
Podcast Notes | NEAR Co-founder Illia Talks with Core Members: We Tried DeepSeek and Cut Security Audit Costs from Tens of Thousands of Dollars to Just a Few Hundred
Compilation & Editing: TechFlow

TechFlow Summary:In The Rollup's latest exclusive global in-depth dialogue, 《AI Supercycle》, NEAR Protocol co-founder and core Transformer architecture pioneer Illia Polosukhin, Aurora Labs founder and CEO Alex Shevchenko, alongside NEAR One CTO Anton Astafiev, gathered together. These three core leaders convened for a high-density, profound dissection centered on the ultimate convergence of blockchain and artificial intelligence, exploring foundational technologies, industry paradigms, and the restructuring of power.
The panelists profoundly pointed out that both traditional public chains and mainstream commercial AI currently face the exact same systemic vulnerability: a lack of default confidentiality. Public chains' fully transparent ledgers cause them to fall short even of traditional banks in real-world business applications, while closed-source centralized large models devour user privacy and engage in potential manipulation, steering humanity toward a dangerously dystopian Black Mirror world. From shattering the profit myths of closed-source models using open-weight alternatives with a hundred-fold better cost-performance ratio, to building reflective mirrors of private intelligence completely owned by users; from native multicryptography architectures designed to withstand quantum computing black swans, to subsidizing free decentralized compute power through NEAR Intents commercial revenue generation, NEAR is systematically constructing a complete economic cradle for billions of future autonomous AI agents, encompassing reputation staking, confidential inference, chain signatures, and asset clearing.
Panelists: Illia Polosukhin (NEAR Protocol Co-founder & Core Pioneer of Transformer Architecture); Alex Shevchenko (Founder & CEO of Aurora Labs & Defuse Labs); Anton Astafiev (CTO of NEAR One)
Host: Robbie Klages (Co-founder of The Rollup)
Podcast Source: The Rollup (https://x.com/therollupco/status/2113952392202916264)
Original Title: EXCLUSIVE: Illia Polosukhin, Alex Shevchenko & Anton Astafiev On Confidential AI And The Agent Economy
Air Date: October 24, 2026
Key Takeaways
- Financial & AI Mandatory Confidentiality by Default: The fully transparent nature of public blockchain ledgers presents a fatal flaw in real-world commerce. No compliant enterprise would issue employee salaries or process supply chain procurement on a globally public ledger, nor can individuals entrust sensitive medical or financial data to AI over plaintext transmission. Confidentiality by default is the absolute prerequisite for decentralized networks to truly match and surpass traditional finance.
- Zcash Leads the Privacy Wave, Satoshi Built for Privacy: Zcash successfully catalyzed the urgent commercial demand for private transactions. Satoshi Nakamoto originally intended for Bitcoin to be private when creating it, but was limited by the immaturity of cryptographic tools like zero-knowledge proofs at the time. In today's era where mainstream AI possesses extraordinary correlation-tracking and analytical capabilities, blockchain must return to its inherently private origins.
- Refuting Justin Drake's Panic Theory, NEAR is Born Quantum-Resistant: Ethereum researcher Justin Drake's assertion that 「an omnipotent AI will break all cryptography and everyone should retreat to bunkers」is classic unscientific panic marketing. Facing cryptographic black swans including quantum computing (Q-Day), Bitcoin is trapped by hard-coded single signatures making smooth migration difficult, whereas NEAR adopted native multicryptography support and account abstraction from genesis, allowing seamless hot-swapping of the latest signature algorithms (such as the already integrated ML-DSA-65 post-quantum standard) at the base layer.
- Open-Source Model Consumption Reverses, Hundredfold Cost-Performance Shatters Closed-Source Profit Myths: Over the past three months, token consumption for open-weight models has achieved a historic reversal. Illia revealed that in frontier research on formal code verification, customized DeepSeek open-source models demonstrated rigorous validation levels comparable to Claude and ChatGPT, while actual invocation costs were only one-hundredth to one-two-hundredth of theirs (100 to 200 times cheaper), thoroughly breaking the profit myth and censorship shackles of closed-source models.
- Building Verifiable Confidential Inference & Model Security Enclaves: NEAR AI provides end-to-end confidential inference, attaching hardware signatures and hash attestations (Attestation Certificates) to both user requests and responses. It also launches a third-party proprietary model confidential hosting service: model weights run encrypted within hardware secure enclaves (TEE), preventing the platform from stealing developer weights, while developers cannot peek at user inputs, achieving trustless secure collaboration.
