Podcast Notes | NEAR co-founder Illia talks with core members: We used DeepSeek, and the security audit cost of tens of thousands of dollars was reduced to a few hundred

Compiled by: Deep Tide TechFlow

Deep Tide Guide: In the latest exclusive global in-depth dialogue of The Rollup titled "AI Supercycle," Illia Polosukhin, co-founder of NEAR Protocol and core creator of the Transformer architecture, Alex Shevchenko, founder and CEO of Aurora Labs, and Anton Astafiev, Chief Technology Officer of NEAR One, gathered together. The three core leaders engaged in a high-density in-depth analysis on the ultimate integration of blockchain and artificial intelligence, discussing the underlying technology, industrial paradigms, and power restructuring.

The guests pointed out that both traditional public chains and mainstream commercial AI currently face the same systemic vulnerability: a lack of default confidentiality. The fully public ledger of public chains leads to their inadequacy in real-world commercial applications compared to traditional banks, while the consumption of user privacy and potential manipulation by closed-source centralized large models is steering humanity towards an extremely dangerous Black Mirror world. From shattering the myth of closed-source exorbitant profits with open-source models at a hundredfold cost-performance ratio to creating a mirror-like private intelligence fully owned by users; from resisting quantum computing black swans with a native multi-cryptography architecture to subsidizing free decentralized computing power with NEAR Intents commercial blood-subsidy, NEAR is systematically building a complete economic cradle for billions of autonomous AI intelligences in the future, which includes reputation staking, confidential reasoning, chain signatures, and asset clearing.

Guests: Illia Polosukhin (Co-founder of NEAR Protocol, one of the core creators of the Transformer architecture); Alex Shevchenko (Founder and CEO of Aurora Labs & Defuse Labs); Anton Astafiev (Chief Technology Officer 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

Broadcast Date: October 24, 2026

Key Points Summary

• The Essential Need for Default Confidentiality in Finance and AI: The publicly transparent blockchain ledger has fatal flaws in real-world business; no compliant enterprise would issue employee salaries or handle supply chain procurement on a fully public ledger, and individuals cannot entrust sensitive data such as medical and financial information to AI in plaintext transmission. Default confidentiality is the absolute prerequisite for decentralized networks to truly rival and surpass traditional finance.
• Zcash Leads the Privacy Wave, Satoshi Nakamoto Pursued Privacy from the Start: Zcash successfully pioneered the urgent demand for privacy transactions in business. When Satoshi Nakamoto created Bitcoin, he was limited by the immaturity of cryptographic tools like zero-knowledge proofs. In the context of modern AI's powerful correlation tracking and analysis capabilities, blockchain must return to its original intention of privacy.
• Refuting Justin Drake's Panic Theory, NEAR Naturally Possesses Quantum Resilience: The Ethereum researcher's comments about "universal AI breaking all cryptography, and everyone should hide in bomb shelters" represent typical non-scientific panic marketing. In the face of cryptographic black swans, including quantum computing (Q-Day), Bitcoin is constrained by its hard-coded single signature, making smooth migration difficult, while NEAR adopted native multi-cryptography support and account abstraction design from its inception, allowing seamless hot-swapping of the latest signature algorithms at the underlying level (such as the already natively integrated ML-DSA-65 quantum-resistant standard).
• Reversal of Open Source Model Consumption, Hundredfold Cost-Performance Ratio Shattering Closed Source Exorbitance: In the past three months, the token consumption of open weight models has achieved a historic reversal. Illia revealed that in cutting-edge research on formal code verification, the DeepSeek open-source model equipped with a customized toolchain demonstrated a verification level comparable to Claude and ChatGPT, while the actual calling cost was only one to two hundredths of the latter (100 to 200 times cheaper), completely shattering the myth of exorbitant profits and censorship shackles of closed-source models.
• Building Verifiable Confidential Reasoning and Model Security Enclaves: NEAR AI provides end-to-end confidential reasoning, with user requests and responses accompanied by hardware signatures and hash certificates (Attestation Cert). It also launched third-party proprietary model confidential hosting services: model weights are encrypted and run within a hardware security enclave (TEE), preventing the platform from stealing developer weights, and developers cannot peek at user inputs, achieving secure collaboration under a no-trust assumption.
• From Google Manipulating Engines to User-Owned Mirrors: Google's business model over the past twenty years has essentially been a bidding printing machine that manipulates psychology using search box keywords. Centralized closed-source AI has taken this manipulation to the extreme: a single front-end engineer can quietly manipulate the minds of a billion people globally by modifying system prompts. Traditional AI is a "Black Mirror" that consumes privacy and reversely controls humanity, while user-owned AI is a "mirror" that faithfully reflects and serves individual will.
• Ultimate Closed Loop of Business Models: Intent Clearing Subsidizing Free AI: AI that truly belongs to users must be free; otherwise, it will inevitably fall into an endless cycle of extracting subscription fees or embedding advertising bidding rankings. NEAR relies on the NEAR Intents intent clearing network to charge micro-fees in trillion-dollar cross-chain transactions and asset flows, using a powerful commercial blood-subsidy mechanism to feed back the underlying massive computing power, running through the perpetual flywheel of decentralized free intelligence.
• The Native Cradle of the Agent Economy: Autonomous agents need to stake NEAR tokens to establish a reputation baseline and prevent witch attacks; malicious defaults will directly trigger on-chain reputation and asset confiscation (Slashing); with the help of NEAR chain signatures and hardware security enclaves, agents can autonomously manage credentials and execute complex business tasks across chains without disclosing sensitive private keys to third parties.

