Why AI Alongside Crypto Will Dominate the Next Decade - NEAR Founder | E193

Ilya Polosukhin defines user-owned AI as AI that works for the user, aligns with their success, and gives them control over data, assets, and choices. He expects AI to become computing’s main interface, mediating transactions among people, services, and agents. Blockchain markets could let agents coordinate and pay one another, while privacy and secure keys help preserve user control. The risk, he argues, is concentrating effort on making one model smarter rather than building an ecosystem that prevents a single company from ruling AI.

Polosukhin co-wrote “Attention Is All You Need” in 2017 with seven others; Transformer models use attention to process text in parallel. He left Google that year to teach machines to code, saying that even with more compute and engineers, it was difficult to launch products that might disrupt Google’s search and advertising businesses. ChatGPT launched in October 2022. NEAR’s path also shifted: difficulties paying international contributors for small data-labeling tasks, when Bitcoin and Ethereum fees could exceed a 15-cent payment, helped prompt the team to explore crypto as global financial infrastructure. It initially set AI aside for lack of compute and GPUs; blockchain development, expected to take six to eight months, took two years.

Polosukhin says NEAR is now building a vertically integrated blockchain and AI stack focused on privacy. NEAR intents can connect chains, enabling users to trade, interact with, and hold assets across them, and NEAR smart contracts can sign transactions on other chains. He describes the network as akin to encrypted Ethereum, with private smart contracts and financial activity such as trading, lending, yield, and perpetuals; he says it recently added 500 tokenized stocks for trading. The broader aim is to pair infrastructure for agent transactions with private AI rather than leave AI and money under one consumer-facing company.

Polosukhin warns that requests to centralized AI providers may expose sensitive information in plain text, and that opting out of training may not guarantee privacy; synthetic data could also carry information from company data into later models. He uses Claude or GPT for research on open information, but says he turns to NEAR AI inference for private personal, financial, healthcare, and crypto-agent data. NEAR AI aims to encrypt prompts and outputs and cryptographically verify the model, system prompt, and code in use. Its verification can report the model, prompt hash, hardware, and server stack; Polosukhin says experts can validate these details, and that Brave’s team validated the stack. He likens the intended user experience to a browser’s green lock.

NEAR AI runs open-weight models including DeepSeek and GLM in a private compute environment, where models can be executed and verified. Polosukhin claims it ranks first on some mathematics benchmarks with open-weight models, partly because those models cost less to run. He says NEAR AI is used in Brave’s Leo AI and Venice, with agents planned for near.com. He also warns that a company controlling the consumer AI interface could centralize money beneath it, an existential threat to Web3. His proposed alternative combines private AI with blockchain ownership: NEAR holders could gain access to private inference and own part of the system. His example of 1.2 billion NEAR and one billion people each holding one token is hypothetical; he argues that limited supply could raise the token’s price and help fund more compute.

Private AI raises safety concerns because private access could enable harmful use, Polosukhin acknowledges. He argues that risks should be assessed through probabilities, mechanisms, and specific scenarios rather than broad claims of AI doom. More capable AI could aid cyberattacks, but also strengthen defenses, including through mathematical models for formal code verification; he says he warned of AI-enabled cyberattack problems more than a year earlier and that such risks have since appeared. He also cautions that prompts could steer answers toward particular beliefs or candidates, while saying he cannot verify that this is happening today and that influence could stop during testing. Separately, Hidia imagines AI using blood tests, DNA, and scans to develop personalized mRNA, CRISPR, or peptide treatments, potentially for cancer or long COVID. He predicts scalable delivery could reach reasonable scale within a couple of years, while noting current personalized medicine is expensive. Hidia expects automation to replace work valued mainly for its outcomes, but says people will continue to value activities such as sport, nature, art, and music for the experience and individual differences.