NEAR founder: AI centralization is an "existential threat to Web3"

Why Polosukhin left Google to pursue "user-owned AI"

Illia Polosukhin, co-founder of NEAR Protocol and CEO of NEAR Foundation, says he co-wrote the "Attention Is All You Need" paper at Google in 2017 because he wanted machines to understand human language. He says Google had the compute, the engineers, and the original research but still could not ship a product like ChatGPT because disruptive AI products threatened the company's own search and ads revenue. He left to start what became NEAR AI, saying "the reality is it's really hard to make disruptive products inside such a big organization."

His core argument against corporate AI is structural: a company's incentive is to make itself successful, not the user. He defines user-owned AI as the opposite — AI "aligned with you," comparing it to "almost like your mom... always aligned with your success" even when it disagrees with the user.

What happens to your data inside centralized AI

Polosukhin says that with frontier AI providers, prompts and outputs sit in plaintext in company databases, and that even unchecking "don't train on my data" may not stop firms from running "synthetic data generation" that absorbs business and personal information into future models. He says agentic AI compounds this risk because agents may feed emails, documents, and internal knowledge to providers, effectively letting a company's full operations "be reconstructed from the data."

He specifically warns against using mainstream chatbots for finance, health, and legal matters:

"Finance, house, legal, because everything you ask OpenAI, it's not privileged." — Illia Polosukhin

He cites a legal case where OpenAI logs were reportedly subpoenaed, arguing this makes such data discoverable by an opponent in a dispute.

NEAR's pitch: private AI plus private money

Polosukhin describes NEAR today as "a vertically integrated stack of blockchain and AI tech focused on privacy," combining a blockchain, an agent-focused markets layer, and private AI infrastructure. He says NEAR AI is already integrated as a provider inside Brave's Leo AI and inside Venice, a private AI app, and that near.com now functions as an on-chain finance app that will add AI agents.

He draws a direct analogy to Zcash, describing it as effectively private Bitcoin, and says NEAR aims to offer "encrypted ETH" — private smart contracts, private trading, and privacy across what he says is now several hundred supported assets, including 500 newly added tokenized stocks, lending, yield, and perpetuals.

How verifiability is meant to replace trust

Asked how a non-technical user could trust NEAR AI over ChatGPT, Polosukhin points to cryptographic verification rather than a simple assurance. He says NEAR can provide a signature confirming which model ran, a hash of the prompt, the hardware used, and the system prompt, comparable to a browser's padlock icon. He notes NEAR currently uses open-weight models like DeepSeek because their openness allows this kind of verification, something not possible with closed models like OpenAI's.

He frames the risk from centralized AI in political terms, warning that a single company controlling a chatbot's system prompt could subtly influence large populations:

"Being a front-end engineer for ChatGPT is probably the simplest way" to influence a billion people's views. — Illia Polosukhin

He concedes "we don't know" whether this is currently happening, saying there's no verifiability today, which is the gap he says NEAR's cryptographic approach is meant to close.

The NEAR token's role, and an open question on sustainability

Polosukhin ties NEAR's economics to privacy access, saying holding NEAR currently grants some amount of private inference, and that as more people hold small amounts, the token becomes scarcer, funding more compute. He argues this ownership model avoids the trap he says OpenAI fell into, moving from a nonprofit mission to a for-profit structure under shareholder pressure. He did not give a specific price target, only reasoning through supply and demand qualitatively.

He acknowledges a direct challenge to his philosophy: Anthropic-style safety arguments that private, unsupervised AI is dangerous. He calls this "a philosophical debate" and suggests companies retaining data "so they can check for safety" is partly a convenient justification, without fully resolving the disagreement in the conversation.

Disagreement with AI doomerism, and bio/cyber risk

Responding to a described view from a guest on another episode, Jordi Alexander, that Silicon Valley figures around Sam Altman have long warned of AI catastrophe without it materializing, Polosukhin says he is "generally very optimistic" and argues doomerism functions well as marketing because it resists nuanced rebuttal. He addresses bio-risk by arguing physical-world production of dangerous materials still faces existing checks and balances, and cyber-risk by arguing AI-assisted defense and formal verification are advancing alongside AI-assisted attacks.

A next frontier: AI-designed personalized medicine

Asked about the next major shift beyond transformers and crypto, Polosukhin points to personalized medicine, describing a future where blood tests, DNA, and scans are fed into AI to generate a specific mRNA, CRISPR edit, or peptide for an individual's condition. He says people have already used this approach to address cancer and long COVID cases that large pharmaceutical companies ignore as unprofitable long-tail problems, though he notes it currently costs significant money and that scaled production — something he compares to "the AWS of medicine" — does not yet exist. He estimates reasonable-scale deployment could arrive within "the next couple of years."