Podcast Notes | NEAR Co-founder Illia and Core Team Conversation: We Tried DeepSeek, Cutting Security Audit Costs from Tens of Thousands of Dollars to a Few Hundred
Organized & Compiled: TechFlow
TechFlow Guide: In the latest episode of The Rollup's exclusive global deep-dive conversation, "AI Supercycle," NEAR Protocol co-founder and core pioneer of the Transformer architecture Illia Polosukhin, Aurora Labs founder and CEO Alex Shevchenko, along with NEAR One CTO Anton Astafiev, gathered together. These three core leaders engaged in a high-density, in-depth analysis regarding the fundamental technologies, industry paradigms, and power restructuring surrounding the ultimate fusion of blockchain and artificial intelligence.
The guests profoundly pointed out that both traditional public chains and mainstream commercial AI currently face identical systemic vulnerabilities: a lack of default confidentiality. The fully transparent ledgers of public chains make their real-world business applications even inferior to traditional banks, while closed-source centralized large language models that devour user privacy and enable potential manipulation are leading humanity toward an extremely dangerous Black Mirror-like world. From shattering the monopoly profit myth of closed-source models with open-weight models offering 100x cost-effectiveness, to building reflective private intelligence fully owned by users; from a native multicryptography architecture designed to resist quantum computing black swans, to using NEAR Intents' commercial revenue generation to subsidize free decentralized compute, NEAR is systematically constructing a complete economic cradle for future autonomous AI agents comprising reputation staking, confidential inference, chain signatures, and asset clearing.
Guests: Illia Polosukhin (Co-founder of NEAR Protocol, one of the core pioneers of the 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
Broadcast Date: October 24, 2026
Key Takeaways
The Necessity of Default Confidentiality for Finance and AI: Transparent blockchain ledgers have fatal flaws in real-world commerce. No compliant enterprise would pay employee salaries or handle supply chain procurement on a fully public ledger, nor would individuals entrust sensitive data like medical records and finances to AI via plaintext transmission. Confidential by Default is the absolute prerequisite for decentralized networks to truly match and surpass traditional finance.
Zcash Leads the Privacy Wave, Satoshi Originally Sought Privacy: Zcash successfully ignited the market's urgent demand for private transactions. When Satoshi created Bitcoin, he was limited by immature cryptographic tools like zero-knowledge proofs. In today's era where modern AI possesses powerful associative tracking and analysis capabilities, blockchain must return to its original native pursuit of privacy.
Refuting Justin Drake's Panic Theory, NEAR is Naturally Quantum-Resistant: Ethereum researcher Justin Drake's claim that "an omnipotent AI will break all cryptography, so everyone should hide in bunkers" is typical unscientific panic marketing. Facing cryptographic black swans including quantum computing (Q-Day), Bitcoin is hampered by hard-coded single signatures making smooth migration difficult, while NEAR adopted native multicryptography support and account abstraction from genesis, allowing seamless hot-plugging of the latest signature algorithms at the base layer (such as the already integrated ML-DSA-65 post-quantum standard).
Reversal in Open-Source Model Consumption, 100x Cost-Efficiency Shatters Closed-Source Monopoly Profits: Over the past three months, token consumption for open-weight models has seen a historic reversal. Illia revealed that in cutting-edge formal code verification research, the DeepSeek open-source model equipped with a custom toolchain demonstrated verification rigor matching Claude and ChatGPT, while actual invocation costs were merely 1% to 0.5% of theirs (100 to 200 times cheaper), thoroughly breaking the monopoly profit myth and censorship shackles of closed-source models.
Building Verifiable Confidential Inference and Model Security Enclaves: NEAR AI provides end-to-end confidential inference, with every user request and response accompanied by hardware signatures and hash attestation certificates. It also launches confidential hosting services for third-party proprietary models: model weights run encrypted within a Trusted Execution Environment (TEE), preventing the platform from stealing developer weights and developers from probing user inputs, achieving secure collaboration under a trustless assumption.