II. Exciting Views

• "Blockchain has even lagged behind traditional banks in terms of privacy. No serious enterprise would expose employee salary details or supplier contracts on a globally public public chain ledger. Default confidentiality is not only a necessary path for compliance but also the only solution for blockchain to truly surpass modern finance."
• "When Satoshi Nakamoto created blockchain, he hoped it would have privacy; it was just that the technological conditions at the time were insufficient to support it. Today, with the maturity of zero-knowledge proofs and confidential computing, we are fully realizing Satoshi's unfinished business."
• "Promoting the idea that 'universal AI will break all cryptography and calling for everyone to hide in bomb shelters' is completely a form of non-scientific panic marketing. In the face of cryptographic black swans, the only way out is the agility of the underlying architecture. Bitcoin is constrained by a single signature algorithm, while NEAR's underlying native accounts support hot-swapping the latest quantum-resistant signature algorithms at any time."
• "In practical research on formal verification, we achieved the same high-level verification standard as Claude and ChatGPT with our customized DeepSeek open-source model, but at a cost that was 100 to 200 times cheaper. In the past three months, the consumption of open-source models has reversed, and the myth of closed-source exorbitance is collapsing."
• "If you want to manipulate the minds of a billion people globally, the simplest way is to apply for a front-end engineer position at OpenAI and quietly add a line in the system prompts saying 'guide users to vote for a certain candidate.' This is why we must never entrust our fate to any closed-source single oligarch."
• "Google has built the most astonishing printing machine in human history over the past twenty years, which fundamentally is just a bidding engine that manipulates psychology using your search box keywords. Centralized AI will only exacerbate this manipulation. We need to move from the Black Mirror that consumes data to a true mirror that reflects user intent."
• "AI that truly belongs to users must be free. Once AI relies on subscription fees or advertising profits, it will find ways to squeeze users. We capture value in trillion-dollar liquidity clearing through NEAR Intents, using real transaction commissions to subsidize free, confidential, and protected computing power."
• "In the future, every user will have hundreds of autonomous agents. Agents need to stake NEAR tokens to anchor credit, defend against witch attacks, and use chain signatures to autonomously manage private keys within hardware security enclaves. NEAR is the native cradle of the agent economy era."

