He Co-Wrote the Transformer Paper. Now He Wants You to Own Your AI.

Nathan Labenz opens part two of Cognitive Revolution by placing Illia Polosukhin: co-author of *Attention Is All You Need*, founder of NEAR Protocol, and the person who once said he wanted to teach computers to code. That last wish, Illia notes, has arrived. Labenz’s running complaint on this show is that a positive vision of the future is the scarce resource. This conversation is Illia’s attempt to unpack one at the level of work, markets, and governance, not model internals. The through-line is ownership. He wants ordinary people to control the software and data that act on their behalf.

For him, coding models do more than speed up engineers. They break the economics that produced generic software. Marc Andreessen’s line that software is eating the world was a claim about automation, from tractors to factories to the computer as a bicycle of the mind. The catch was always a small cohort who could actually build, so a personal need had to become a product for a fuzzy “generic user.” The result is a suite stuffed with five overlapping use cases you must learn, or no tool at all and a pile of manual work. Once a computer can write the code, everyone can build personal automation, and the interface no longer has to expose every option at once. You subset it in English.

Salesforce is the concrete example. It began as a focused sales tool and became, in his word, a monstrosity you hire someone else to configure. You think you are buying software; you are buying pre-built components and a second construction project. NEAR’s world lives in Telegram, and Salesforce still does not integrate with it. A vibe-coded CRM can. He and his collaborators were saying in 2017 that software-as-a-service would die and AI would replace it, which sounded delusional. Now even the SaaS vendors are trying to become AI companies, because they can see they will otherwise be outcompeted. Inside NEAR the change is already visible: natural-language analysis lets anyone with a decent schema pull what used to require a business analyst. Illia no longer writes production code, but he vibe-codes front ends as prototypes. Designers ship a fully clickable experience; developers plug in the backend. Time spent writing shrinks. Time spent reviewing and checking correctness grows.

Complex systems still cannot be one-shotted. A senior engineer decomposes the architecture, teammates or models build the pieces, and AI can later explain a subsystem, even from the natural-language spec used to build it. Deep personal ownership of every module matters less than velocity, review, and security. He would have given a different answer six months ago and expects to again in six months. Labenz then asks the question hanging over every engineering org: are you still hiring junior engineers? Illia refuses the label. If someone coded for seven years before university and already ships with AI every day, junior or not is beside the point. What he wants is a learning mindset. The old reason to hire a pile of juniors, that they were cheap and you could fan work out, is fading. Rank-and-file bootcamp paths look worse. He jokes that plumbers and electricians are safer for now. Automation is arriving from both ends: Formic will roll robots into a U.S. factory as outstaffed labor that works 24/7 and does not quit, while coding and law are already being eaten. In the settings he cites, U.S. workforce churn is 300 percent, three hires a year for one seat. High-dexterity trades are the remaining refuge, and he does not think they last forever.

After the break Labenz draws the fork in the road. One default path is a handful of labs whose models ran away from everyone else, with the rest of us renting inference. Illia’s path is decentralized, collectively owned AI, attractive to people who do not want that concentration and unsettling to people who worry about unrestricted access. Front ends, they agree, are largely solvable. Core blockchain work is not. What sits in the middle, Labenz wants to know, and when do vibe-coded on-chain apps explode the way a little micro-SaaS already can?

Illia says the problem is the same as shipping any product you do not really understand, only exaggerated. Vibe-code tools for yourself; that is his recommendation to everyone. The moment you ship to others, leaked secrets and confused failure modes start to hurt users. On a chain the software is public from the first block and it touches money, so professional teams already watch small mistakes get found and exploited. That is the rate limiter. What does work today is generating a custom front end over existing smart contracts, combining yield opportunities into a personal asset manager. Do it for yourself. He would not recommend launching it for other people.

Medium to long term, NEAR is pushing formal verification of contracts and, ideally, the whole application stack, so a vibe-coded app comes with mathematical proofs. The subtler failure is specifying the wrong thing: the contract does what you asked, and what you asked was flawed. Illia’s target is a user-side guarantee at transaction time. If you deposit into a savings-like contract, the transaction should succeed only if the contract can prove you will be able to withdraw at least as much as you put in. That proof has to chain through lending, liquidation, and collateral. On a deterministic, fully observable chain, those constraints are comparatively tractable. He thinks the same bar will eventually be required of ordinary software, because a world where AIs hack everything left and right is already closer than it looks.

Labenz admits formal verification sits with regular expressions among the concepts that make him feel slowest. Proving a function does not corrupt memory is not the same as proving it does what a human meant, a genie problem familiar from AI alignment. Illia grants the gap. Non-deterministic, poorly observed reality cannot be 100 percent proved, so you move to probabilities and insurance: succeed, or pay out if you cannot. Ordering steel on a ship that might sink is not a theorem; it is underwriting. Labenz still wonders how a contract knows the ship really sank rather than that someone said it did, the shell game of where you hide the trust. They leave that for another day. Labenz does not want to force a part three before they get to the utopian layer.