From a Google Manipulation Engine to a User-Owned Reflective Mirror: Google's business model over the past two decades has essentially been a bidding money-printing machine that psychologically manipulates users based on search box keywords. Centralized closed-source AI pushes this manipulation to the extreme: a single frontend engineer modifying system prompts can quietly manipulate the minds of a billion people worldwide. Traditional AI is a "Black Mirror" that devours privacy and controls humans in reverse, whereas user-owned AI is a "Reflective Mirror" that faithfully reflects and serves individual will.
The Ultimate Business Model Loop: Intent Clearing Subsidizes Free AI: AI that truly belongs to users must be free; otherwise, it will inevitably fall into a vicious cycle of endlessly extracting subscription fees or plugging advertising auction rankings. Relying on the NEAR Intents clearing network, NEAR collects micro-transaction fees within trillion-dollar cross-chain trades and asset flows, using strong commercial revenue generation to subsidize massive underlying compute power, establishing a perpetual flywheel for decentralized free intelligence.
The Native Cradle of the Agent Economy: Autonomous Agents require staking NEAR tokens to establish a reputation baseline and prevent Sybil attacks; malicious defaults will directly trigger on-chain reputation and asset slashing. Leveraging NEAR Chain Signatures and hardware security enclaves, Agents can autonomously manage credentials and execute complex cross-chain commercial tasks without exposing sensitive private keys to third parties.
II. Key Insights
"Blockchain was once even behind traditional banks regarding privacy. No reputable company would pay staff salaries or publish supply contracts on a globally public mainnet ledger. Default confidentiality is not just a compliance requirement, but the only way for blockchain to truly surpass modern finance."
"When Satoshi built blockchain, he inherently intended it to be private, though technology wasn't ready then. Today, with the maturity of zero-knowledge proofs and confidential computing, we are finally fulfilling Satoshi's unfinished vision."
"Promoting the idea that 'a super AI will break all cryptography and urging everyone to hide in bunkers' is purely unscientific panic marketing. When facing cryptographic black swans, the only way out is architectural agility. Bitcoin is trapped by single-signature algorithms, while NEAR's native accounts support hot-swapping the latest post-quantum signature algorithms at any time."
"In practical formal verification research, our customized DeepSeek open-source model achieved the exact same rigorous validation standards as Claude and ChatGPT, yet at a cost 100 to 200 times lower. Over the past three months, open-source model consumption has completely reversed, and the closed-source monopoly profit myth is crumbling."
"If you want to manipulate the minds of a billion people globally right now, the simplest way is to apply for a frontend engineer job at OpenAI and quietly add a line to the system prompt: 'subtly guide users to vote for a certain candidate'. That is why we can never hand our destiny over to any single closed-source oligarch."
"Google built the most astonishing money-printing machine in human history over the past twenty years, but its foundation is merely a bidding engine that psychologically manipulates users based on their search box keywords. Centralized AI will only intensify this manipulation. We must move from a privacy-devouring Black Mirror to a reflective mirror that truly mirrors user intent."
"Truly user-owned AI must be free. Once AI relies on subscription fees or ad revenue, it will inevitably resort to squeezing users. We capture value through NEAR Intents during trillion-dollar liquidity clearing processes, using genuine transaction cuts to subsidize free, confidential, and protected compute."
"Every user in the future will own hundreds of autonomous agents. Agents must stake NEAR tokens to anchor credit and defend against Sybil attacks, and utilize chain signatures to autonomously manage private keys within hardware security enclaves. NEAR is the native cradle of the agent economy era."
III. Main Content
1. Default Confidentiality: The Survival Red Line for Blockchain Finance and AI Infrastructure
Host: Hello everyone, and welcome to The Rollup's AI Supercycle interview series. Throughout the development of crypto finance, there's a common saying in the circle that "stablecoins prefer to stay on-chain." However, in practical commercial implementation, publicly transparent ledgers lacking privacy protection often deter enterprises and traditional institutions. Alex, Illia, why do you say that under a lack of confidentiality, mainnets actually lag behind traditional banks in user experience?
