy.
35:15 The Problem with Open Weights vs. Open Process: Critiquing the current state of open-weight models and proposing a shift toward open training processes that allow for monetization and reproducibility.
58:50 Formal Verification and AI Agent Composability: The necessity of formal verification to ensure that as AI agents begin calling other AI services, they adhere to predictable, verifiable, and secure properties.