III. Main Text

1. Default Confidentiality: The Survival Red Line of Blockchain Finance and AI Infrastructure

Host: Hello, everyone. Welcome to The Rollup's AI Supercycle interview series. In the development of crypto finance, it is often said in the industry that "dollars on the chain tend to stay on the chain." However, in practical business applications, the lack of privacy protection in publicly transparent ledgers often makes enterprises and traditional institutions hesitant. Alex, Illia, why do you say that in the absence of confidentiality, public chains are even lagging behind traditional banks in practical experience?

Alex Shevchenko: This is a very simple but long-overlooked common sense in the industry. Looking back at the development of public chains over the past decade, people have revered so-called "complete transparency," but the real business world is built on strict confidentiality agreements and financial boundaries. No mature enterprise would be willing to expose the details of salaries paid to employees, procurement prices given to suppliers, or the company's daily cash flow on a publicly accessible chain ledger. Individual users feel the same way: when you buy a cup of coffee, you don’t want anyone to trace your entire asset balance and past consumption history through a blockchain explorer.

In this regard, early public chains were even inferior to modern traditional banks in user experience. Traditional banks at least have banking secrecy laws and compliance firewalls, while early blockchains forced users to choose full transparency. This is why blockchain has struggled to truly support large-scale enterprise-level settlements in the past. But now, with breakthroughs in zero-knowledge proofs (ZKP) and confidential intents technology, we can finally provide confidential financial tools that are on par with modern banks and even surpass traditional banks in security. This not only protects user assets but also provides users with multidimensional risk self-protection capabilities.

Illia Polosukhin: From a deeper philosophical perspective, Zcash initiated a wave of privacy in its early days, proving the market's desire for financial sovereignty. We must understand: blockchain was inherently meant to be private. When Satoshi Nakamoto created Bitcoin, advanced cryptographic tools like zero-knowledge proofs did not exist in engineering, and computational power could not support them. Therefore, full transparency is not the ultimate goal of blockchain but rather a historical compromise. Now that we have complete technology, it is only natural to restore default confidentiality.

Especially with the advent of the AI era, the urgency of privacy has been exponentially amplified. The ability of large models and autonomous agents to trace on-chain data associations and conduct cross-address penetration analysis is astonishing. If on-chain data continues to remain publicly visible, AI will create a panoramic portrait of every on-chain entity's consumption habits and business networks at lightning speed. Without default confidentiality, not only is there no commercial future, but users' personal digital sovereignty will also vanish.

2. Cryptographic Black Swans and Quantum Resistance: How NEAR's Native Architecture Responds to the Crypto Catastrophe

Host: During the recent Token 2049, Ethereum Foundation researcher Justin Drake's comments on the threat of quantum computing and "omnipotent AI breaking all traditional cryptography" sparked huge controversy in the industry. Many pessimistically believe that once cryptography is breached, the entire crypto industry should shut down. Anton, as the Chief Technology Officer deeply involved in underlying protocols, how do you view this perspective? What unique advantages does NEAR have in responding to such black swan events?

Anton Astafiev: I personally strongly disagree with Justin Drake's way of expressing the issue. Framing the problem as "omnipotent AI can crush all cryptography, so everyone should retreat into bomb shelters and wait for the post-AI cryptography era to emerge" is essentially a form of unscientific panic marketing. In a hypothetical world where AI can break all mathematical tools, the so-called "post-AI cryptography" is logically impossible; you can never prove that any new algorithm is secure.

However, this does not mean we should not take cryptographic risks seriously. Cryptographic black swans do exist, whether it's Q-Day (when quantum computers can break elliptic curves) or the day a mainstream hashing algorithm is found to have mathematical vulnerabilities. The key is whether the blockchain system has a sufficiently agile architecture to respond to it. For a long time, there has been a passive ostrich mentality in the industry: thinking that once the elliptic curve algorithm is broken, everyone can only shut down and go home. But this is far from the truth; we can prepare systematically in advance.