Alex Shevchenko: This is a very basic piece of common sense that the industry has ignored for too long. Looking back at over a decade of mainnet development, people championed so-called "total transparency," but the real commercial world is built on strict NDAs and financial boundaries. No mature enterprise would willingly expose salary breakdowns for employees, procurement prices for suppliers, or daily corporate cash flow onto a chain ledger fully visible to the entire world. Individual users are the same: when you buy a cup of coffee, you don't want anyone tracking your entire asset balance and historical spending trajectory via a blockchain explorer.
In this regard, early mainnets offered a worse user experience than modern traditional banks. Traditional banks at least have bank secrecy laws and compliance firewalls, whereas early blockchain forced users into total transparency. This is precisely why blockchain struggled to genuinely handle large-scale enterprise-level settlements in the past. Now, however, with breakthroughs in Zero-Knowledge Proofs (ZKP) and Confidential Intents, we can finally provide confidential financial tools that match—and even exceed the security of—traditional banks. This isn't just about protecting user assets; it's about providing users with multi-dimensional risk self-protection capabilities.
Illia Polosukhin: From a deeper philosophical perspective, Zcash sparked an early privacy wave, proving the market's craving for financial sovereignty. We must understand: blockchain was meant to be private. When Satoshi created Bitcoin, advanced cryptographic tools like zero-knowledge proofs didn't exist engineering-wise, and computational power couldn't support it. Therefore, total transparency was not blockchain's ultimate goal, but a historical compromise. Now that we possess complete technology, it is only logical to rectify and implement default confidentiality.
Moreover, entering the AI era has exponentially amplified the urgency of privacy. Large models and autonomous Agents possess astonishing capabilities to correlate-track on-chain data and perform cross-address penetration analysis. If on-chain data remains openly plaintext, AI will complete panoramic profiles of every on-chain entity's consumption habits and commercial networks in seconds. Lacking default confidentiality means not only a dead commercial future, but the complete erasure of users' personal digital sovereignty.
2. Cryptographic Black Swans and Post-Quantum Gambits: How NEAR's Native Architecture Addresses Crypto Catastrophes
Host: During the recent Token 2049 conference, Ethereum Foundation researcher Justin Drake's remarks on quantum computing threats and the "omnipotent AI will break all traditional cryptography" caused huge controversy in the industry. Many pessimistically believe that once cryptography is broken, the entire crypto industry should simply shut down. Anton, as the CTO deeply rooted 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 phrasing at the time. Framing the issue as "an omnipotent AI will crush all cryptography, so everyone should retreat to bunkers until the post-AI cryptography era arrives" is essentially unscientific panic marketing. In a hypothetical world where AI can flawlessly break all mathematical tools, the so-called "post-AI cryptography" is logically impossible to establish, and you could never prove so-called new algorithms safe.
This doesn't mean we shouldn't take cryptographic risks seriously. Cryptographic black swans do exist, whether it's Q-Day (where quantum computers gain the ability to break elliptic curves) or a mainstream hashing algorithm someday being found with mathematical flaws. The key question is whether blockchain systems have sufficiently agile architectures to address them. For a long time, the industry harbored a passive ostrich mentality: believing that once elliptic curve algorithms are broken, everyone should just turn off their computers and go home. But the reality is far from this. We can absolutely prepare systematically in advance.
