Unlimited Is a Four-Letter Word: What OpenAI's ChatGPT Bet Teaches Decentralized AI About Trust, Computation, and Governance

Sixty-two percent. That's the number OpenAI wants the world to internalize as proof of safety. In a single announcement, the company claims that switching default models to something called GPT-5.6 Luna, giving free users unlimited text chat, and adding a "Think" button with an adjustable reasoning slider will cut the frequency of at least one factual error per response by 62%. The number is precise. It is also, as far as anyone outside OpenAI can tell, entirely unverifiable. I've spent the better part of a decade designing governance frameworks for DAOs, tokens, and decentralized markets. I've seen multisig contracts fail because everyone trusted the code but nobody trusted the people. I've watched a treasury drain in real time while the community argued about which of the seven signers had the right to veto. So when a company with the most concentrated AI power in history tells me it has reduced errors by roughly two-thirds, I don't ask whether the number is plausible. I ask where the proof lives, who audited it, and what happens when the model changes its mind. Code is law, but people are the soul. And right now, the soul of the most important AI product on Earth is controlled by a single corporate entity. Let's be clear about what the announcement actually contains. It is not a new model release. It is a product strategy reshuffle designed to answer three questions. How do you serve millions of free users without going bankrupt? How do you move from spectacle to daily habit? And how do you turn free users into paying subscribers without making the free tier feel punitive? The answers, as reported, are elegant. Default the product to a lighter, faster model named Luna. Give free users and users of a new "Go" plan unlimited text chat. Add a Think button with a reasoning slider so users can voluntarily spend more compute on harder questions. And polish GPT-5.6 Sol for Plus and Pro subscribers, making its answers more focused, less cluttered, more consistent in tone, and less likely to hallucinate. On the surface, this is a consumer-first move. Under the surface, it is a profound statement about how centralized AI plans to organize access to the most valuable resource on Earth: thought itself. As a governance architect, I look at this announcement and see something different than a PR beat. I see a unilateral parameter change in a protocol that has billions of users. In the DAO world, changing a default is a governance event. You put the proposal on-chain. You let the community review the new routing logic. You publish a transparent audit trail showing why the previous default was replaced and how the new one was tested. OpenAI did none of that. There is no forum, no vote, no verifiable commit hash, no on-chain record of the old model's outputs next to the new model's outputs. The switch to GPT-5.6 Luna was made silently, behind closed doors, and then announced as a gift. That should terrify anyone who thinks decentralizing the digital economy matters. The default model in a system like this is the most powerful governance lever that exists. Defaults shape behavior more than any pop-up warning or terms-of-service agreement. When a user opens ChatGPT and starts typing, they are not choosing Luna. They are accepting the default. They are outsourcing the decision about what constitutes a good answer, a true answer, a useful answer, to a product manager's A/B test. In decentralized networks, we obsess over who gets to set the default. That's why we have token-weighted votes on DAO parameters. That's why we demand governance forums and timelocks. We know that whoever controls the default controls the ecosystem. OpenAI just exercised that control, and the only cost was a blog post. Let's talk about the Think button and its slider, because this is where the design becomes dangerously seductive. On its face, giving users control over reasoning depth is a great UX pattern. The default path is fast and free. If you need more thoughtful analysis, you move the slider and the model "thinks" harder. This feels like agency. It feels like the user is deciding how much computation to buy. But it is a complete illusion. A slider is not a market. It is not a governance mechanism. It is a rate-limiting interface attached to a centrally managed compute budget. The user does not know how many tokens are being spent at each level. They do not know whether "level 5" is ten times more compute than "level 2" or simply a different temperature setting. The slider maps to internal reasoning effort parameters that OpenAI owns and can change at any time. I have audited enough protocols to know that every apparent freedom is hiding a constraint. In DeFi, Aave and Compound let users choose between different interest rate models, but those curves are arbitrary. They are not drawn from order books or real capital markets; they are parameterized formulas chosen by the protocol team. The interest rate slider, if it existed, would feel like a free market. In reality, it is a permissionless illusion within a walled garden. OpenAI's Think slider is exactly the same. It gives users a fake sense of control over reasoning depth while the actual production function — the cost of a single inference, the latency, the batch size, the token budget — remains hidden. This is not a critique of the slider's utility. It is a critique of the opacity. If I can't verify what happens when I move the slider, then the slider is just a rhetorical device. And the "Think" mode introduces a security question that the announcement completely ignores. In the past few years, we have learned that chain-of-thought reasoning can be extracted from models if users ask sufficiently clever prompts. When a user forces a model to "think harder," the model may generate more intermediate reasoning steps. Some of those steps may be hidden from the final answer, but some may leak into the response. This is a direct attack surface for adversarial users. If OpenAI has not added additional guardrails around the Think mode, then the company is