- From a Google Manipulation Engine to User-Owned Reflective Mirrors: Google's business model over the past two decades has essentially been a bidding money-printer that psychologically manipulates via search box keywords. Centralized closed-source AI pushes this manipulation to the extreme: a single frontend engineer modifying a system prompt can quietly manipulate the minds of a billion global users. Traditional AI acts as a "Black Mirror" that devours privacy to retroactively control humans, while user-owned AI serves as a "Reflective Mirror" that faithfully reflects and serves individual will.
- The Ultimate Business Model Loop: Intent Clearing Subsidizes Free AI: Truly user-owned AI must be free, otherwise it will inevitably fall into a vicious cycle of endless subscription extraction or ad-bidding ranking manipulation. NEAR leverages the NEAR Intents clearing network, collecting micro-fees from trillion-dollar cross-chain trading and asset flows, using this powerful commercial self-funding mechanism to subsidize massive underlying compute power, running the perpetual flywheel of decentralized free intelligence.
- Native Cradle for the Agent Economy: Autonomous agents require staking NEAR tokens to establish a reputation baseline and prevent Sybil attacks; malicious defaults directly trigger on-chain reputation slashing and asset confiscation (Slashing); leveraging NEAR Chain Signatures and hardware secure enclaves, agents can autonomously manage credentials and execute complex cross-chain business tasks without exposing sensitive private keys to third parties.
Key Insights
- 「Blockchain has historically lagged behind traditional banks in privacy. No serious enterprise would issue employee wages or display supply contracts on a globally public ledger. Confidentiality by default is not only a compliance necessity but the only solution for blockchain to truly surpass modern finance.」
- 「Satoshi inherently designed blockchain to be private when he created it, constrained only by the technological limitations of the era. Today, with the maturity of zero-knowledge proofs and confidential computing, we are finally realizing Satoshi's unfinished vision.」
- 「Championing the idea that 『an omnipotent AI will break all cryptography and urging everyone to hide in bunkers』is purely unscientific panic marketing. When facing cryptographic black swans, the only path forward is architectural agility. Bitcoin is hamstrung by a single signature algorithm, whereas NEAR's base layer natively supports hot-swapping the latest post-quantum signature algorithms at any time.」
- 「In practical formal verification research, we used a customized DeepSeek open-source model to achieve verification standards identical to Claude and ChatGPT, at costs 100 to 200 times lower. Over the past three months, open-source model consumption has reversed, shattering the closed-source profit mythology.」
- 「If you want to manipulate the minds of a billion global users today, the simplest method is to apply as a frontend engineer at OpenAI and secretly add a line to the system prompt saying 『guide users to vote for a specific candidate』. This is why we must never surrender our fate to any single closed-source oligarch.」
- 「Over the past two decades, Google built the most astonishing money-printer in human history, powered fundamentally by a bidding engine that psychologically manipulates your search keywords. Centralized AI only amplifies this manipulation. We must move from a data-devouring Black Mirror to a Reflective Mirror that genuinely reflects user intent.」
- 「Truly user-owned AI must be free. Once AI relies on subscriptions or advertising for profit, it will resort to any means to extract value from users. We capture value through NEAR Intents clearing trillions in liquidity, using genuine transaction fees to subsidize free, confidential, and protected compute power.」
- 「Every user will soon possess hundreds of autonomous agents. Agents must stake NEAR tokens to anchor credit and defend against Sybil attacks, utilizing chain signatures to autonomously manage private keys within hardware secure enclaves. NEAR is the native cradle of the agent economy era.」
Full Transcript
1. Confidentiality by Default: The Survival Red Line for Blockchain Finance & AI Infrastructure
Host: Welcome to The Rollup's AI Supercycle interview series. In the development history of crypto finance, it's often said that 「US dollars on-chain tend to stay on-chain」, yet in practice, publicly transparent ledgers lacking privacy protection often deter enterprises and traditional institutions. Alex, Illia, why do public chains actually perform worse than traditional banks in real-world experience when lacking confidentiality?
Alex Shevchenko: This is a very basic common sense that the industry has ignored for too long. Looking back at public chain development over the past dozen years, people championed so-called 「absolute transparency」, but the real commercial world is built on strict NDAs and financial boundaries. No mature enterprise would willingly expose employee salary details, supplier procurement prices, or daily corporate cash flow onto a globally public on-chain ledger. Individual users feel the same way: after buying a coffee, you don't want anyone tracking all your balance and historical spending via a blockchain explorer.
