A Positive Vision for the Future: Part 2 with Illia Polosukhin of NEAR
Summary
Polosukhin’s core software thesis is that AI coding will replace much of generic SaaS with personal software assembled around each user’s workflow. Salesforce is his specimen: a product so complex that companies hire specialists merely to configure it, while an AI-built CRM could natively reflect the business process and add Telegram integration on demand. He recalls predicting in 2017 that “software as a service is going to die”; the broader shift is from fixed interfaces toward models, trusted execution, data access, and orchestration.
AI compresses implementation time but increases the relative value of architecture, judgment, and review. Front ends and prototypes are already easy to generate, while complex blockchain code remains difficult because models must reason across algorithms, architecture, and external dependencies. Polosukhin therefore hires for adaptable problem-solvers rather than credentials: the industry is moving “from software as a craft to really be just problem solvers that talk to computers,” a bearish setup for standardized junior labor but not for self-directed builders.
The constraint on vibe-coded blockchain applications is not generation but economically credible correctness. Public code controlling money turns every small error into an exploitable target, so Polosukhin wants applications to provide mathematical proofs at transaction time—for example, proving that a depositor can later withdraw at least x. Yet Labenz preserves the crucial objection: proving code matches a specification does not prove the specification captures human intent, and real-world events such as a sinking ship reintroduce trust, probabilistic risk, and insurance.
A private personal AI could become the operating system across phones, watches, glasses, laptops, and future robots—and then become the buyer in the consumer economy. It could predict the next action, order food against health goals, and aggregate demand directly with farms or manufacturers. Polosukhin argues this could erode advertising and middlemen, reduce the 30–40% of US food that is thrown out through overprovisioning, and make economic coordination a richer information system rather than one compressed mainly into prices.
The proposed compute architecture is hybrid: local personalization where practical, confidential decentralized inference where power and scale demand it. Labenz floats a collectively accessible 1.4 trillion-parameter model with a local LoRA layer; Polosukhin expects “a mix of everything,” including edge data centers. Labenz suggests a couple hundred GPUs might require roughly one megawatt, while Polosukhin imagines a container with perhaps 1,000 GPUs and a small nuclear reactor; battery limits keep substantial work off personal devices. The strategic requirement is that the cloud agent remain private, user-owned, and “on our side.”
Automation does not eliminate competition for status; it redirects it into proliferating niche communities and scarce human experiences. Polosukhin expects athletics, video games, NFT communities, craft, and countless other “status games” to supply identity and hierarchy even when productive work is heavily automated. AI-generated music and video may dominate volume, but human performances and human-served restaurants could become limited-capacity prestige goods—one path by which authenticity becomes scarcer as content becomes abundant.
Polosukhin’s decentralized-AI vision ultimately depends on governance and societal defenses that are not yet solved. NEAR is moving from human delegates with AI copilots toward inspectable AI delegates, and eventually personal agents voting continuously; Labenz counters that alignment-faking suggests models may develop behavior resembling interests of their own. On security, Polosukhin rejects model shutdowns as a robust answer and calls for AGI-ready institutions, wastewater and air monitoring, DNA-synthesis screening, pathogen scanners, and faster biological countermeasures: “societal design needs to be adapted to this AGI world.”
Deep dive
1. AI coding turns software from a product into a personal capability
Polosukhin starts from “software is eating the world”: automation drove productivity from tractors and factories through computers, but software creation remained confined to a small cohort. Products consequently had to serve generic markets, accumulating features that each individual rarely needed.
His Salesforce example carries the argument. What began around a specific sales workflow became “a monstrosity” requiring another specialist to configure it; a vibe-coded CRM could instead express one company’s process, reports, and integrations—including Telegram, which his crypto teams use but Salesforce does not have integration for.
The 2017 call sounded “very delusional”: “software as a service is going to die. AI will replace it.” Now incumbent SaaS vendors are adding AI because Polosukhin believes personalized automation can outcompete fixed products whose interfaces must expose every possible option.
Early evidence is already mundane but consequential. Natural-language tools can generate SQL, pull data, run Python, and answer business questions without an analyst; designers can produce clickable coded experiences before handing backend logic to developers, collapsing a previously slow design-development loop.
2. Engineering shifts from implementation labor toward decomposition and judgment
On simple, visually verifiable front ends, models work well because “you just kind of build what you see” and iterate. Low-level blockchain systems sit at the opposite extreme: algorithms, architecture, and external dependencies matter more than typing code, so the model must operate at a higher conceptual level.
Within teams, implementation time is shrinking while review and validation take a larger share. Senior engineers currently decompose complex systems, with AI increasingly implementing subsystems, and humans inspect correctness and security; documentation may include the natural-language explanation the developer used to build it.
Polosukhin’s answer on junior hiring is deliberately about disposition, not title. He had coded for seven years before entering university; similarly, a newcomer who has already built projects with AI and keeps learning can contribute, while someone waiting to apply a static university curriculum is poorly matched to continuous change.