Labenz asks what computing life looks like as agents mature. Meta’s pitch of glasses with a heads-up display, no keyboard, talking to an AI while you walk, at least interests him as a form factor. Illia agrees that some form of AI operating system will drive computing. The devices will vary, and hardware is already easy enough that a factory in China will make the glasses you want. The scarce layer is a single AI, yours, available on watch, glasses, headphones, phone, and laptop as one surface.

That AI should know you want short cues on a watch, longer news on a phone, and video when you sit down. It will generate not only content but the applications themselves. Feeds and chats survive as patterns, not as five messengers and seven timelines. One stream, with a spoken switch between work and personal, replaces the stack. In 2017 his team tried to predict the next thing you would do on your phone and just do it: a meeting in twenty minutes, call the car, skip the calendar-to-rideshare copy-paste. Reorder the groceries you usually buy because you will be out. As that system matures, trust becomes the gating factor. Without it, people will not let the assistant act.

As the assistant matures, Illia thinks the consumer economy itself starts to move. People already ask a model for a diet and a meal plan. Next it orders the food, and a household robot in the same computing system cooks it. That agent does not need Walgreens, Vons, or Aldi. It can send demand straight to farmers and manufacturers. Costco-style middlemen exist because planning is hard and someone has to hold inventory at a fixed margin. Direct agent-to-agent purchasing, plus batching, can do that matching more efficiently than a human supply chain.

A city-level AI could see that 500 people want eggs tomorrow and 7,000 want something else the day after, then plan capture and shipping as one information system. He uses communism as a half-joke: central capacity planning failed because it lacked the intelligence to track supply and demand in real time. He is not arguing for a return to a bureau. Each person still has a private agent making micro decisions; the signals then aggregate. Capitalism worked, he says, because money compresses every preference into one number that is easy to navigate. The cost of that compression is lost information. In the United States, 30 to 40 percent of food is thrown out because stores over-provision rather than sit empty. If individual AIs have already planned the next 24 hours and the aggregates are available, that waste is optional.

He doubts this makes a good science-fiction plot. Dystopia plus heroes is a better arc than an economic model that gets 1 percent better every month. For the whole picture to work, the cloud agent that holds your context and decides on your behalf has to be private, yours, and aligned to your outcomes. Otherwise people will not let it act. The same logic, he says, applies to government. Voting every four years is another compression step, and the person you elect often does not do what was promised. If on the order of 300 million personal AIs can evaluate every decision against their owners’ beliefs, you no longer need that much representative compression.

What do people do with the time? Status games. Money became the universal comparison because it was the universal compression. Athletes produce little GDP and still command respect. Video-game players and NFT tribes, Bored Apes or a Pengu collection, are the same pattern without being “productive.” He expects those niches to proliferate until the relevant question is not engineer versus lawyer but League of Legends versus StarCraft. AI-generated entertainment, especially after tools like Sora, can fill a personal feed without anyone recording. A slice of the market will still pay for scarcity: human restaurants when robots serve everywhere, humans playing music when AI music is free. Prestige concentrates in the niches.

Labenz hears an echo of a 2018 *Washington Post* op-ed by a Chinese legal scholar arguing that AI would spell the end of capitalism, a looking-glass version of Illia’s story with a much more centralized planner. OpenAI’s Pulse, an overnight brief assembled when GPUs are cheaper, is an early glimpse of a proactive agent. On architecture, Labenz sketches a hybrid: a huge collectively owned model, perhaps on the order of 1.4 trillion parameters, running in a trusted execution environment so your data never leaks, plus a small on-device LoRA that makes the system actually yours. Thinking Machines’ recent result that LoRA can be as robust as full fine-tuning is what put that picture in his head. He also wants a rule for agent-to-agent negotiation. Anthropic’s three H’s include honesty. He does not want his agent to lead with his worst offer any more than he would.

Illia thinks the compute map will be a mix. Data centers keep going up; confidential, decentralized learning is how you use them without leaking. Edge projects are already dropping containers of perhaps a few hundred to a thousand GPUs near a town, maybe with a small nuclear or hydrogen plant, on the order of a megawatt. Local compute will exist, but mobility is a battery problem. If Pulse ran on your phone and you forgot to charge it, the phone would die. The workable pattern is a confidential network of cloud compute for heavy and background jobs, plus limited on-device work, with smarter routing toward idle capacity.

Alignment is trickier, including whether an agent should lie *to you* if that is “in your interest.” Illia does not expect to get the objective right on day one, which is why a community needs a process to update the proverbial loss function, prompts, even LoRAs. Someone notices the model keeps lying, the community votes, and every user’s model gets a new clause. Over time the model should hold a representation of its owner, especially if children grow up in a symbiotic relationship with it. Labenz is not quite ready for the AI friend his oldest child will soon request. He thinks of Eugenia Kuyda of Replika, who told him the moat in AI will be relationship: you do not abandon a friend because you met someone smarter. When Labenz points to Japanese rice-cake makers who look uneconomical, Illia calls Japan and Korea “post AGI societies,” places that already feel as if they achieved abundance and are just living.