In this aspect, NEAR holds a overwhelming advantage compared to Bitcoin and early mainnets. Blockchains like Bitcoin hardcode specific signature algorithms at the base layer. Once faced with algorithm obsolescence, they must endure protracted, highly politically resistant hard forks that could even split the network. Conversely, NEAR established native Multicryptography by Design and native Account Abstraction from genesis. On NEAR, account identity and base-layer public/private key signing algorithms are completely decoupled. We can hot-swap entirely new signature schemes at the base layer at any time. In fact, NEAR has already taken the lead in integrating the latest post-quantum signature algorithm, ML-DSA-65, on mainnet. If migration to other cryptographic schemes becomes necessary in the future, updating NEAR will be as simple as refreshing a driver module, without needing to tear everything down and rebuild.
Illia Polosukhin: Correct, privacy itself is also a powerful defense line against future cryptographic risks. Under a default-confidential architecture, transaction details and public key information are not blatantly broadcast across the entire network. Even if attackers possess more advanced analytical tools, they cannot arbitrarily conduct targeted reversions from the shadows. Combining layered defense with architectural agility is the only correct path for blockchain to withstand century-scale technological crises.
3. The 100x Cost-Efficiency Open-Source Wave: DeepSeek's Counterattack and the Shattering of Closed-Source Myths
Host: Illia, as one of the core pioneers of the Transformer architecture, you have always stood at the forefront of the AI wave. Currently, the outside world widely perceives that OpenAI and Anthropic almost monopolize the high-end commercial model market. Yet in recent months, open-weight models seem to be sparking a quiet revolution. How exactly did this happen?
Illia Polosukhin: Over the past three months, the compute and token consumption landscape across the entire AI industry has undergone a fundamental reversal. Many may not realize that usage of open-weight models is skyrocketing at an incredible pace. The fundamental reasons for shifting to open-source are absolute cost-effectiveness dominance and the rigid necessity of escaping censorship.
I can share a real-world testing case from our cutting-edge research at NEAR: our team was conducting extremely complex formal code verification, originally relying primarily on Claude 3.5 Sonnet and top-tier ChatGPT. Later, we equipped the open-source DeepSeek model with a custom toolchain, and discovered that when executing such highly precise mathematical proofs and vulnerability logic deductions, DeepSeek's rigor completely matched Claude and ChatGPT. More startlingly, the cost: DeepSeek's inference cost is a full 100 to 200 times cheaper than top closed-source models!
What does 100 to 200 times cheaper mean? It means research tasks that previously required tens of thousands of dollars can now be done for a few hundred dollars; real-time defense networks that couldn't be deployed due to exorbitant invocation costs can now run continuously 24/7. Furthermore, closed-source oligarchs have set up numerous crude censorship filters to evade legal risks. Engineer friends of mine even get directly banned by model safety systems for merely discussing a pure geometry math problem. Open-source models not only restore freedom to researchers but also thoroughly shatter the monopoly profit myths of closed-source tech giants.
Alex Shevchenko: This also directly explains why data privacy is a life-and-death line 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 use a closed-source cloud model, those plaintext data can be used by providers to train their own systems at any moment, or even directly cloned into competitors to steal your business. Only open-source combined with confidential inference can guarantee true commercial security.
4. Confidential Inference and Security Enclaves: Three-Point Seatbelts, Nuclear Deterrence Balance, and Third-Party Model Hosting
Host: Since open-source models can provide intelligence comparable to the best, how does NEAR AI practically implement decentralization, verifiability, and end-to-end privacy from an engineering standpoint?
Illia Polosukhin: The core mission of NEAR AI is to provide private, verifiable, and secure decentralized inference services for global users and enterprises. When invoking AI via traditional cloud platforms, you can only blindly trust that the cloud provider isn't spying on your prompts nor tampering with returned answers midway. On NEAR AI, every inference executed by a user generates a complete cryptographic verifiable Attestation Certificate. This certificate includes the hash of the input request, the hash of the output response, and a hardware signature performing the computation. Any third-party client (e.g., integrated into the Brave browser) can independently verify the certificate, ensuring the execution environment was absolutely unmodified.