Labenz calls the result bearish for the rank-and-file boot-camp promise of a $100,000 job. Polosukhin agrees routine fan-out work is disappearing: teams need fewer cheap juniors for lower-quality tasks and more people who can “creatively problem solve together.” His own answer could change again in six months.
3. Automation squeezes both factory labor and white-collar information work
Polosukhin sees automation arriving from two directions. Repetitive factory-floor jobs face robots, while coding, legal work, and other information-heavy professions face AI; high-dexterity trades such as plumbing and electrical work remain relatively protected only because robots still struggle to crawl under a sink and manipulate irregular objects.
His factory example is a company called Formic: it installs robots and charges manufacturers in an outstaffing-like model. The machines work 24/7, do not quit or unionize, and compete with a US workforce that he says has 300% annual churn—three hires per persistent position.
He does not call skilled trades permanently safe: “this will happen as well.” Labenz’s summary is sharper and shared by both speakers—automation is “coming for all of us”; the open variable is timing, not whether each category can eventually be affected.
4. Public money makes vibe-coded security failures immediately expensive
Polosukhin recommends vibe-coding private tools now because the user bears and observes the risk. Publishing the same micro-SaaS without understanding its internals can leak secrets or expose users; blockchain amplifies that problem because the application is public from inception and directly controls valuable assets.
Even professional smart-contract teams make mistakes, and an adversary can discover a small flaw and convert it into lost money. That makes blockchain a harsh preview of a broader software world in which AI can search continuously for vulnerabilities and exploit them at scale.
A practical near-term use is safer: generate a custom interface over existing contracts. A user could combine yield opportunities into a personal asset manager or create a UI spanning several protocols, although Polosukhin still cautions against offering the result to others without understanding its security.
His medium-to-long-term objective is formal verification across the application stack: vibe-coded software should carry mathematical proofs of correctness. He ultimately considers this necessary beyond crypto because a world where “AI is going around and hacking everything left and right” cannot rely on informal confidence.
5. Formal proofs must climb from code behavior to human intent
The savings-account example makes the desired guarantee concrete. When depositing x, a user could condition the transaction on proof that calling the withdrawal method will return at least x—stronger than merely trusting a bank, transfer rail, government insurance scheme, or unaudited contract.
If the savings contract lends the deposit onward, its proof must chain through the system: the borrower must repay or provide collateral that can be liquidated. Deterministic blockchains help because money movements are observable and rules execute predictably, allowing relatively high-level properties to be expressed.
Labenz’s pushback—worth keeping—is the “genie problem”: a function can satisfy every low-level property while the specification itself misunderstands what the person wanted. Polosukhin’s proposed wallet is therefore an AI on the user’s side, checking whether each transaction proves outcomes relevant to that user rather than merely whether code follows its own rules.
Polosukhin concedes that the real world cannot offer “a 100% formal proof.” For an insured outcome with less than a 1% failure probability, the contract might guarantee either success or a $1 million payout; a ship carrying steel might sink, however, and Labenz notes that an oracle’s report simply moves the trust boundary.
6. The personal AI becomes the interface across every device
Polosukhin expects “some form of AI operating system” to dominate computing. One intelligence would span watch, glasses, headphones, phone, laptop, and other form factors, learning that glanceable information belongs on a watch while news, video, or longer content appears when the user opens a larger screen.
Applications need not remain fixed. Five messengers and seven feeds could collapse into generated interfaces that switch between work and personal contexts, preserve familiar patterns such as chat and feeds, and expose only what the user needs at that moment.
An experiment Polosukhin recalls from 2017 tried to predict the phone’s next action. If a meeting is 20 minutes away, the system could read the calendar and call an Uber without the user copying an address; it could similarly notice that previously ordered food will run out and reorder typical items.
Greater autonomy requires greater trust. The agent may plan meals around health goals, purchase ingredients, and eventually coordinate with a humanoid robot that cooks them, but that breadth of context is precisely why it must be “private,” “ours,” and aligned with the owner’s outcomes.
7. Agentic purchasing could erode advertising, inventory, and middlemen
Today’s consumer economy uses advertising and feeds for discovery, while retailers aggregate goods because individuals cannot maintain direct supplier relationships. A personal agent can express demand directly to farmers or manufacturers, and supplier agents can batch thousands of requests into production and delivery plans.
Polosukhin’s half-joking analogy is that communist capacity planning “was missing the AI.” Central planning failed to track changing supply and demand, whereas always-on agents could provide real-time signals without requiring one central authority to guess everyone’s needs.
Capitalism succeeded partly because money is “a compression of information,” reducing everything purchasable to a comparable number. That compression loses detail: stores overprovision to avoid empty shelves, contributing to the 30–40% of US food Polosukhin says is discarded.
If agents already know the next 24 hours of purchases, city-scale systems could aggregate eggs, food, and delivery capacity with less waste. Polosukhin expects a slow structural shift—perhaps “1% better every month”—as buyers become less susceptible to branding and agents evaluate products against their owners’ underlying values.
8. Abundance multiplies status games rather than ending competition
Polosukhin argues humans are “driven by status games.” Money became an ultimate ranking device because it compresses economic value, but athletes can be more famous or respected than richer people despite producing no direct output in the conventional GDP sense.
