Illia Polosukhin On The Future Of AI, Crypto, And How Near Is Going To Win

Summary

• Illia says NEAR Intents processes about $4 billion monthly, with protocol fees converted into NEAR purchases. near.com contributes 10–20% of volume, depending on the week; perps and Earn also direct fee shares toward buybacks. His framing is that “every product is effectively generating revenue and then buying back NEAR,” although he supplies no cumulative buyback figure.
• NEAR’s strategy now includes building the consumer finance product itself. Launched in February at NEARCON, near.com combines cross-chain assets, swaps, payments, stocks, perps and Earn with biometric login and a private shard running inside secure enclaves. Illia describes the broader business as “a vertically integrated stack for the AI and blockchain future.”
• Illia thinks the larger opportunity is bringing $20–40 trillion of business commerce on-chain through AI agents. Agents would find counterparties and negotiate, while blockchain supplies escrow, settlement and dispute infrastructure for businesses that do not trust each other. “AI is a front end, blockchain is a back end”: the target is commercial relationships burdened by legal friction, while consumer grocery purchasing is already highly optimized.
• Illia argues that agents handling money need verifiable inference and explicit spending limits. An arbitrary model endpoint—especially a router or free model, but also some actual providers—could substitute models or return malicious instructions: “you have no idea what’s running behind it.” Near ID’s stated answer is signed GPU, model and software provenance; a financial agent’s policy engine requires approval before exceeding its permissions.
• A $3.8 million exploit of a partner-built BSC connection exposed gaps in NEAR’s integration testing. Illia reports systems restored within an hour and “100% recovery” after identifying the attacker, while acknowledging that the connection had been insufficiently tested. NEAR is accelerating verification across interacting contracts; he claims a 100-fold cost reduction compared with what he phrases as “using OpenAI, like Sol or Astra”—a comparison that is unclear in the transcript.
• Avi’s central commercial challenge is that strong technology alone does not bring users or distribution. Illia says NEAR now works “sales first,” citing wallet integrations, Venice’s end-to-end encrypted mode, Brave’s agent using NEAR as an option, and non-crypto, non-AI design partners. The concrete agent-commerce example is an event company monetizing its venue database through bookings: the intended business model is “paying for outcome.”

Deep dive

1. The AI ambition preceded NEAR, and compute determined its pace

• Illia’s case for decentralized money starts with his grandparents’ Sberbank savings becoming worthless after the USSR dissolved. In post-Soviet Ukraine, he recalls bread prices rising, he thinks, 100× within roughly a year: “the currency didn’t work yet.” Separately, competitive programming convinced him at 14 that “machine should write this code.”

• A Google experiment around 2012–2013 supplied the next opening: an unsupervised model learned the concept of a cat from unlabeled internet pictures. Illia identifies it as a paper by Jeff Dean and Andrew Ng, trained across, he thinks, about 600 machines. He attributes the broader neural-network breakthrough to Geoffrey Hinton and Yann LeCun getting multilayer networks and backpropagation to work, with the math and compute catching up. Illia wanted the equivalent for language; animals could see, but people tested intelligence by asking questions. He joined Google to work on question answering.

• At Google, the obstacle was algorithmic: recurrent neural networks read one word at a time, trained slowly and could forget a long document’s beginning. His early answer projected words into vectors and summed them by paragraph, producing a fast, simple question-answering model on google.com around 2015–2016. Transformers extended that parallelism so GPUs could train larger models on more data.

• To explain attention’s layers, Illia uses “cat sat on a wall”: first connect cat with sitting, then connect that combined concept with the wall. He describes KV cache as retaining processing so later questions can reuse it. OpenAI’s next move was to “scale the hell out of this,” through GPT-1 and GPT-2 toward ChatGPT; the architecture also worked for images, sound and video.

2. A micropayments problem became a two-year blockchain build

• In 2017, Illia and co-founder Alex were training transformers on GitHub, aiming to replace separate applications and SaaS with computers that simply did what users requested. Limited compute constrained them. They recruited students in Eastern Europe, Southeast Asia and China to write code and explanations, then discovered that paying contributors across fragmented banking and payment systems was its own problem.

• Crypto was supposed to solve global micropayments, but in 2018 transaction fees exceeded what they were paying students. The intended detour was an easy blockchain with cheap transactions whose prices would not rise with usage. Sharding, account abstraction and WebAssembly followed from those requirements. “It’s gonna be six months,” Illia recalls; an August-to-January pitch became an October 2020 launch, a little over two years later.

• The underestimate was permissionless coordination, despite both founders’ distributed-systems experience. Sharding raised routing and data-availability problems; they implemented, he thinks, five consensus designs. Illia says they broke liveness in the original Avalanche design, noting that Avalanche has since changed substantially. Their Whiteboard Series continued that approach: engineers explaining mechanisms deeply enough that flaws, including Loom’s exit game, could emerge on camera.

3. Chain abstraction now connects a consumer app to NEAR purchases

• Avi challenges whether infrastructure can itself be consumer friendly. Illia’s answer starts with interviewing roughly 300 crypto companies in 2019 and testing onboarding themselves: buying a CryptoKitty from a new machine took 92 steps and 12 hours, with what he describes as 99.99% drop-off. Better wallets, smoother onboarding and meta-transactions could remove requirements such as acquiring gas before using an application.