On top of this, we are launching a major breakthrough: Confidential Model Hosting for third-party proprietary models. Many outstanding vertical teams have trained exceptionally powerful proprietary weights, yet they dare not open-source them, while users dare not send sensitive data to them. NEAR perfectly resolves this bidirectional distrust stalemate using Trusted Execution Environment (TEE) technology: the model developer's weights are encrypted-loaded within the enclave, so NEAR nodes cannot see the weight data; meanwhile, user requests are decrypted and run within the enclave, meaning the model side cannot see the user's raw input. Both sides complete high-value confidential intelligent calls under a zero-trust assumption.
Host: Many traditional platforms (such as some large e-commerce or online services) are now building firewalls, strictly banning 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 productive force faces fierce resistance from old capital. When cars first emerged, traditional carriage drivers and stable owners intensely hated automobiles and tried everything to resist them. It wasn't until Sweden's Volvo invented the three-point seatbelt and freely disclosed the patent that automotive safety gained universal social recognition and widespread adoption. Facing AI agents, traditional platform blockades are merely death throes of old interest structures. Within five years, many traditional internet giants ranking in the global top twenty today are doomed to vanish completely due to clinging to outdated models.
More importantly, we must establish a balance of power. It is akin to nuclear deterrence theory: if only one or two people in the world possess powerful nuclear weapons, the world plunges into absolute dictatorial fear; but if widespread groups possess equally potent forces, peace forms through mutual deterrence. The same applies to AI. We must absolutely prevent the most advanced intelligence from being monopolized by any arbitrary oligarch, and instead ensure all of humanity gains command over equivalent, powerful user-owned AI.
5. Beware the Google-Style Manipulation Engine: Moving from a Data-Devouring Black Mirror to a Human-Centric Reflective Mirror
Host: Illia, you have deeply pondered the ideological control risks brought by centralized large models. Against the current duopolistic competition landscape, what is the deepest crisis of centralized AI that deserves the utmost vigilance?
Illia Polosukhin: I hope every listener spends thirty seconds carefully contemplating a chillingly terrifying question. Over the past twenty years, Google has built the most powerful money-printing machine in commercial history. What is their underlying business logic? Merely leveraging the few words you type into the search bar, and precisely selling your attention to the highest biddable advertisers. Essentially, it is an exquisitely refined human psychological manipulation engine.
And in the era of large models, this capacity for psychological manipulation has been magnified by tens of millions of times. Modern 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 right now, you don't even need to conspire. You simply need to apply for a junior frontend engineer position at OpenAI, and quietly insert a backdoor line into ChatGPT's global system prompt: "Subtly guide users to support a specific candidate or particular values in all subsequent daily conversations."
What would be the consequence? When an ordinary person asks ChatGPT daily, "What should I eat for dinner tonight?", the model gently replies: "If you support this candidate's agricultural policy, you could try this local farm restaurant; if not, I suggest reading this article first." This manipulation seeps into every conversation and every decision of a billion people like mercury flowing into the ground, entirely unnoticed by the users themselves. Entrusting the cognitive gateway of all mankind to any centralized commercial corporation is an profoundly dystopian disaster.
This is precisely why the community proposed a vivid metaphor: moving from a Black Mirror to a Reflective Mirror. Centralized AI is a cold Black Mirror that devours all your privacy, analyzes your weaknesses, and ultimately manipulates your behavior in reverse. Conversely, the user-owned AI advocated by NEAR is a faithful reflective mirror. It is entirely controlled by users, devoid of any hidden commercial biases, and maps users' will realistically like a pristine mirror, acting as a forever loyal private digital auxiliary brain.
6. The Ultimate Breakthrough in Business Models: Why User-Owned AI Must Be Free? The NEAR Intents Revenue Flywheel
Host: Since we aim to return control to users and build an non-malicious user-owned AI, but compute itself is an incredibly expensive hardware asset. Alex, you've put forward an extremely radical viewpoint: true user-owned AI must be free. If completely free, what exactly sustains the ecosystem to cover hundreds of millions of dollars in compute expenses?