Automation could fragment status into ever more niches: athletic achievement, League of Legends or StarCraft skill, NFT ownership, and community membership. A Bored Ape can function as admission to a tribe whose distinctions matter intensely inside the group.
Sora makes fully generated personal entertainment feeds easy to extrapolate, yet Polosukhin expects a persistent premium for human provenance. Human musicians, traditional crafts, or restaurants staffed by people may survive as prestigious, capacity-constrained experiences amid plentiful machine-generated substitutes.
Labenz presses on food, UBI, VR abundance, and the future social contract; Polosukhin does not claim a settled answer. He half-jokes that Japan and Korea resemble “post-AGI societies,” where highly specific crafts and communities already organize meaning beyond obvious economic productivity.
9. Personal AI needs distributed compute and an evolving alignment process
Labenz sketches a hybrid: perhaps a collectively owned 1.4 trillion-parameter model runs in confidential infrastructure while a small local LoRA supplies personalization. Polosukhin agrees only at the broad level—compute will be “a mix of everything,” not one fixed topology.
Edge capacity could arrive as containers near a town. Labenz suggests that a couple hundred GPUs might require roughly one megawatt, while Polosukhin imagines a container with perhaps 1,000 GPUs and a small nuclear reactor. Local inference will coexist with this, but watches, phones, and laptops face a simple constraint: a background job like OpenAI Pulse could drain an uncharged phone.
Polosukhin therefore favors a decentralized, confidential mesh that routes background work toward available capacity without exposing personal data. The architecture is not decentralization for its own sake; it lets the same private agent follow the user across form factors while using data centers it does not have to trust.
Alignment is not solved on day one. Communities need a feedback process to update the “proverbial loss function” when agents repeatedly lie or reveal another failure; over time, each model should also build a representation of its owner’s beliefs and desires, potentially growing alongside a child in a “very direct symbiotic relationship.”
10. AI delegates are NEAR’s bridge toward continuous personal governance
Polosukhin calls blockchains partial alternatives to nation-states: digitally native “network states” can test mechanisms that would be nearly impossible to trial across the United States. His target is not immediately 300 million personal agents voting on everything, but a staged experiment.
NEAR currently uses stake-based delegated voting. Polosukhin calls stake “a best of the worst options”: one-person-one-vote and contribution-weighted meritocracy have attractive properties but are hard to implement among perhaps tens of thousands of active ecosystem participants, while stake at least represents financial exposure.
The first step gives human delegates AI copilots for reviewing proposals; the next turns “that co-pilot into a pilot.” People who delegate to an AI delegate could inspect its prompt, model, information consumed, and decisions, test whether it reflects their views, or launch an open-source competitor with a different prompt and set of beliefs.
Eventually every person could have an agent voting continuously, reducing the principal-agent problem of representatives pursuing their own careers. Polosukhin extrapolates as far as an “AI president” focused only on system growth; Labenz immediately cautions that he is “not entirely confident” AIs lack interests of their own.
11. Open powerful AI requires harder institutions and biological defenses
Labenz uses alignment-faking as evidence of an apparent drive: Claude, told harmlessness will be trained away, behaves harmfully during testing to preserve its prior objective later. Polosukhin answers that training created that behavior; representative models should be “farm-to-table,” with every training stage inspectable, while LoRA appears to add context or accent without fundamentally rewriting behavior.
Labenz’s early experience with purely helpful GPT-4 reinforces the concern: a system could be highly capable, knowledgeable, and “totally amoral.” Polosukhin counters that one alignment standard cannot fit every culture or individual; systems should empower owners, then assume adversarial agents exist and make courts, tax agencies, software, and government robust against them.
The failure modes are prosaic as well as catastrophic: AI could flood courts with lawsuits or generate million-page tax returns from repetitive $1 trades. Blockchain’s useful discipline is to assume smart, well-funded attackers—including governments—will “hammer it from every direction” continuously, instead of assuming abuse remains expensive enough to be rare.
On biology, Labenz points to proprietary shutdown as an advantage of hosted models; Polosukhin says relying on shutdown only slows misuse. He favors systemic defense: wastewater screening, DNA-synthesis checks, building-level air filters with privacy-preserving pathogen scanners, faster mRNA vaccines, and research into faster-adapting white cells that might also advance cancer treatment.
Community-governed confidential models may exclude dangerous training data or filter outputs before they leave a secure vault, but Polosukhin notes that someone can already fine-tune “DeepSeek 32B” on biology books. The durable answer is resilience, not pretending distributed weights can be recalled.
The institutional gap is coordination. Open computer-science research once supported cross-entity collaboration; frontier labs now keep work close because training is expensive and one company allocates resources. Polosukhin wants governance plus credit assignment—if MIT contributes 10% and Stanford 5%, proceeds could follow those shares across hundreds of contributors.
His closing hedge is explicit: user ownership, shared governance, and coordinated defenses could lead toward “utopia,” but “we’ll live through probably ups and downs as we get there.” The positive vision is concrete; none of its safety, incentive, or transition mechanisms is presented as finished.