ILLIA POLOSUKHIN Co-author of the paper that powers every AI you use 173,000+ citations on "Attention Is All You Need" Co-founder, NEAR Protocol CEO, NEAR Foundation Building user-owned AI $500M+ raised $20B+ peak market cap Code on Mars Code in the Arctic Building Open Superintelligence ILLIA POLOSUKHIN Co-author of the paper that powers every AI you use 173,000+ citations on "Attention Is All You Need" Co-founder, NEAR Protocol CEO, NEAR Foundation Building user-owned AI $500M+ raised $20B+ peak market cap Code on Mars Code in the Arctic Building Open Superintelligence
Illia Polosukhin, co-founder of NEAR Protocol
NEAR FOUNDATION CEO
Profile / Founder & Engineer

Illia
Polosukhin

The man who wrote "Attention Is All You Need" - then left to prove it.

Before ChatGPT had a name, before OpenAI was a household word, Illia Polosukhin and seven colleagues at Google wrote a paper that changed the architecture of intelligence. Then all eight of them walked out the door.

Transformer Co-author NEAR Protocol User-Owned AI Ukraine
173K+ Paper Citations
8 Co-authors (all left Google)
$500M+ NEAR Ecosystem Raised
$20B+ Peak Market Cap
2018 NEAR Founded

The paper that ate the world

In June 2017, a group of eight researchers at Google published a 15-page paper and titled it after a Beatles lyric. "Attention Is All You Need" - arXiv:1706.03762 - was presented at NeurIPS that December. It introduced the Transformer architecture: no recurrence, no convolutions, just attention mechanisms all the way down.

Within two years, GPT-2 ran on it. Within three, BERT did too. Today, if it understands language - ChatGPT, Gemini, Claude, LLaMA, Mistral, Grok - it's a Transformer. The eight names on that paper essentially wrote the operating system that modern AI is running. Illia Polosukhin was one of them.

Then, almost immediately, all eight left.

"For-profit can very quickly become for-control."

- Illia Polosukhin, 'Attention' Substack, 2025

The exodus from Google after that paper is one of the stranger footnotes in tech history. The authors dispersed into the industry like seeds in wind. Polosukhin took a different angle than most. While others went deeper into AI companies, he went sideways - into blockchain. Not because he doubted AI, but because he'd already seen the problem it was building toward.

He started NEAR in 2017 with his friend and fellow engineer Alexander Skidanov. The first idea was NEAR.ai - an AI startup focused on program synthesis, teaching machines to write code. It was a small operation in a WeWork in San Francisco, two engineers working on a problem they found fascinating. Then they realized: the infrastructure to make AI trustworthy didn't exist. So they'd build it.

By 2018 they'd pivoted. NEAR.ai became NEAR Protocol - a layer-1 blockchain built for scale, speed, and developer accessibility. The AI DNA stayed buried in the architecture. Seven years later, it's become the thesis.

In November 2023 Polosukhin took the title of CEO at NEAR Foundation. The timing was deliberate. The AI boom was reaching temperatures that made him uncomfortable - not because of what AI could do, but because of who would own it. Corporate AI, he argued in his Substack newsletter (launched August 2025, named "Attention" - because of course it is), doesn't optimize for users. It optimizes for shareholders. And those two things are not the same thing.

His current project: NEAR AI. Privacy-preserving AI inference in hardware-backed confidential enclaves. Private Chat that even NEAR's servers can't read. IronClaw, a Rust-based agent harness that hit 12,000+ GitHub stars in its first year. Partnerships with Brave, OpenMind, and Phala, reaching a combined 100 million users. He calls it "User-Owned AI." The pitch is simple: your AI should work for you, not the company that sells it to you.

He grew up in Ukraine, taught himself to code as a teenager, and took freelance jobs before university. He's still coding. His GitHub username - ilblackdragon - predates his Google career, his papers, and his blockchain. It's the one thread that runs through everything.

arXiv:1706.03762
"Attention Is All
You Need"

The paper was submitted June 12, 2017. It dispensed with recurrence - the way neural networks had been thinking sequentially, like reading one word at a time - in favor of attention mechanisms that process relationships across an entire sequence simultaneously.

The result: you could train much faster, on much more data, with much more parallelism. A 100-million-parameter model could beat the state of the art in translation in 3.5 days on 8 GPUs. Today's models run on tens of thousands of GPUs - but the architecture is the same.

The title - a riff on "All You Need Is Love" by The Beatles - was the kind of in-joke that academic papers rarely allow themselves. The paper is cited more than 173,000 times. For context: most highly-cited papers in history reach tens of thousands. This one, in eight years, has climbed into the top tier of scientific papers ever published on any topic.

Attention(Q, K, V) = softmax(QKT / √dk) × V

The scaled dot-product attention formula from the paper. Every ChatGPT response runs through a version of this.

Eight Equal Authors

Names listed in randomized order - the paper's note on equal contribution.