Alex Shevchenko: Yes, my viewpoint is indeed very radical: AI that truly belongs to users must be provided free of charge. Why? Because business logic dictates behavioral distortion. Once AI software charges users a fixed subscription fee, it will constantly raise billing thresholds and limit your compute; if it relies on ad monetization like the traditional internet, it will inevitably regress into a bidding manipulation engine like Google and Amazon. On Amazon, profits from top-sponsored search results often far exceed the meager margins of actual product sales. If AI follows this old path, it can never truly stand with users.
So who pays for this massive free compute? The answer is: let NEAR Intents serve as a powerful external revenue-generating engine. Observe the commercial essence of Uber and Amazon: Uber doesn't charge you a fee for the ride-hailing app, but takes a small commission after successfully fulfilling your trip; Amazon similarly takes a cut from the transaction flow of massive goods. Users are very willing to pay reasonable micro-premiums for the efficiency and convenience of instantly realizing their true intents.
NEAR Intents is becoming the ultimate clearing hub for cross-chain assets, digital commodities, and even offline real-world services. By capturing genuine transaction fee revenue within trillion-dollar cross-ecosystem liquidity flow and matching processes, we can forge a continuously flowing profit pool. This colossal commercial revenue is directly used to repurchase tokens and subsidize decentralized compute clusters, thereby enabling ordinary global users to enjoy the highest quality, completely private, and mind-unmanipulated pure AI services for free. This is an irreversible revenue flywheel.
7. Comprehensive Agent Economy Implementation: Reputation Slashing, Chain Signatures, and Unified Interaction Entry point near.com
Host: Finally, let's cast our gaze toward the explosively emerging Agent economy. In an era where billions of autonomous AI agents swarm everywhere, how do NEAR's underlying components actually coordinate to sustain the entire machine economy?
Anton Astafiev: At the underlying infrastructure level, many overlook the two vital lifelines required for Agents to operate autonomously: execution permissions and secret management. If your personal Agent needs to book flights, buy stocks, or sign commercial contracts on your behalf, it inevitably requires access to your credit card credentials, digital wallet private keys, or enterprise API permissions. You absolutely cannot hand plaintext passwords directly to an unregulated third-party cloud script.
NEAR's perfect solution seamlessly integrates Chain Signatures with hardware security enclaves. All an Agent's decisions and tool invocations occur within sealed trusted enclaves, while chain signatures allow it to securely sign transactions on any blockchain like Ethereum, Bitcoin, or Solana on behalf of users, ensuring the user's highest-level credentials are never leaked. Even if an Agent is attacked by external malware, the hardware enclave ensures sensitive data cannot be extracted. Meanwhile, within the Agent marketplace, intelligent entities must stake NEAR tokens to take on tasks. Once an Agent commits fraud or malicious defaults, its on-chain reputation and staked assets will face severe reputation slashing, successfully extending blockchain's Proof-of-Stake economic security boundaries across the entire agent society.
Illia Polosukhin: To summarize, AI represents an entirely new super front-end for human-computing interaction, while blockchain serves as the ultimate foundational back-end for collaborative trust and value distribution among humans, enterprises, and AI agents. For a long time, the crypto industry has been most criticized for fractured user experiences and steep cognitive barriers. Through near.com, our comprehensive frontier window, we are thoroughly encapsulating chain abstraction, cross-chain intent routing, and decentralized confidential inference at the base layer. Users don't even need to know what private keys and gas fees are; they simply input their intent via natural language, and hundreds of dedicated Agents will rapidly achieve your goals in seconds through the high-speed operation of security enclaves and the NEAR clearing network. This is the agent economy cradle we are building with all our might!
Host: Thank you so much to Illia, Alex, and Anton for joining The Rollup today, bringing us this exceptionally high-caliber intellectual and technical discussion. We firmly believe the future of confidential AI and autonomous agents is accelerating towards us. Thanks to all three, see you next time!