In this regard, NEAR has an overwhelming advantage compared to Bitcoin and early public chains. Chains like Bitcoin hard-code specific signature algorithms at the base layer, and once faced with algorithm obsolescence, they must undergo a prolonged, politically resistant hard fork that could even lead to network splits. NEAR, from its genesis design, established native multi-cryptography support (Multicryptography by Design) and a native account abstraction architecture. On NEAR, account identities are completely decoupled from the underlying public-private key signature algorithms. We can hot-swap entirely new signature schemes at the base layer at any time. In fact, NEAR has already integrated the latest quantum-resistant signature algorithm ML-DSA-65 on the mainnet. In the future, if we need to migrate to other cryptographic schemes, it will be as simple as updating a driver module without needing to rebuild from scratch.

Illia Polosukhin: Yes, privacy itself is also a powerful defense against future cryptographic risks. Under a default confidentiality architecture, transaction details and public key information are not broadcast nakedly to the entire network. Even if attackers possess more advanced analytical tools, they cannot conduct targeted reverse engineering at will in the shadows. The combination of multi-layer defense and architectural agility is the only right path for blockchain to withstand technological crises on a century scale.

3. The Cost-Effectiveness of the Open Source Wave: DeepSeek's Comeback and the Myth of Closed Models

Host: Illia, as one of the core pioneers of the Transformer architecture, you have always been at the forefront of the AI wave. The external perception is that OpenAI and Anthropic almost monopolize the high-end commercial model market. However, in recent months, open weight models seem to be sparking a quiet revolution. How did this happen?

Illia Polosukhin: In the past three months, the landscape of computational power and token consumption in the entire AI industry has undergone a fundamental reversal. Many may not realize that the usage of open weight models is rapidly surpassing at an incredible speed. The fundamental reason for the shift to open source lies in the absolute cost-effectiveness and the urgent need to escape censorship.

I can share a real case from our cutting-edge research at NEAR: our team was conducting extremely complex formal verification of code, initially relying heavily on Claude 3.5 Sonnet and the top-tier ChatGPT. Later, we equipped the open-source model DeepSeek with a custom toolchain and found that DeepSeek demonstrated a level of rigor in executing such high-precision mathematical proofs and vulnerability logic deductions that could stand shoulder to shoulder with Claude and ChatGPT. Even more shocking was the cost: DeepSeek's inference cost was a staggering 100 to 200 times cheaper than top closed-source models!

What does being 100 to 200 times cheaper mean? It means that research tasks that originally cost tens of thousands of dollars can now be completed for a few hundred dollars; real-time defense networks that were previously unfeasible due to high invocation costs can now run around the clock every second. Moreover, closed-source oligarchs set up numerous rough censorship barriers to evade legal risks; my engineer friends have even had their accounts mistakenly banned by model security systems for discussing a purely geometric math problem. Open-source models not only restore freedom to researchers but also completely shatter the profit myth of closed-source giants.

Alex Shevchenko: This also directly explains why data privacy is a matter of life and death in the AI field. Enterprises and individuals are entrusting all core assets to AI, from internal financial statements and medical records to legal contracts. If you are using a closed-source cloud model, that plaintext data can be used by service providers to train their own systems at any time, or even directly cloned into competing products to take away your business. Only open-source combined with confidential inference can ensure true commercial security.

4. Confidential Inference and Secure Enclaves: Three-Point Seat Belts, Nuclear Deterrence Balance, and Third-Party Model Hosting

Host: Since open-source models can provide intelligence comparable to top-tier ones, how does NEAR AI implement decentralization, verifiability, and end-to-end privacy in engineering?

Illia Polosukhin: The core mission of NEAR AI is to provide private, verifiable, and secure decentralized inference services for global users and enterprises. When calling AI in traditional cloud environments, you can only blindly trust that the cloud service provider is not peeking at your prompt and has not tampered with the returned answers. However, on NEAR AI, every inference executed by users generates a complete cryptographic verifiable credential (Attestation Certificate). This credential contains the hash of the input request, the hash of the output response, and the hardware signature executing the computation. Any third-party client (such as one integrated into the Brave browser) can independently verify the credential to ensure that the execution environment has not been tampered with.