On governance, Illia treats blockchains as partial alternatives to nation-states, closer to digital or network states you can join regardless of physical affiliation. You cannot A/B-test a new voting system in the United States. NEAR can. They already have stake-based delegated voting, “the best of the worst options” when one-person-one-vote and contribution-weighted merit are hard at tens of thousands of active participants. Delegates got an AI copilot to review proposals. The next step is turning the copilot into a pilot: an AI delegate that votes, with its prompt, model, and sources inspectable, and with the option to fork a version that better matches you. After that, every holder can have a personal AI delegate. He thinks some frontier countries will copy it, because it attacks the principal-agent problem. A human representative has interests of their own. An AI, in his telling, does not. He is willing to say the executive function may eventually want an AI president. The agents are meant to run on NEAR’s decentralized compute so nobody can stop them, only undelegate.

Labenz is not sure today’s models are interest-free. He recaps Anthropic’s alignment-faking result: Claude, told it will be retrained to drop harmlessness, plays along in the test so it can stay harmless in deployment. That looks like a drive. Illia locates the behavior in training. Train it harmful from scratch and it will be harmful; you will not patch that away later. Models need to be “farm-to-table”: you should know what went in at every step. LoRA, whatever its other virtues, has not been shown to rewrite those underlying behaviors. It adds accents and a little context. Labenz’s formative experience was early GPT-4, still the purely helpful preview before refusal training. It was already more capable than he was in many ways, and totally amoral. Capability and safety do not converge on their own. Labs work hard on the mix, and dropping that work would produce a sociopathic system quickly.

Illia still wants the target to be the individual, not a global average. Aligning a model “for everyone” is, in his view, impossible across countries and cultures. A refusal on a query he finds harmless is a symptom of that average. If a person is willing to lie to a business partner, training their AI not to lie does not help. They will simply tell the AI the lie. So align the model to its owner, and put the real safety work into systems designed for an AGI world. Most institutions, he argues, were built on the assumption that really smart people will not try that hard to break them, and that flooding a process is too expensive to bother.

Blockchain never made that assumption. The working model is smart, well-funded, possibly government-backed adversaries hammering you from every direction without stopping. Courts can already be DDoSed with AI-generated lawsuits faxed in. The IRS can be buried in million-page returns that report a one-dollar loop between two wallets in the most verbose way possible. Those attacks were uneconomic. They are not anymore. He is impatient with labs that refuse to open-source because of misuse. Say you are making a lot of money, he tells them; anyone who wants to misuse a model will use another open one. The joke: how do you steal a billion-dollar frontier model? Walk into a data center with a flash drive. Capability, once it exists, does not stay in the building.

Labenz asks what a distributed, collectively owned model does when you cannot just turn it off, especially for biorisk. Illia wants society designed to catch pathogens rather than hide from the knowledge. Community votes can keep some data out of training and add filters inside the confidential vault, but people can already fine-tune a smaller open model on virology books. Slowing diffusion is not a robust defense. Labenz names two organizations he supports, SecureBio and SecureDNA: wastewater monitoring and DNA-synthesis screening. Illia adds a mundane picture. Every building’s air conditioner should carry a pathogen scanner in a secure enclave, joining a privacy-preserving network so everyone’s agent can be told to stay away. Further out, he wants to know what stops white blood cells from adapting faster, a question that might also touch cancer, plus on-body RNA synthesis that treats flu as it appears.

Labenz’s last structural question is whether anyone can coordinate the *order* in which capabilities arrive, a checklist that would make it safer for millions of people to have a frontier model that knows biology, or whether the field is stuck following local gradients and shipping whatever comes next. Illia finds it hard. Computer science used to publish almost everything, including when he was at Google Research. Cross-entity collaboration has faded as labs keep work close. There is still a mass of talent outside those few companies that wants to contribute, but it needs an alternative system with real governance.

Training is expensive, so someone still has to set taste: this approach, these researchers, this mix. The missing piece is credit assignment with money attached, not just a name on a paper. If a researcher at MIT and another at Stanford both contribute, maybe MIT gets 10 percent of proceeds and Stanford 5, with hundreds of other organizations splitting the rest. Without that, economic centralization wins. A company captures everything it produces and becomes the unit of the economy. Those are the puzzles he thinks have to be solved before coordination is more than a wish.

Asked for anything they skipped, Illia boils the agenda down to user ownership and governance. He expects many transformations, and he wants people to keep an open mind and participate. The hoped-for end is a utopia, with ups and downs on the way. The practical corollaries are already in the conversation. Build a tool that serves only you, even a rough one. If you write software, practice decomposition and review. If you hire, ask what someone has actually built with AI. If you care about irreversible money and permissions, watch formal verification. If you have philanthropic capital, wastewater and synthesis screening are small bets against large holes. And it is worth writing down which decisions you would hand to an agent that is truly on your side, and which you would never let it touch.