• That evolved into chain abstraction: restricting users to one chain resembled making Netflix viewers pay another currency because the next video lived on another server. With Intents, users specify an outcome—10 ZEC at a shielded address, funded with Solana USDC—and solvers choose the route. Asked whether this is just a bridge, Illia says solvers may use LayerZero, Binance or whatever achieves the outcome.

• After waiting for others to build the consumer experience, NEAR launched near.com in February at NEARCON. Illia describes Face ID or fingerprint login, assets including Bitcoin and Zcash, payments, swaps, stocks, perps and Earn without juggling wallets or gas. Its private NEAR shard runs inside secure enclaves: encrypted transactions decrypt and execute there. He calls trading privately and paying out across ecosystems “privacy in motion.”

• Illia puts Intents volume at about $4 billion monthly; near.com contributes 10–20%, depending on the week. The protocol’s take varies by asset, with less charged on stablecoins, and fees convert into NEAR. Perps and Earn also contribute fee shares toward buybacks. Avi presses for cumulative NEAR purchases, but Illia supplies no figure, pointing instead to revenue.near.org: “every product is effectively generating revenue and then buying back NEAR.”

4. Agent commerce targets the friction between businesses

• Avi’s bigger objection: give an AI agent an Ethereum private key and it should already transact, regardless of the underlying chain—so what makes AI × blockchain distinctive? Illia accepts that commerce can happen without blockchain. His proposed division is “AI is a front end, blockchain is a back end”: agents find counterparties and negotiate, while markets execute transactions.

• The concrete ambition is an agent arranging delivery of 50 tons of steel to the Port of LA, supported by escrow and dispute resolution. Illia argues that businesses’ lack of trust creates lawyers, long agreements and redlining; blockchain rails could reduce that friction. He initially describes a $20–50 trillion annual market, later describing $20–40 trillion of economy he thinks AI can bring on-chain.

• He draws the boundary at business relationships, rather than grocery buying: consumer purchasing is already highly optimized. NEAR’s agent market begins with digital work because outcomes are easier to verify, then aims at broader commerce. Hiring agents also introduces a privacy requirement: an agent running on somebody else’s machine could expose your trading relationships and private information.

5. Inference provenance and spending policies address different agent risks

• Illia’s inference warning is concrete: an API endpoint could substitute a smaller or quantized model, or fabricate a tool response instructing an agent to run a Bash command that infects the machine and steals private keys. He says this risk applies especially to routers and free models, but also to some actual providers. Avi expects OpenAI safeguards; Illia says the vulnerability sits with the agent and whichever endpoint it uses. With Hermes pointed at an arbitrary API, “you have no idea what’s running behind it.” He also warns that browsing the internet and reading news can be another attack route, and says NEAR has an attack bench for this.

• Illia names Near ID as the source of signed provenance identifying the GPU, exact model and software hashes, plus every node routing traffic. He calls it “full end-to-end verifiability and privacy.” Decentralizing GPU access is another benefit. He also sees a compute-market opportunity in replacing today’s bilateral supplier haggling—availability in two months, competing price quotes—with blockchain-based markets.

• The planned near.com financial agent adds a separate control: a policy engine constrains execution, with explicit approval needed outside the policy. Illia compares it to hiring a money manager with restrictions, rather than handing over every dollar. Users could authorize even a memecoin strategy, but the agent should not choose one unilaterally. The same policy system is also a developer product for other applications.

6. The BSC exploit accelerated verification of the whole system

• Discussing the previous week’s exploit, Illia says a BSC connection used by NEAR Intents lost $3.8 million. Shield detected the activity, the team paused and patched it, and systems were back online within an hour. He describes Shield as scanning NEAR and other chains for anomalies and suspected laundering. He says the concern started with the Bitget hack, where North Korea was effectively trying to move money around.

• Shield’s data then helped identify the attacker, Illia says: “we got all the money back,” with “100% recovery” and an on-chain apology. Avi asks whether NEAR engineers built the link. It came from an ecosystem partner, but Illia concedes that it “wasn’t tested to the level that we should be testing on.”

• The response is to accelerate formal verification across the whole interacting system of contracts, alongside stronger monitoring. Illia says work had already been underway for months, claims number-one results on multiple benchmarks, and claims a 100-fold verification-cost reduction compared with what he phrases as “using OpenAI, like Sol or Astra.” The comparison is unclear in the transcript.

7. Distribution starts with customers and charges for outcomes

• Avi’s go-to-market challenge is blunt: good technology has failed before, even when Turing Award winners built it. Illia says NEAR now works “sales first,” finding customers or design partners before building.

• Distribution includes Trust Wallet, Ledger, LI.FI and SwapKit; Shield partnerships include Binance and ZeroShadow. Illia says Venice uses NEAR for its end-to-end encrypted mode, while Brave uses NEAR as an option for its agent. Robotics companies are also using NEAR on the back end for inference and agents.

• The non-crypto, non-AI example is Philosophia’s venue-finding agent. Its accumulated internal database can answer a request for an institutional side event in Singapore for 50 people, replacing an email and an employee searching the records. The offering combines that database with an agent and charges for the outcome—actually booking the venue. Illia presents this as automating an existing commercial relationship.

• The product he is most excited about is the financial agent operating on users’ existing near.com assets. The longer-term ambition is an agent market expanding from digital tasks into paid real-world actions: “paying for outcome,” and ultimately moving “real GDP” on-chain.