Ashish Vaswani Noam Shazeer Niki Parmar Jakob Uszkoreit Llion Jones Aidan N. Gomez Łukasz Kaiser Illia Polosukhin

All eight left Google after the paper was published. Each went on to found or lead companies that shaped the AI industry.

173K+ Citations
2017 Published
NeurIPS Presented

Built on Transformers

GPT series, ChatGPT, BERT, Claude, Gemini, LLaMA, Mistral, Grok, DALL-E, Stable Diffusion, Sora, Vision Transformers (ViT)

Performance (2017)

28.4 BLEU on EN-DE translation - 2 points above prior state of the art. 41.8 BLEU on EN-FR. Trained on 8 P100 GPUs in 3.5 days.

Model Size (2017)

~100 million parameters in the base model. GPT-4 (2023): estimated 1.8 trillion. Same architecture. Different scale.

From freelance teenager
to building the future of AI

He didn't start in Silicon Valley. He started in Ukraine, writing code for hire before he had a degree, before he had a title, before anyone knew what a Transformer was.

Early 2000s
Starts coding as a teenager in Ukraine. Takes freelance software projects - early self-starter instinct that never quite left.
2014
Joins Google Research as Engineering Manager. Works on machine learning and natural language processing at Google Brain.
June 2017
Co-authors "Attention Is All You Need" with seven Google colleagues. Paper introduces the Transformer. Presented at NeurIPS 2017.
2017
Leaves Google. All eight co-authors depart. Co-founds NEAR.ai with Alexander Skidanov - initially focused on program synthesis (AI that writes code).
2018
Pivots NEAR.ai to blockchain. NEAR Protocol is born in a WeWork office in San Francisco. The problem: decentralized infrastructure for trustworthy AI.
2019-2022
NEAR raises $12.1M, then $21.6M led by Andreessen Horowitz. Mainnet launches 2020. Token hits $20.42 all-time high in January 2022. Ecosystem raises $500M+.
Nov 2023
Named CEO of NEAR Foundation. Takes formal leadership of the Swiss non-profit that stewards the NEAR ecosystem.
2025
Launches "Attention" Substack. NEAR AI products ship: Private Chat, NEAR AI Cloud. Co-authors "Decentralized Confidential Machine Learning" paper with Skidanov.
2026
IronClaw (Rust AI agent harness) crosses 12,100+ GitHub stars. Publishes vision for agent harnesses as the new operating system layer.

Kharkiv Polytechnic Institute

National Technical University of Kharkiv, Ukraine. Computer Science and Engineering. Before that: self-taught, freelance, and relentlessly curious.

Why blockchain after the Transformer?

NEAR started as an AI company. The original idea - program synthesis, machines writing code - required a foundation that didn't exist: open, verifiable, user-controlled infrastructure. So they built the infrastructure. Eight years later, Polosukhin is using that infrastructure to bring AI home.

Alexander Skidanov

NEAR Protocol's other half. Former Microsoft engineer. Co-authored "Decentralized Confidential Machine Learning" with Polosukhin in 2025. The two have been building together since 2017.

NEAR Protocol: the blockchain
that started as an AI company

$500M+ Total Ecosystem Raised
$20.42 All-Time High (Jan 2022)
2020 Mainnet Launch
100M+ NEAR AI Partner Users

NEAR Protocol is a proof-of-stake, sharded blockchain built for developer accessibility and scale. Written in Rust. Consensus via Nightshade sharding. Smart contracts in Rust, AssemblyScript, and JavaScript.

The positioning has shifted over the years. In 2020 it was "fast, cheap, developer-friendly." By 2024, under Polosukhin's leadership, the narrative settled: NEAR is "the blockchain for AI." Chain abstraction - the ability to work across multiple blockchains without the user knowing which one they're on - became the flagship technical thesis.

The Foundation, a Swiss non-profit, raised $12.1M in 2019 (Metastable Capital, Accomplice), $21.6M in 2020 (led by Andreessen Horowitz), and additional rounds hitting $500M+ total. At the 2022 peak, the NEAR token was worth $20.42 and the ecosystem's market cap hit over $20 billion.

In 2024, Polosukhin cut Foundation headcount by ~40% and refocused the organization. Leaner. More ecosystem-oriented. More honest about where the value actually sits.

Proof-of-Stake Nightshade Sharding Rust Chain Abstraction NEAR Intents AI-Native Swiss Foundation
NEAR AI Products (2025-2026)
Private Chat

End-to-end encrypted AI chat. Not even NEAR can read it. Hardware-backed confidential enclaves.

NEAR AI Cloud

Privacy-preserving AI inference. Your model runs in a trusted execution environment.

IronClaw

Rust-based AI agent harness. Security and privacy by design. 12,100+ GitHub stars. Polosukhin's bet that AI agents need secure, open infrastructure.