handing out a free tool for prompt-injection and systematic extraction of hidden reasoning. It is also handing out a tool for deep fake explanations. A model that thinks longer can produce more convincing wrong answers, more elaborate rationalizations, and more plausible fake news. Lowering the fact-error rate on everyday queries means nothing if the Think mode can be used to manufacture high-confidence misinformation. Now let's look at the word that deserves the most suspicion: unlimited. OpenAI is giving free users unlimited text chat. I want to be honest, because I love the ambition. I also want to be clear that "unlimited" is the single word most abused in consumer technology. The minute a company says unlimited, it has already started building the limits. There are server-side rate limits, time-of-day caps, hidden token ceilings, and fair-use policies that can yank the rug whenever the cost center gets too big. This is not unique to OpenAI. Every cloud service has done it. Every "unlimited" data plan has done it. The word exists not to describe a technical condition but to create an emotional condition: the feeling of abundance that makes a user change their defaults. In my own experience launching a community fund, we told members they had "unlimited" voting rights on the DAO treasury. What we meant was that anyone could submit a proposal, but the multisig had a seven-day delay and a quorum requirement. The word unlimited made the community feel empowered. The code made them feel the quorum. The same dynamic is playing out with ChatGPT. Free users will get "unlimited" chat, but it will be unlimited within a basket of constraints: a maximum number of requests per hour, a maximum context window, a maximum number of Think-button activations, a maximum response length. OpenAI is not lying. It is simply using the word as marketing. The limits will live in the backend, and they will shift based on server load, investor pressure, and competitive threats. The reason "unlimited" is available at all is an infrastructure story. To make a free unlimited tier work, OpenAI must have reduced the unit cost of a chat interaction below the threshold that makes free users bankrupting. That means Luna is almost certainly a distilled, pruned, or quantized version of the larger model. It is not the strongest model. It is the most cost-efficient model for the common cases. This is a brilliant engineering decision. It is also exactly the kind of decision that a decentralized protocol would document transparently. A DAO would say, "We have two models: a large generalist and a small specialist. Here are the benchmark comparisons, the inference costs, and the sampling rules. The router will send short queries to the small model, and long reasoning tasks to the large model." OpenAI does not do this. It just gives the small model a nice name, Luna, and lets its halo effect make it feel like an upgrade. Let's talk about the 62% claim. It is the most emotionally resonant number in the announcement. It suggests that the company has made a meaningful dent in hallucination. But I have spent enough time in crypto to know that when a team reports a large percentage improvement, the first question is: over what benchmark, with what percentage of the test set? A reduction from 0.32 errors per response to 0.12 errors per response is 62.5%. If you sample enough everyday questions, that might be real. But if you sample only easy questions, you are measuring a model's ability to parrot familiarity, not its ability to reason under uncertainty. If you measure the frequency of "at least one factual error per response" on a set of prompts where models are already good, then small absolute gains become huge relative gains. The metric could be perfectly accurate and still misleading. In the decentralized world, we would never accept a claim like this without an independently reproducible test. We would ask for the evaluation harness. We would ask for the prompts. We would ask for the temperature settings and the random seeds. We would ask for the model version and the system prompt. OpenAI will provide none of that publicly. The 62% figure is a black box. It is designed to be believed, not verified. And here is the damage: every time a user sees that number, they are being trained to trust a centralized oracle. They are being taught that truth comes from a vendor. That is the exact opposite of the epistemic foundation that decentralized networks are trying to build. Trust isn't verified on-chain; it is verified by repetition. And the repetition is purchased by a marketing budget. Sol's improvements are also part of this trust-building project. The announcement describes a model that gives more focused answers, reduces unnecessary formatting, stays consistent in tone, and lowers error rates. These are all qualities associated with a polished, predictable, even bureaucratic intelligence. There is no mention of new capabilities, no mention of higher reasoning ceilings, no mention of creative breakthroughs. The direction is toward standardization. In a centralized company, this is called alignment. In a governance context, I would call it the imposition of a single voice. When a model becomes consistent in tone, it stops representing the diverse perspectives of the people who generated its training data. It starts representing the preferences of the alignment team and the legal department. That might be safer for the company, but it is a form of cultural monoculture. The Sol update is not just a quality-of-life improvement. It is an exercise in tightening the boundaries of what can be said, how it can be said, and who controls the aesthetic of acceptable thought. Let me bring this back to decentralized AI, because that is where the governance stakes become clear. There is a growing ecosystem of projects trying to create open, crowd-sourced AI networks. I have been part of discussions where people dream about a future where models are not controlled by one company but by a fluid community of miners, validators, and curators. These projects face enormous technical and economic challenges. Inference verification is hard. Preventing