A Positive Vision for the Future: Part 2 with Illia Polosukhin of NEAR

Episode Summary

Executive Summary: Ilya Polosukhin argues that AI coding assistants will shift software from standardized, UI-heavy products to personal automation built on demand, changing work, commerce, and governance. He envisions AI agents as private representatives that negotiate, plan, and vote for individuals, while formal verification, decentralized confidential compute, and stronger biosecurity defenses become essential in an AI-rich world.

Main Topics: AI turns software into personal automation (Priority: 5/5): Polosukhin says machine coding enables everyone to build their own tools, bypassing monolithic SaaS interfaces and replacing generic software with personalized workflows and automations. Impact on work and engineering roles (Priority: 4/5): He predicts AI will reduce time spent coding and increase time spent reviewing, architecting, and decomposing complex systems; junior roles become less about hierarchy and more about problem-solving mindset. Decentralized, user-owned AI and compute architecture (Priority: 5/5): He expects a mix of cloud, edge, and confidential decentralized compute, with personal AIs spanning devices and remaining private and aligned to individual users rather than centralized providers. Economic and consumer-market transformation (Priority: 5/5): He argues AI agents will reduce dependence on advertising and middlemen by enabling direct buyer-seller coordination, more precise demand planning, and personalized consumption. Governance through AI delegates and digital states (Priority: 4/5): Polosukhin describes NEAR’s experiments with stake-based delegates and AI copilots, moving toward AI representatives that vote on behalf of users in network-state-like governance systems. Formal verification, trust, and safety (Priority: 5/5): He emphasizes formal proofs for smart contracts and software, but notes that high-level intent remains hard to specify; probabilistic guarantees and insurance will be needed for real-world systems. Biosecurity and system hardening for AGI-era risks (Priority: 5/5): He warns that AI lowers the cost of hacking, DDoS-like abuse, and bio-risk generation, so society must redesign defenses, screen DNA synthesis, monitor air/pathogens, and assume adversarial use.

Key Arguments: AI coding assistants make software creation accessible to non-specialists, enabling personal software and automation rather than one-size-fits-all SaaS. As software becomes conversational and customizable, major platforms like Salesforce may be replaced or heavily abstracted by user-built systems. Engineering work is shifting from writing code to reviewing AI-generated code, designing architecture, and ensuring correctness/security. Junior hiring becomes less about seniority labels and more about whether someone is a strong, adaptable problem solver who can work with AI. The most immediate automation is happening at both ends: repetitive factory work and high-end white-collar work like coding and law; skilled physical trades remain harder to automate for now. The future computing interface is an AI operating system spanning glasses, phones, watches, laptops, and headphones, with one private assistant across all devices. AI agents will change markets by directly coordinating purchases, capacity planning, and logistics, reducing roles for advertisers and intermediary retailers. Governance can be improved by AI delegates that represent users continuously rather than periodic human voting, reducing principal-agent problems. Formal verification can prove specific properties of smart contracts and transactions, but user-intent alignment still needs governance and probabilistic safeguards. Biosecurity must be redesigned for a world where malicious actors can use AI to scale attacks, generate lawsuits, or attempt bioweapon design.

Data Points: Year referenced for belief that AI would replace SaaS: 2017 - Polosukhin said he and his team were already saying in 2017 that SaaS would die and AI would replace it. Timeframe for improvement in AI coding capabilities: changes every 6 months - He said his answer on AI coding would have been different six months ago and likely will be different again in six months. Current blockchain ecosystem participation: tens of thousands of active participants - He described NEAR governance at a scale where tens of thousands of active citizens make direct voting impractical. US workforce churn cited: 300% - He said U.S. workforce churn can mean hiring three people per job per year because workers keep quitting. Scale of AI delegate voting today: stake-based delegates - He described current NEAR governance as stake-based delegated voting with AI copilots for delegates, moving toward AI pilots. Number of GPUs in edge container example: 1,000 GPUs - He described an edge-compute container that could be dropped into a town with roughly 1,000 GPUs. Alternative edge-compute example size: 100,000 GPUs (rejected as too large for edge example) - He contrasted a hypothetical larger data-center deployment with the smaller, local 1,000-GPU container model. Approximate power for a couple hundred GPUs: about 1 megawatt - He estimated that a few hundred GPUs would require around one megawatt of power. Governance cadence in the U.S.: every 4 years - He used U.S. elections as an example of low-frequency compressed representation compared with always-on AI delegates. Food waste in the U.S.: 30-40% - He said 30-40% of food is thrown out because stores over-provision due to imperfect demand planning.

Pivotal Quotes: "Everyone now is able to build their personal software. Everyone now is able to build their own personal automation." — Ilya Polosukhin: On why AI coding changes software from generic applications to individualized tools and workflows. "We need formally verified software to really secure that." — Ilya Polosukhin: On the need for mathematical guarantees as AI-generated software becomes widespread and adversarial use increases. "It needs to be ours. It needs to be on our side." — Ilya Polosukhin: On the design requirement for personal AI assistants that act as private agents for individual users.

Implications: The episode argues for a future of personal AIs, decentralized compute, and AI-mediated governance, but only if software, infrastructure, and biosecurity are rebuilt for adversarial conditions and individual ownership.