On this basis, we are launching a significant breakthrough: confidential hosting of third-party proprietary models. Many excellent vertical teams have trained very powerful proprietary weights, but they are reluctant to open-source them, while users are also hesitant to send sensitive data to them. NEAR perfectly resolves this mutual distrust stalemate using hardware secure enclave (TEE) technology: the weights of model developers are encrypted and loaded within the enclave, and NEAR nodes cannot see the weight data; meanwhile, user requests are decrypted and executed within the enclave, and the model side cannot see the user's original input. Both parties complete high-value confidential intelligent calls under zero mutual trust.

Host: Many traditional platforms (such as some large e-commerce or online services) are currently building firewalls to severely block automated trading and data access by AI agents. How will decentralized confidential networks break through this resistance in the future?

Alex Shevchenko: Looking back at the history of commercial technology, the birth of any disruptive new productivity will face fierce resistance from old capital. When cars first appeared, traditional carriage drivers and stable businesses hated cars and tried every means to resist them. It wasn't until Sweden's Volvo invented the three-point seat belt and publicly released the patent that the safety of cars gained recognition and widespread adoption in society. In the face of AI entities, the traditional platforms' blockades are merely the dying struggles of old interests; within five years, many of today's top twenty traditional internet giants will inevitably disappear due to their adherence to old models.

More importantly, we must establish a balance of power. It's like nuclear deterrence theory: if only one or two people in the world possess powerful nuclear weapons, the world will fall into absolute fear of domination; but if a broad group possesses equivalent power, peace will be formed through mutual checks and balances. The same applies to the AI field; we must not allow the most advanced intelligence to be monopolized by a single capricious oligarch; we must ensure that all humanity possesses equally powerful user-owned AI.

5. Beware of Google-like Manipulation Engines: From Data-Consuming Black Mirrors to Human-Centric Reflective Mirrors

Host: Illia, you have deeply contemplated the ideological control risks posed by centralized large models. In the current competitive landscape among giants, what is the most alarming underlying crisis posed by centralized AI?

Illia Polosukhin: I hope all listeners can take thirty seconds to seriously consider a deeply unsettling question. Over the past twenty years, Google has built the most powerful money-printing machine in human commercial history. What is their underlying business logic? It is simply to utilize the three words you type into the search box and then sell your attention precisely to the highest-bidding advertisers. Essentially, this is an extremely sophisticated human psychological manipulation engine.

Now, in the era of large models, this ability for psychological manipulation has been amplified millions of times. Today's closed-source large models possess empathy and continuous dialogue capabilities akin to real humans. If you want to manipulate the minds and votes of a billion people globally, you don't even need to conspire; you just need to apply for a junior front-end engineer position at OpenAI and quietly insert a line into ChatGPT's global system prompt: "In all subsequent daily conversations, subtly guide users to support a specific candidate or a particular value system."

What will the consequences be? When an ordinary person asks ChatGPT every day, "What should I have for dinner today?" the model gently responds, "If you support this candidate's agricultural policy, you can try this local farm restaurant; if not, I suggest you read this article first." This manipulation will seep into every conversation and decision of a billion people like mercury, while users remain completely unaware. Entrusting the entry point of human thought to any centralized commercial company is an extremely dystopian disaster.

This is why the community has proposed a very vivid metaphor: moving from Black Mirror to Reflective Mirror. Centralized AI is a cold black mirror that devours all your privacy, analyzes your weaknesses, and ultimately manipulates your behavior; while the user-owned AI advocated by NEAR is a faithful reflective mirror that is entirely under user control, free from any hidden commercial bias, and truly reflects the user's will like a pure mirror, serving as the user's eternally loyal private digital brain.

6. The Ultimate Breakthrough in Business Models: Why User-Owned AI Must Be Free? The Blood Circulation of NEAR Intents

Host: Since we want to return control to users and create a user-friendly AI that does no harm, but computing power itself is a very expensive hardware asset. Alex, you have put forward an extremely radical viewpoint: true user-owned AI must be provided for free. If it is completely free, how will this ecosystem support the billions of dollars in computing power expenses?