free-riding is hard. Compensating data contributors is hard. But the problem is not only technical. The problem is that the demand for reliability and trust is being captured by a centralized player at exactly the moment when capital-intensive AI infrastructure is becoming impossible for a pure DAO to replicate. OpenAI's unlimited free chat is not a gift. It is a loss leader designed to lock in a habit before a decentralized alternative can mature. Every user who gets used to instant, unlimited, free text chat is a user who will not understand why a decentralized AI network has pay-per-inference fees, latency delays, and variable quality. The production cost that OpenAI hides behind venture capital subsidies is the same cost that a sustainable protocol must pass on to its users. This is a dangerous asymmetry. The centralized company can subsidize its way to user acquisition, crush the competition, and then raise prices later once the defaults have become reinforced. Decentralized networks cannot do that. They do not have a single treasury that can burn billions on user adoption. They have communities that must make trade-offs transparently. There is a second lesson from this announcement for decentralized AI projects. It is about the importance of model routing transparency. Luna and Sol represent a routing decision: some queries go to the lighter model, some go to the heavier model. In a decentralized protocol, routing is a critical governance parameter. Who decides which tasks get routed to which model? How are those decisions audited? What happens when the router is wrong? These questions should be answered in advance, not after a scandal. I have seen DAOs fail because they treated routing as an engineering detail. A governance structure that creates an oracle for model selection is at least as important as the model itself. OpenAI is making that point unintentionally by giving Luna a name and a persona. Luna is not a product. Luna is a governance choice in disguise. The Think slider also teaches us something about the token economics of AI. When a user has a way to dial compute consumption up and down, the marginal cost of their interaction becomes a variable. This is a perfect recipe for a decentralized credit system. Imagine a protocol where users spend tokens not just per request but per unit of reasoning effort. The slider could be tied to a token curve, with high reasoning levels burning more credits and low reasoning levels burning fewer. That kind of design is impossible under OpenAI's centralized pricing because users cannot see the actual compute consumption. But a protocol with verifiable inference could make every slider position a smart contract event. The user would know exactly how many tokens are being allocated to each reasoning step. The model would have to produce a proof of computation, perhaps a ZK-proof, that demonstrates the claimed reasoning depth was actually executed. This is the future that decentralized AI promises: not just open models, but open economics and open computation. I am well aware that ZK proofs for AI are still painfully expensive. I have spent time studying the proving costs on today's zkVM and zkML stacks, and the numbers are not pretty. In a bull market, people forget that gas prices eventually fall and that proving costs remain. But the same was true for ZK rollups a few years ago. The theoretical promise was there, but the operators were bleeding money every block. Over time, specialized hardware and better cryptographic protocols brought the cost down. The same trajectory is possible for verifiable inference. The question is whether decentralized AI will get the sustained funding and engineering attention that rollups did. If the answer is yes, then in a few years the idea of an auditable Think slider will be a commodity. If the answer is no, then OpenAI will continue to own the only sliders that matter. There is also a geopolitical angle that I cannot ignore. The announcement says nothing about where the free unlimited tier will be available. But if OpenAI is rolling this out primarily in North America and Europe, it reinforces the divide between the global AI-rich and AI-poor. A user in a county with stricter regulation might get a less capable model, a more limited Think slider, or no free unlimited tier at all. This is a governance decision as much as a technical one. It mirrors the existing rifts in financial services, where decentralized money is supposed to offer equal access but often fails because of regulatory capture. MiCA is a perfect example: Europe's attempt at crypto regulation gives the appearance of clarity while the compliance costs crush small projects. A similar regulatory dynamic is coming to AI. The largest firms will be able to hire armies of compliance officers. Decentralized AI networks will be left with the choice between following a centralized rulebook and being shut out of the market. The free unlimited announcement may be the first whisper of that future. I want to pause and acknowledge the contrarian case. As much as I believe in decentralization, there is a real argument that OpenAI is doing the responsible thing. The model with lower fact error rates is objectively better for a daily assistant. The Think slider gives users agency over their own task complexity. The Sol improvements make the product more usable. And the free unlimited tier is a genuine attempt to democratize access to a powerful tool. Not every centralized decision is evil. Some of them are pragmatic. Some of them are even humane. If I am in a hospital and my insurance company has an arbitrage opportunity, I do not care whether the routing algorithm is decentralized. I care about whether the model gives me the right answer. The danger is not centralization itself. The danger is opacity at the moment of maximum power. However, the contrarian case does not remove the need for adversarial accountability. It simply changes the question. Instead of asking "Is OpenAI's strategy good or bad?" we should ask "What kind of governance infrastructure would make a system like ChatGPT as trustworthy as it pretends to be?" The answer, I think, is a