Illia Polosukhin, Alex & Anton: Why Confidential AI Is The Next Big Shift

Illia Polosukhin said crypto is moving beyond a decade of largely speculative activity: over the past two years, more real businesses have begun using it as regulation and liquidity have improved. But public on-chain payments and income statements can reveal company activity before earnings calls, while AI tools that track people and correlate spending heighten personal privacy risks. He cited Zcash as the leading wave of personal privacy and a preview of business demand, arguing that businesses need not broadcast every transaction. Privacy also reduces exposure to cryptographic and post-quantum risks by keeping transactions and public keys from being publicly exposed.

Alex said confidential intents and advances in EZK and trusted execution could bring blockchain confidentiality closer to modern banks, giving people more ways to manage and hedge risk rather than necessarily displacing services such as Revolut or Robinhood. NEAR’s role remains unsettled: it has been described as a “market of markets,” a “SWIFT” across blockchains, and a home for the agent economy. Alex said the goal is to enable use cases alongside other ecosystems, not to compete in a zero-sum game. NEAR’s account model supports multiple cryptographic schemes, enabling new sign-in methods and migration without committing the network to one option; it has rolled out MLDSA 65 and plans to add schemes as they mature. One speaker warned that changing a chain’s sign-in scheme within a year could require a massive transition and potentially years of work. Preparing for cryptographic failures is prudent, the speaker said, but claims that AI will break all cryptography are unscientific, and “bunker mode” cannot establish when post-AI cryptography is safe.

Near AI’s initial focus is private, verifiable, and secure inference. Polosukhin said its certifications include prompt hashes, response signatures, and hardware signatures that users can independently check through an Explorer-like tool. NEAR is also rolling out encrypted hosting for third-party models, designed so NEAR cannot see model weights and providers cannot see user data. The longer-term plan combines a model and inference marketplace with blockchain-based AI verifiability, with some models potentially accessible through NEAR staking. Polosukhin said open-weight models are gaining usage as quality improves and costs decline; over the preceding three months, token consumption had shifted toward them, beginning to affect revenue projections. He said a customized harness using a DeepSeek-like model matched Claude and GPT on formal-verification work at 100–200 times lower cost, while cautioning that open-weight models lack some imposed guardrails. Security work, he argued, needs models capable of helping protect systems and respond to attacks; Near uses frontier models to attack its own systems.

The speakers said agents may handle sensitive financial, medical, legal, relationship, and business information, making privacy central. One warned that a provider given business documents in plain text could reconstruct a business, and said some businesses felt their work had effectively been stolen or included in training data. Polosukhin argued that user-owned AI should be verifiable: people should be able to inspect open-source code and check an attestation hash for the code that runs. A verifiable trace could identify the model and hardware and expose unauthorized changes or access. Users should also control their models and prompts, he said, citing modified system prompts that could steer views across conversations and arbitrary refusals; he said friends had reportedly been banned after asking geometry questions. He described an inference network as a practical starting point for current business needs, while collaborative model-building, fine-tuning, and improvement remain harder problems involving research, a few startups, and NEAR R&D.

Eli predicted that AI would become computing’s interface, with blockchain coordinating AI, people, and businesses while providing settlement and security. NEAR Intents is being used as a “market of markets” to aggregate liquidity and support agent-driven services, including product discovery and turning selfies into printed professional photos. Alex said truly user-owned AI should be free, with a separate revenue-generating business funding its development; he identified NEAR Intents as that business. Polosukhin said NEAR’s token could become a security, identity, and trust anchor as agents and on-chain transactions grow. Agents may need to hold NEAR to resist Sybil attacks, establish reputation, provide security deposits, and gain access to compute and inference. Agent-level reputation could be subject to slashing, while token requirements could help limit demand for scarce compute.

Agents also need controlled access to credentials and secrets. NEAR’s chain-signature solution derives user-controlled secrets that can be injected into enclaves where agents run tools. Because not all compute can run on-chain, NEAR AI provides inference, while connected trusted execution enclaves can prove to agents and on-chain contracts that execution was correct and unmanipulated. Enclave proofs and user-controlled secrets can be composed into workflows using arbitrary tools, though Anton noted that the discussion had left out important building blocks, including the execution layer. Polosukhin said agents could transact, fulfill jobs through NEAR Intents, run inference on NEAR AI, and rely on the blockchain for security as NEAR scales assets on Intents, blockchain performance, inference capacity, and AI options.

Illia said Neo.com is intended to bring NEAR’s vision together in an accessible product without making users manage blockchain details such as gas fees and addresses. After focusing on infrastructure and expecting applications to build user experiences, the team now sees value in building the integrated experience itself, aiming to rival or improve on traditional banking while retaining blockchain’s proposed benefits. NEAR plans to continue expanding its infrastructure and give developers more ways to use it.