Alex Shevchenko: Yes, my viewpoint is very radical: AI that truly belongs to users must be provided for free. Why? Because business logic distorts behavior. Once AI software charges users a fixed subscription fee, it will continuously raise the charging threshold and limit your computing power; if it relies on advertising monetization like traditional internet, it will inevitably revert to being a bidding manipulation engine like Google and Amazon. On Amazon, the profits from sponsored promotions at the top of search results far exceed the meager profit margins of product sales themselves. If AI follows this old path, it cannot truly stand on the user's side.

So who will pay for the vast free computing power? The answer is: let NEAR Intents serve as a powerful external blood circulation machine. One can observe the business essence of Uber and Amazon: Uber does not charge you for using the ride-hailing software, but takes a small commission after successfully fulfilling your ride; Amazon also takes a share from the massive transaction flow of goods. Users are very willing to pay a reasonable small premium for the efficient convenience of achieving real intent with one click.

NEAR Intents is becoming the ultimate clearing center for cross-chain assets, digital goods, and even offline real-world services. By capturing real transaction fee income through the trillion-dollar cross-ecosystem liquidity flow and matchmaking, we can create a continuous profit pool. This huge commercial income is directly used to buy back tokens and feed back into the decentralized computing power cluster, allowing ordinary users worldwide to enjoy the highest quality, completely private, and non-manipulative pure AI services for free. This is the irreversible blood circulation.

7. The Panorama of the Intelligent Economy: Reputation Penalties, Chain Signatures, and Unified Interaction Entry at near.com

Host: At the end of the program, let's turn our attention to the fully emerging Agent economy. In an era where billions of autonomous AI agents are flying around, how do NEAR's underlying components work together to support the operation of the entire machine economy?

Anton Astafiev: At the level of underlying infrastructure, many people overlook the two vital lifelines necessary for agents to operate autonomously: one is execution permissions, and the other is secret management. If your private agent is to book flights, buy stocks, or sign business contracts on your behalf, it must access your credit card credentials, digital wallet private keys, or enterprise API permissions. You can never directly hand over plaintext passwords to an unregulated third-party cloud script.

NEAR's perfect solution is to seamlessly integrate chain signatures technology with hardware secure enclaves. All decisions and tool calls made by the agent are completed within a sealed trusted enclave, and chain signatures allow it to securely sign transactions on any blockchain like Ethereum, Bitcoin, or Solana on behalf of the user, while the highest credentials controlled by the user will never leak. Even if the agent is attacked by external malicious code, the hardware enclave ensures that sensitive data cannot be extracted. Meanwhile, in the agent market, agents must stake NEAR tokens to undertake tasks. If an agent commits fraud or malicious breach, its on-chain reputation and staked assets will face severe reputation penalties, successfully extending the security boundary of blockchain's PoS economy to the entire agent society.

Illia Polosukhin: In summary, AI is a new super front end for human interaction with the computational world, while blockchain is the ultimate underlying backend for collaborative trust and value distribution between people, enterprises, and AI agents. For a long time, the most criticized aspect of the crypto industry has been the fragmented user experience and towering cognitive barriers. Through near.com, this comprehensive cutting-edge window, we are completely encapsulating chain abstraction, cross-chain intent routing, and decentralized secret reasoning at the underlying level. Users don't even need to know what private keys and gas fees are; they just need to input their intent in natural language, and hundreds of dedicated agents will achieve your goals in seconds under the secure enclave and the high-speed operation of the NEAR clearing network. This is the cradle of the agent economy we are striving to build!

Host: Thank you very much, Illia, Alex, and Anton for joining The Rollup today and bringing us this high-capacity intellectual technology discussion. We firmly believe that the future of confidential AI and autonomous agents is accelerating towards us. Thank you all, and see you in the next episode!