layer of independent verification. There should be a public registry of model versions, including hash commitments. There should be a community-maintained benchmark harness that runs against the same public model endpoints on a daily basis. There should be a way for users to file "governance challenges" when a model's output changes in an unexpected way, with a public record of the response and the system's stated confidence. None of this has to be adversarial in the sense of trying to destroy OpenAI. It can be cooperative in the sense of building a verification layer that all AI providers, centralized or decentralized, can plug into. The blockchain community has given the world a toolkit for proving that things happened when they happened and that the state remains consistent. The AI community needs exactly that toolkit, but for model behavior. Let me be concrete about what I would do if I were designing a governance prototype for a ChatGPT-like product tomorrow. I would start with a signed log of model versions. Every release would produce a cryptographic fingerprint of the weights, the config, and the system prompt. The default model router would be a separate module with its own versioned rules, so users could audit why a query went to Luna instead of Sol. The Think slider would emit a signed receipt on each activation, recording the requested effort level and the actual compute budget allocated. The fact-error benchmark would be run by an independent DAO using a publicly curated set of prompts, with the results published on-chain. The 62% figure would be one claim among many, always compared against the open benchmark. Finally, there would be a community veto mechanism: if a change to defaults or safety filters degrades performance for a protected class of users, a proposal could be put on-chain to revert to a previous version. This would make trust a process, not a promise. This is where the phrase "decentralization is a verb" becomes real. Decentralization is not a thing you buy. It is a behavior you keep performing. OpenAI's announcement is spectacularly centralized because it treats users as passive receivers of an improved model. The user does not get to participate in the improvement. The user does not get to propose a model version. The user does not get to audit the failure mode. The user just consumes a better answer. The verb "to decentralize" has no object. You cannot decentralize a product. You can only decentralize the decision-making about the product. And that is exactly what OpenAI did not do. Every design choice in this announcement — from Luna as the default, to the slider as the interface, to Sol as the premium polish — reinforces a hierarchy in which the company decides and the user accepts. Do not mistake my criticism for a dismissal of the product value. I use AI tools constantly. I am writing this article with AI assistance in my research process. I believe that every person on Earth should have access to a powerful, low-friction, intelligent assistant. The fact that OpenAI is pushing toward that goal is genuinely good. What I worry about is the path. If we become habituated to a single centralized assistant that defines truth, tone, and reasoning depth behind closed doors, we will lose the instinct to question. That instinct is the foundation of democratic governance. It is also the foundation of science, journalism, and personal autonomy. We cannot build a decentralized economy on top of an epistemic monoculture without repeating every historical tragedy of centralized power. The technology is not the only problem. The default is. The next time you see a shiny new AI announcement, ask yourself who voted for the default. Ask where the proof of the model's safety lives. Ask whether the number could survive independent inspection. Ask what happens when the company decides that "unlimited" is no longer affordable. Ask whether the value of the network accrues to the users or to the shareholders. And then ask what you are going to do about it. Because the answer cannot be just a tweet. It cannot be just a blog post. It has to be a governance primitive. It has to be a mechanism that lets people verify, challenge, and rebuild the rules of the systems they depend on. Code is law, but people are the soul. If we let the code become a black box, we lose the soul too. I don't know exactly what OpenAI will do next. I suspect Luna and Sol will evolve, the Think slider will get more granular, and the unlimited tier will eventually become mildly limited. I also suspect that decentralized AI will not capture the mass market in the next year or two. The capital gap is too large. But I do know that the window for building the governance layer is open right now. Every protocol that wants to provide verifiable inference has a chance to learn from ChatGPT's user experience while avoiding its opacity. The tools exist: cryptographic commitments, public benchmarks, DAO-governed model routers, tokenized reasoning credits, and a growing community of people who care about accountable technology. The question is whether we can move fast enough to make trust a default feature of the next generation of AI. The answer depends on whether we treat decentralization as a verb, not a noun, and start doing the hard work of building a more transparent default. So here is my recommendation to every DAO, every AI researcher, every token holder, and every user who cares about the future of intelligence. Stop treating AI as a utility and start treating it as a public good with governance layers. Ask for proofs. Demand independent audits. Build sliders that are backed by auditable token consumption. Make model routing a community decision. And do not accept a 62% claim from anyone unless you can verify it yourself. We have the cryptography. We have the markets. We have the governance patterns from the DAO experiments that failed and the ones that survived. The only thing missing is the will to apply them. If we can't do that, then we will have traded the promise of decentralized abundance for a perfectly polished, perfectly centralized, perfectly obedient oracle. That is a trade I am not ready to make.