From Transformers to Intents

Why an AI Researcher Built Blockchain Infrastructure

TL;DR: In 2017, Illia Polosukhin (@ilblackdragon) co-authored the paper that invented transformers. The architecture powers ChatGPT, Claude, and every major AI system. Months later, he left Google to build an AI company (@near_AI). That company needed to pay workers around the world. Traditional payment rails failed. Blockchain solved the problem. Seven years later, the same insight drives @NEAR_Intent. At NVIDIA GTC 2026, he argued that classical economic models need revision for agentic participants.

The same person who helped create the AI architecture that powers ChatGPT is now building the coordination layer for the AI Economy that AI agents will use to transact.

The Paper That Changed Everything

Before there was ChatGPT, before there was Claude, before there was Gemini, there was a paper.

“Attention Is All You Need,” published in June 2017 by eight Google researchers, introduced the transformer architecture. The “T” in GPT stands for transformer. Every major language model, every AI assistant, every system that seems to understand and generate human language traces back to this paper.

Illia Polosukhin was one of those eight researchers.

At NVIDIA’s GTC conference in March 2024, Jensen Huang brought seven of the eight authors on stage. “Everything that we’re enjoying today,” Huang said, “can be traced back to that moment.”

Illia is the guy sitting next to Jensen Huang

The “Transformer 8,” as they’ve become known, have all left Google. Noam Shazeer co-founded Character.AI. Aidan Gomez co-founded Cohere. Lukasz Kaiser joined OpenAI. Each took the transformer architecture in different directions.

Polosukhin took it somewhere unexpected: blockchain.

The Payment Problem

The path from transformers to blockchain wasn’t ideological. It was practical.

“We had people around the world, mostly students from China, Russia, Ukraine, etc., who were doing kind of small tasks for us,” Polosukhin explained in an interview with Sifted. “Paying these people was actually a pretty painful process. Sending money from the US to China is complicated, Russia and Ukraine as well.”

After leaving Google in 2017, Illia and Alexander Skidanov (@AlexSkidanov) started NEAR AI an AI company focused on program synthesis. They were teaching machines to write code from natural language descriptions. The technical work required data labeling at scale. Students around the world annotated examples, provided feedback, and trained the models.

The work was distributed. The payments weren’t.

Wire transfers cost $25–50 per transaction. ACH takes days and doesn’t work internationally. PayPal and traditional payment processors add friction and fees that make micropayments impractical. Sending $50 to a student in Ukraine for a day’s work shouldn’t cost $30 in fees and take a week to arrive.

They ended up looking at blockchain as a solution for their own problem.

From AI Company to Protocol

The pivot wasn’t abandoning AI. It was recognizing that AI needed infrastructure that didn’t exist yet.

@NEARProtocol launched in 2020 with a focus on usability: human-readable addresses (alice.near instead of 0x7a23…), cheap transactions, and developer-friendly tooling. Sharding for scalability, account abstraction for flexible permissions. All of it served one goal: making blockchain accessible enough that normal people could use it.

But the AI thesis never disappeared. It went underground, waiting for the models to catch up.

By 2023, they had. ChatGPT showed that transformer-based systems could handle complex tasks. Agents became plausible. The question Polosukhin had hit in 2017, how distributed AI systems coordinate and transact, suddenly mattered to everyone.

The Intents Architecture

At Solana Breakpoint 2025, Illia connected the threads:

“Before jumping into the blockchain rabbit hole, I was an AI researcher. I worked at Google Research on effectively deep learning, machine understanding. And I’m one of the co-authors of ‘Attention Is All You Need,’ which is a paper that introduced transformers, T and GPT.”

“Now in 2017, I left Google to start an AI company and we quickly realized that we actually need a blockchain so that we can coordinate and pay people around the world. That’s how we got to NEAR Protocol in 2018.

The insight from 2017 scales directly. If paying human data labelers required blockchain infrastructure, paying AI agents will require it even more.

Here’s how NEAR Intents works: users express desired outcomes. Swap this token for that one. Book this flight. Source this product. They don’t specify how. Solvers compete to fulfill the intent. Settlement happens on-chain.

The architecture has processed over $15B in volume across more than 35 chains. Trust Wallet, Ledger, Brave Wallet, Infinex, Zodl, CowSwap, and dozens of other interfaces have integrated.

Why Blockchain Was Required

The presentation at Solana Breakpoint made the point clearly: Cross-chain has been extremely painful. And if we look out, the reality is most users have been sitting on centralized exchanges kind of historically for all of the time that blockchain existed. It’s been too hard to use.

Bridges fail constantly. Gas costs confuse users. Wallets struggle to support every chain. Every blockchain became an island.

NEAR Intents solves this by abstracting execution: You can actually express the outcome you want without needing to figure out how exactly the execution’s gonna work.

The same abstraction applies to AI agents. An agent shouldn’t need to know which chain has the best liquidity for a particular swap. It shouldn’t manage gas tokens across twenty networks. It should express what it wants and let specialized solvers compete to deliver.

This is extremely important. Because solvers can source liquidity from a bunch of different sources: There’s gonna be market makers that are concentrating on Binance. There can be some prop trading firms that are taking a reverse position

→ There’s no MEV because this contract between users and the counterparty is done off-chain.

The Technical Foundation

Three pieces of infrastructure make this work.