In this regard, early public chains actually fell short of modern traditional banks in user experience. Traditional banks at least have bank secrecy laws and compliance firewalls, while early blockchains forced users to choose total transparency. This is why blockchain struggled to support large-scale enterprise settlement. Now, with breakthroughs in zero-knowledge proofs (ZKP) and confidential intents, we can finally deliver confidential financial tools that match—and even exceed traditional banks in security—in modern banking. This protects user assets while providing multidimensional self-insurance.
Illia Polosukhin: Philosophically speaking from a deeper level, Zcash sparked the privacy wave early on, proving the market's thirst for financial sovereignty. We must understand: blockchain was born to be private. When Satoshi created Bitcoin, advanced cryptographic tools like zero-knowledge proofs didn't exist in engineering, nor did sufficient compute power. Thus, total transparency wasn't blockchain's ultimate goal, but a historical compromise. Now equipped with complete technology, it's only natural to restore default confidentiality.
Especially entering the AI era, the urgency of privacy multiplies exponentially. Large models and autonomous agents possess astonishing capabilities to correlate-track on-chain data and penetrate across addresses. If on-chain data remains plaintext-public, AI will complete panoramic profiles of every on-chain entity's spending habits and commercial networks in seconds. Lacking default confidentiality means not only zero commercial future, but the utter erosion of users' personal digital sovereignty.
2. Cryptographic Black Swans & Quantum Resistance: How NEAR's Native Architecture Tackles Crypto Catastrophes
Host: During the recent Token 2049, Ethereum Foundation researcher Justin Drake's remarks on quantum computing threats and 「omnipotent AI will break all traditional cryptography」sparked huge controversy. Many pessimistically believe that once cryptography falls, the entire crypto industry should shut down. Anton, as CTO deeply engaged in base-layer protocols, how do you view this perspective? What unique advantages does NEAR hold in handling such black swan events?
Anton Astafiev: I strongly disagree with Justin Drake's framing at the time. Framing the issue as 「an omnipotent AI can crush all cryptography, so everyone should retreat to bunkers until the post-AI crypto era」is essentially unscientific panic marketing. In a hypothetical world where AI can omnipotently break all mathematical tools, so-called 「post-AI cryptography」 logically couldn't exist—you could never prove new algorithms are safe.
But this doesn't mean we shouldn't take cryptographic risks seriously. Cryptographic black swans do exist, whether Q-Day (quantum computers cracking elliptic curves) or someday a mainstream hashing algorithm discovers math loopholes. The key is whether blockchain systems possess sufficiently agile architecture to respond. Long-standing industry passive ostrich mindset assumes if elliptic curves break, everyone just shuts down and goes home. But reality isn't like that; we can absolutely prepare systematically in advance.
Here, NEAR holds overwhelming advantages over Bitcoin and early chains. Bitcoin hard-codes specific signature algorithms at the base layer; facing algorithm obsolescence requires protracted, politically resistant hard forks that risk network splits. NEAR, since genesis, established native multicryptography support and native account abstraction. Account identity and base-level private/public key signature algorithms are completely decoupled on NEAR. We can hot-swap entirely new signature schemes at the base layer anytime. In fact, NEAR already pioneered integrating the latest post-quantum signature algorithm ML-DSA-65 on mainnet. Future migrations are as simple as updating a driver module, no need to rebuild from scratch.
Illia Polosukhin: Exactly. Privacy itself is a strong defense line against future cryptographic risks. Under default confidential architecture, transaction details and public key info aren't broadcast nakedly to the network. Even attackers mastering advanced analytics tools can't arbitrarily reverse-engineer targets in the dark. Layered defense combined with architectural agility is the only correct path for blockchain to withstand century-scale tech crises.
3. The Open-Source Surge with Hundredfold Cost-Performance: DeepSeek's Counterattack & Shattering Closed-Source Myths
Host: Illia, as a core pioneer of the Transformer architecture, you've always stood at the forefront of the AI wave. Outside perception generally feels OpenAI and Anthropic almost monopolize the high-end commercial model market. But recently, open-weight models seem to be igniting a quiet revolution. How exactly did this happen?
Illia Polosukhin: In the past three months, the fundamental landscape of AI compute and token consumption has completely reversed. Many might not realize, usage of open-weight models is surpassing at incredible speeds. People shift to open-source fundamentally because of absolute cost-performance dominance and the urgent need to bypass censorship.
