NEAR Co-founder Illia Polosukhin: Verifiable AI Turns Judgments into Code
NEAR Protocol co-founder Illia Polosukhin stated that on-chain "code is law" could only cover simple calculations in the past, while real transactions require judgments: whether funds come from theft, if trusted sources have reported, and whether a jurisdiction is enforcing laws seriously or abusing power. Previously, this relied on manual approval; now it can be assessed using known inputs and known model deployments, with results providing proof after evaluation.
Polosukhin is one of the eight authors of the 2017 Google Brain paper "Attention Is All You Need" and later co-founded NEAR with Alexander Skidanov. By 2026, his public stance is that the main users of blockchain will become AI agents rather than humans. NEAR AI Cloud integrates inference into Intel TDX and NVIDIA confidential GPU enclaves, with memory closed to hosts, operators, and NEAR itself, returning hardware signature proofs with each request, binding the code and model run at that time.
The verification path is public: pull proof reports from gateways or directly connected interfaces, then verify GPU evidence, TDX quotes, signature keys, and nonces using NVIDIA and Intel verifiers. Intel Trust Authority is connected as an independent verifier from operators. Open-weight models like Qwen, DeepSeek, and GLM run within enclaves; cutting-edge closed-source models are forwarded through gateways also within TEE, with providers only seeing prompts from the gateway, not binding user identities.
The payment layer converts staking rewards into inference quotas: the principal remains redeemable, and rewards pay for private inference without needing to bind cards. The same stack also supports resident agents and private agent products on the enterprise side. NEAR AI Cloud has been integrated as an inference provider into OpenRouter, offering verifiable confidential inference through privacy gateways.
"Code is law" is rewritten here as: writing compliance judgments as reproducible model calls, rather than manual approvals after multiple signatures. Inputs, model hashes, and hardware proofs become on-chain verifiable evidence packages. What remains unresolved is whether the model itself is fair, if training data is contaminated, and whether jurisdiction standards can be compressed into a single forward computation.
Buy-side protocols and agent wallets need to write freezing, sanctions, and source reviews into contracts; sell-side provides verifiable inference from NEAR AI and enclave computing power. Funding shifts from per-use cloud billing to staking rewards for quotas, with event-driven actions coming from agents starting real transfers. Beneficiaries are the inference layer that can provide proof and the NEAR ecosystem; those under pressure are custodians still relying on manual compliance counters and centralized interfaces that cannot prove "this is the model that ran."
Source: Public Information
ABAB AI Insight
Polosukhin transitioned from a Transformer author to an L1 founder, aiming to turn the attention mechanism into an industry and then create a settlement layer needed by the industry on his own chain. NEAR initially focused on sharding and developer experience, later shifting the main narrative to agents, intent transactions, and confidential inference, allowing a NEAR account to sign assets on other chains without traditional cross-chain bridges. This transforms the identity of the 2017 paper's authors into the default account layer for AI agents.
The capital path subsidizes inference with token staking rewards rather than raising another round of cloud business. NEAR locks at validators, turning rewards into confidential inference quotas, which are then fed to agents to execute cross-chain intents. The motivation is that once agents hold assets, they must prove which model and chip were used for the judgment, or the contract cannot be automatically released. Resources are transferred between protocol inflation, stakers, and GPU enclaves.
Similar positions include Chainalysis and TRM selling on-chain labels to exchanges, and banks using manual suspicious transaction reports to do the same. NEAR aims to move label generation from analytics companies into verifiable inference. The industry phase is shifting from "writing transfers as scripts" to "writing discretion as provable model calls," still expanding and not yet mastering regulatory pricing power.
This is a technical replacement. Manual approval nodes are replaced by model evaluations within enclaves. The mechanism is that simple Boolean logic cannot cover disputes over illicit funds and jurisdictions, and contracts cannot pause to wait for lawyers, so judgments are encapsulated as reproducible inferences, with proofs replacing signatories. Replacement does not eliminate politics; it merely moves politics from chat rooms into model cards and proof chains.
ABAB News · Cognitive Laws
- Once judgments can be proven, approval positions will be replaced by code.
- For laws to go on-chain, model hashes must first be written into contracts.
- To automate complex worlds, inference must first be verifiable.