"In today's attention economy, the user is the product. Web2 and centralized AI companies are mining our attention and data for profit."

- Illia Polosukhin, 'Your Data Can and Will Be Used Against You', December 2025

The philosophy behind
the infrastructure

He's not anti-AI. He's not anti-corporate by default. He just thinks there's a version of the future where AI serves the people who use it, and a version where it doesn't. He's trying to build the first one before the second one becomes the default.

User-Owned AI

AI systems should be owned and controlled by users, not the corporations that deploy them. Corporate AI optimizes for profit, not user benefit - and those objectives diverge over time. Cryptographic guarantees, not terms of service.

Blockchain + AI as One Stack

The combination isn't a marketing angle. It's architectural: blockchain provides verifiable ownership, transparent computation, and economic rails. AI provides intelligence. Together, you get AI that's both capable and accountable.

Explicit Governance

"Implicit governance is a slow killer for blockchains." He believes every organization needs clear, transparent, on-chain governance - and that AI agents may become the most reliable mechanism to enforce it, because they can't be bribed.

Open Superintelligence

His long-term target: superintelligent AI that is open-source, user-owned, and not controlled by any single corporation or government. Not sci-fi. A design goal that shapes every product decision NEAR AI makes today.

Agent Harnesses as OS

Graphical interfaces are going away. Natural language-driven AI agents will replace them. The new operating system isn't Windows or macOS - it's the harness that coordinates those agents. IronClaw is his prototype.

AI Risk - The Human Version

He's skeptical of existential AI doom narratives focused on machine consciousness. His concern is more immediate: very human actors using AI as a tool of power and control. The risk isn't the robot - it's the organization that owns the robot.

What Illia Polosukhin
actually says

"
Corporate-owned AI is getting richer and more powerful, optimizing primarily for profit rather than user benefit.
'Attention' Substack, August 2025
"
For-profit can very quickly become for-control.
'Attention' Substack, August 2025
"
Implicit governance is a slow killer for blockchains.
'Unwritten Rules of the Game', October 2025
"
It's impossible to remove self-interest in people, and so any governance system must operate with this assumption as a baseline.
'Unwritten Rules of the Game', October 2025
"
In today's attention economy, the user is the product. Web2 and centralized AI companies are mining our attention and data for profit.
'Your Data Can and Will Be Used Against You', December 2025
"
AI agents will replace most interfaces and change computing as we know it.
'Attention' Substack, 2025

Ten things about Illia
you didn't know you needed

  • 01 He started coding as a teenager in Ukraine doing freelance work - before a degree, before Google, before the paper. The builder instinct came first.
  • 02 The title "Attention Is All You Need" is a deliberate play on The Beatles' "All You Need Is Love." It was 2017 and Google researchers were feeling playful. They had no idea the paper would be cited 173,000+ times.
  • 03 All eight co-authors listed their names in randomized order. Equal credit. When the paper became one of the most important in AI history, nobody could claim the head of the table.
  • 04 All eight left Google. Each went on to found or lead significant AI companies. It's one of the most consequential talent dispersals in tech history.
  • 05 His GitHub username - ilblackdragon - predates his Google career and his famous paper. It's the most stable identity he has.
  • 06 NEAR Protocol started as NEAR.ai - an AI startup. It became a blockchain because they needed the infrastructure that didn't exist. Now it's circling back to AI. The whole arc looks planned in retrospect. It wasn't.
  • 07 His code is in the GitHub Arctic Code Vault - physically stored in an abandoned coal mine in Svalbard, Norway, preserved for 1,000 years. The Transformer architecture will outlast most things humans build.
  • 08 He's a NASA Mars 2020 mission contributor. His code made it to Mars. He is still not done building things for Earth.
  • 09 He named his Substack newsletter "Attention" - a direct callback to the paper. 2,000+ subscribers. Published weekly. The man commits to a bit.
  • 10 IronClaw - his Rust-based AI agent harness - hit 12,100+ GitHub stars in its first year. For a new infrastructure project, that's extraordinary signal. Developers are paying attention.

What he's been building
lately

MAR 2026

Agent Harnesses as the New OS

Published his most-read Substack essay arguing that AI agent harnesses - like IronClaw - are the new operating system. Natural language in, coordination out. Graphical interfaces are a dead end.

DEC 2025

NEAR AI Cloud Launch

Private Chat and NEAR AI Cloud ship with partners Brave, OpenMind, and Phala. 100M+ combined users. Privacy-preserving AI inference using hardware-backed confidential enclaves.

AUG 2025

'Attention' Newsletter Launches

Substack newsletter "Attention" debuts with Illia's thesis on user-owned AI, blockchain governance, and the path to Open Superintelligence. Grows to 2,000+ subscribers.

MAR 2025

Decentralized Confidential ML Paper

Co-authored with Alexander Skidanov. Explores the technical foundations for user-owned AI via confidential computing and blockchain infrastructure. The research arm of NEAR AI.