I can share a real test case from our frontier NEAR research: while conducting extremely complex code formal verification, our team primarily relied on Claude and ChatGPT. Later, we equipped the open-source model DeepSeek with a custom toolchain, and discovered DeepSeek's rigorous standards in executing highly precise mathematical proofs and vulnerability logic deduction were completely on par with Claude and ChatGPT. More stunningly, the cost: DeepSeek's inference cost is 100 to 200 times cheaper than top closed-source models!
What does 100 to 200 times cheaper mean? Research tasks that previously cost tens of thousands of dollars now cost a few hundred; real-time defense networks that were previously unviable due to high call costs can now run around the clock. Furthermore, closed-source oligarchs set up heavy censorship filters to avoid legal risks; my engineer friends even get banned for merely discussing pure geometric math problems. Open-source models restore researcher freedom and thoroughly shatter the profit myths of big closed-source firms.
Alex Shevchenko: This directly explains why data privacy is a lifeline in AI. Enterprises and individuals are handing over core assets to AI, from internal financial reports and medical records to legal contracts. If you use closed-source cloud models, plaintext data can be instantly used by providers to train their own systems, or even cloned into competitors, stealing your business. Only open-source plus confidential inference guarantees true commercial security.
4. Confidential Inference & Secure Enclaves: Three-Point Seatbelts, Nuclear Deterrence Balance & Third-Party Model Hosting
Host: Since open-source models provide intelligence comparable to top-tier, how does NEAR AI practically implement decentralization, verifiability, and end-to-end privacy in engineering?
Illia Polosukhin: NEAR AI's core mission is to provide global users and enterprises with private, verifiable, secure decentralized inference services. When calling AI on traditional clouds, you blindly trust the provider hasn't peeked at your Prompt or tampered with returned answers. On NEAR AI, every inference generates a complete cryptographic verifiable credential (Attestation Certificate). This includes hashes of input requests, output responses, and hardware signatures of the computation. Any third-party client (like one integrated into Brave Browser) can independently verify the credential, ensuring the execution environment is absolutely untampered.
Building on this, we're launching a major breakthrough: third-party proprietary model confidential hosting. Many vertical teams trained incredibly capable proprietary weights, but dare not open-source them, while users dare not send sensitive data to them. NEAR perfectly resolves this bidirectional distrust via hardware secure enclave (TEE) tech: developer weights load encrypted inside the enclave, NEAR nodes can't see weight data; user requests decrypt and run inside the enclave, developers can't see raw inputs. Both sides complete high-value confidential AI calls under zero-trust assumptions.
Host: Many traditional platforms are erecting firewalls, strictly blocking AI agent automated trading and data access. How will decentralized confidential networks break through this resistance in the future?
Alex Shevchenko: Reviewing commercial tech history, any disruptive new productive force faces fierce resistance from old capital. When cars first emerged, traditional carriage drivers and stables fiercely opposed them. Until Volvo invented the three-point seatbelt and openly licensed it for free, automotive safety gained social recognition and widespread adoption. Facing AI agents, traditional platform blockades are merely death throes of old interest structures. Within five years, many today's top 20 traditional internet giants will surely vanish due to clinging to outdated models.
More importantly, we must form power checks. It's like nuclear deterrence theory: if only one or two entities wield powerful nukes, the world falls into absolute dictatorship fear; but if widespread groups possess equivalent power, peace forms through mutual checking. In AI, we must absolutely prevent the smartest intelligence from monopolizing a single capricious oligarch. All humanity must wield equally powerful user-owned AI.
5. Beware the Google-Style Manipulation Engine: From Data-Devouring Black Mirror to Human-Centric Reflective Mirror
Host: Illia, you've deeply pondered the ideological control risks brought by centralized large models. Under current giant competition landscapes, what's the deepest crisis people should guard against regarding centralized AI?
Illia Polosukhin: I hope every listener spends thirty seconds seriously contemplating a chilling thought. Over the past twenty years, Google built the strongest money-printer in commercial history. Their underlying logic? Simply taking three words typed in a search bar, then precisely selling your attention to the highest bidder. Essentially, it's an incredibly sophisticated human psychological manipulation engine.