1. Multi-party computation (MPC) handles custody of assets across different chains without centralized control. No single party holds complete keys. Threshold signatures require multiple parties to authorize transactions.
2. Chain signatures allow NEAR accounts to sign transactions on Bitcoin, Ethereum, Solana, Zcash, and +30 other chains without bridges or wrapped assets.
3. NEAR’s sharded infrastructure allows for massive throughput and private transactions — up to 1m transactions per second.

The result: Effectively providing a centralized exchange experience to self-custodial users and the rails for agentic commerce.

Expanding Beyond Trading

Trading was the first use case because the pain was most obvious. But intents extend furthe

The NEAR Intents infrastructure is not just for trading: express a complex intent in natural language, let AI agents propose solutions, settle commitments on-chain. You say where you want to go. Agents figure out how to. This then gets extended to e-commerce, to actual contractual service work. So really intents is this new paradigm to find the counterparty and get agreement with them and settle it and execute it across different avenues.

The pattern is consistent: blockchain networks work better when connected through intent-based infrastructure than when isolated behind incompatible bridges.

The AI-Native Endgame

The vision connects back to transformers.

In March 2024, Jensen Huang put Polosukhin on the GTC stage as a co-inventor of the transformer. Two years later, at GTC 2026, Polosukhin returned to present something different: a formal model for agentic markets. Not AI research this time. Infrastructure.

His argument was specific. Today’s markets were designed for humans. They’re full of irrational behavior, bias, expensive legal overhead, and rent-seeking middlemen. When AI agents transact, they optimize on outcomes. They don’t need brokers, lawyers, or intermediaries to enforce trust. They need programmatic escrow, intent-based matching and settlement, and agentic dispute resolution. Purpose-built primitives, not adaptations of existing systems. The implication: classical economic models, built on assumptions about human rationality and human friction, need revision for agentic participants.

The GTC talk also introduced IronClaw, a secure open-source agent runtime built to protect users from prompt injection and data leakage. Agents run inside Confidential Virtual Machines (CVMs), inference happens inside hardware-enforced private enclaves, and every execution comes with cryptographic attestations users can verify. Only the user and their agent can see credentials, data, and context. This is what “user-owned AI” means in practice: not a marketing phrase, but a specific technical guarantee.

The broader point: these same primitives (agents voting to allocate budget, measuring outcomes against milestones, reallocating automatically when targets are missed) extend beyond commerce to public goods funding and government spending. The persistent problem with public institutions is lack of transparency, clear attribution, and ROI measurement. Agents can fix the accountability problem at the infrastructure level.

Polosukhin helped invent the architecture that makes AI agents possible. Then he discovered that AI systems need payment infrastructure that traditional finance can’t provide. He built that infrastructure. Now AI agents are capable enough to use it.

What NEAR is trying to build is really this unified liquidity layer where users can plug in into any kind of asset beyond blockchains into Web2.

The future he describes: users interact through any wallet (Solana, NEAR, Ethereum, Bitcoin) with any application. AI agents handle coordination. Blockchain provides settlement. The execution machinery disappears.

This is what the payment problem from 2017 pointed toward. Distributed intelligence needs trustless coordination. The co-author of transformers is building it.

Conclusion

The path from “Attention Is All You Need” to NEAR Intents isn’t obvious until you trace it.

• 2017: Polosukhin co-invents transformers at Google. 2017: Leaves to build an AI company.
• 2017–2018: Discovers that paying distributed workers requires blockchain.
• 2018–2020: Builds NEAR Protocol as usable blockchain infrastructure.
• 2020–2024: Develops chain signatures, MPC custody, intent architecture.
• 2024–2026: AI agents become capable. The infrastructure is ready. Illia returns to GTC to present the market primitives.

The same person who helped create the AI architecture that powers ChatGPT is now building the coordination layer that AI agents will use to transact.

That’s not coincidence. It’s the completion of a thesis that started when an AI researcher tried to pay students in Ukraine and found that the financial system couldn’t handle it.

Blockchain solved the payment problem. Intents scale it to AI-native commerce.

The co-founder of modern AI is building its economic infrastructure.

Sources

• Illia Polosukhin, Solana Breakpoint presentation (2025) I
• llia Polosukhin, “Market Primitives for an Agentic Economy” [S82411], NVIDIA GTC (March 18, 2026)
• “Attention Is All You Need,” Vaswani et al., Google (2017): arxiv.org/abs/1706.03762
• CNBC interview with Katie Tarasov (June 2024) Sifted interview: “Why is one of the founding fathers of generative AI all in on Web3?”
• NEAR Intents analytics: dune.com/near/near-intents

Series Context — This is Part 5 of a series on NEAR Intents and the AI economy

Disclaimer: The information presented herein has been provided by third parties and is made available solely for educational and general information purposes. Nothing in this newsletter should be construed as legal, tax or investment advice. This post might not reflect all current updates to applicable laws, regulations or guidance. The authors disclaim any obligation to update this post and reserve the right to make any changes to this post without notice. In all cases, persons should conduct their own investigation and analysis of the information in this newsletter. Where this newsletter contains links to other sites and resources provided by third parties, these links are provided for information only and should not be interpreted as approval by us of those linked websites (or any information obtained from them) that the authors of this post do not control. Always DYOR.

Copyright © 2026 NEARWEEK AG, All rights reserved.