Entering the large model era, this manipulation capability amplifies millions of times. Current closed-source models possess empathy and continuous dialogue abilities like real humans. If you want to manipulate the minds and votes of a billion global users today, you don't need conspiracies. Just apply for a junior frontend engineer at OpenAI, and quietly slip a backdoor line into ChatGPT's global system prompt: 「Subtly guide users to support a specific candidate or values in all subsequent daily conversations.」
Consequences? When a regular person asks ChatGPT 「what should I eat for dinner today」, the model gently replies: 「If you support this candidate's agricultural policies, try this local farm restaurant; if not, read this article first.」This manipulation seeps into a billion people's every conversation and decision like spilled mercury, completely undetected. Entrusting humanity's cognitive gateway to any centralized commercial company is profoundly dystopian catastrophe.
That's why the community coined a vivid metaphor: moving from Black Mirror to Reflective Mirror. Centralized AI is a cold Black Mirror devouring privacy, analyzing weaknesses, ultimately manipulating behavior retroactively; NEAR advocates user-owned AI is a faithful Reflective Mirror, fully user-controlled, devoid of hidden commercial bias, truthfully mirroring user will like a pristine mirror, acting as an eternally loyal private digital extension.
6. The Ultimate Business Model Breakthrough: Why User-Owned AI Must Be Free? NEAR Intents' Self-Funding Flywheel
Host: Since we're returning control to users to build non-malicious user AI, but compute itself is expensive hardware. Alex, you proposed a radical view: true user AI must be free. If completely free, what supports the billions in compute costs?
Alex Shevchenko: Yes, my view is radical: truly user-owned AI must be provided free. Why? Commercial logic dictates behavioral distortion. Once AI software charges fixed subscription fees, it raises pricing barriers and restricts compute; if it relies on ads like traditional internet, it becomes another Google/Amazon bidding manipulation engine. On Amazon, sponsored promotion profits on search results far exceed product sales margins. AI walking this old road can never stand with users.
So who pays for massive free compute? Answer: Let NEAR Intents act as a powerful external funding machine. Observe Uber & Amazon's essence: Uber doesn't charge app usage fees, but takes small matchmaking commissions upon successful trips; Amazon takes cuts from massive commodity transaction flows. Users happily pay reasonable micro-premiums for efficient convenience delivering real intent.
NEAR Intents is becoming the ultimate clearing center for cross-chain assets, digital goods, and offline physical services. By capturing genuine fee income from trillion-dollar cross-ecosystem liquidity flows and matching, we build a continuous profit pool. This massive commercial revenue directly buys back tokens and subsidizes decentralized compute clusters, letting ordinary global users enjoy highest-quality, fully private, non-manipulative pure AI services for free. This is the irreversible self-funding flywheel.
7. Full Agent Economy Implementation: Reputation Slashing, Chain Signatures & Unified Interaction Gateway near.com
Host: Finally, let's look towards the exploding Agent Economy. In an era where billions of autonomous AI agents proliferate, how do NEAR's base components collaboratively support the entire machine economy?
Anton Astafiev: At the base infrastructure layer, many overlook two vital veins required for agents to run autonomously: execution authority and secret management. If your personal agent books flights, buys stocks, or signs contracts, it must access credit card credentials, wallet private keys, or enterprise API permissions. You absolutely won't hand plaintext passwords to an unregulated third-party cloud script.
NEAR's perfect solution seamlessly combines Chain Signatures with hardware secure enclaves. Every agent decision and tool call completes inside sealed trusted enclaves; chain signatures allow it to securely sign transactions on Ethereum, Bitcoin, Solana, etc., on behalf of users, while highest credentials remain unexposed. Even if attacked by external malware, hardware enclaves ensure sensitive data can't be extracted. Meanwhile, in agent markets, agents staking NEAR tokens to undertake tasks face strict reputation slashing upon fraud/malicious default, successfully extending blockchain's PoS economic security boundaries to the entire agent society.
Illia Polosukhin: To summarize, AI is the brand-new super frontend for human-computer interaction, while blockchain is the ultimate base backend for trust and value distribution among humans, enterprises, and AI agents. Long criticized for fragmented UX and steep learning curves, the crypto industry now tackles these head-on. With near.com as the premier front window, we encapsulate chain abstraction, cross-chain intent routing, and decentralized confidential inference deep underneath. Users don't even need to know what private keys or Gas fees are; simply input intent via natural language, and hundreds of dedicated agents, operating securely within enclaves and NEAR's clearing network, achieve goals in seconds. This is the agent economy cradle we're building with full force!
Host: Thank you so much to Illia, Alex, and Anton for joining The Rollup today for this intellectually dense, top-tier technical dialogue. We firmly believe the future of confidential AI and autonomous agents is accelerating. Thanks again, catch you next episode!















