# Private AGI, Public Scraps: Who Really Gets the Future? Canonical URL: https://themodernhermit.blog/private-agi-public-scraps-who-really-gets-the-future/ Published: 2026-11-07T22:47:13+00:00 Updated: 2026-11-07T22:55:03+00:00 ## Lede The AI industry could eventually build intelligence capable of changing the world, keep its greatest advantages behind closed doors and continue charging everyone else for a restricted imitation. ## Words used - **AGI:** Artificial general intelligence, broadly understood as AI capable of performing a wide range of intellectual tasks at or beyond human ability. There is no universally agreed test for when it has been achieved. - **Frontier model:** An AI model among the most capable available at a particular time. - **Internal model:** An AI system developed or used within a company that may not be available to outside customers. ## Hermit Off Script My suspicion about frontier AI is becoming increasingly uncomfortable. I don't think the most powerful models will stop improving, or that the companies building them will eventually discover some great technological limit and give up. Quite the opposite. I suspect the real race will increasingly happen behind closed doors, where companies develop much more capable systems for their own use while the rest of us receive whatever restricted versions they decide to release. And here's my prediction: they could eventually reach AGI internally without ever giving ordinary customers access to anything remotely as useful. Imagine a company possessing an intelligence capable of developing technologies, improving its own systems, solving difficult scientific problems and accelerating research far beyond anything its customers can accomplish. Why would it simply hand that advantage to everyone for a monthly subscription? It could use that intelligence to develop products, run operations, [sell expensive services](https://themodernhermit.blog/ai-pro-plans-humanity-sales-slogan/), improve the next generation of models and strengthen its position against every competitor. Meanwhile, I could still be sitting at home with a supposedly intelligent assistant that needs instructions repeated, produces unreliable results and has to be supervised to complete a serious professional job. I've already complained about the model circus and the [endless repairs](https://themodernhermit.blog/smarter-ai-same-repairs-another-model-upgrade/). This goes much further. What worries me is the possibility that the public versions won't merely lag behind by a few months. They could become permanently [restricted products](https://themodernhermit.blog/ai-limits-reset-not-mistakes/), carefully designed to be useful enough to keep subscriptions coming but never powerful enough to provide the independence or economic advantage available to their owners. And how would I know? I can't inspect the company's most advanced internal systems. I can only judge what I'm permitted to use, what it costs and how often it actually works. Maybe restrictions will be justified by safety. Some certainly will be. But I can imagine safety becoming a very convenient explanation for an arrangement that also protects an extraordinary commercial advantage. The customers provide money, usage, feedback and, where their settings permit it, information that may help improve future models. The companies collect revenue from subscriptions, APIs, enterprise contracts and the enormous businesses built around AI infrastructure. And if their internal systems become powerful enough, they might not even need to sell their best intelligence directly. They could sell the discoveries, services and products it creates instead. I cannot prove that this is where the industry will end up, and I hope I'm wrong. Genuine competition, open models and independent scrutiny might prevent it. But the possibility deserves more attention than another announcement about how close we are to superintelligence. What happens if the companies actually reach it and decide that the greatest benefits should remain their property? They get the intelligence. We get the monthly bill. ## What does not make sense - **The frontier and the product are treated as the same thing.** A company's most advanced internal capabilities and the system available to an ordinary subscriber can differ considerably. Public model releases do not reveal everything being developed internally. - **The most useful intelligence could become a competitive secret.** A system capable of accelerating research and creating valuable technologies might be worth more to its owner than the subscription revenue from releasing it widely. - **Safety and commercial advantage can point in the same direction.** Restrictions may be justified by genuine concerns about misuse. They can also concentrate valuable capabilities among organisations already powerful enough to develop or obtain them. - **The public may celebrate a breakthrough without receiving it.** Announcing a powerful model does not necessarily mean making its capabilities generally available. - **Access can be selective without being described as a monopoly.** Selected research institutions, partners and organisations may receive models that other customers cannot use, even when access restrictions have legitimate reasons. - **The economics could reward keeping the best models private.** A company might earn more by using an exceptionally capable system to develop commercial products than by allowing competitors and ordinary customers to use the same system. - **AGI could become an internal business advantage before a public service.** If general intelligence is achieved, there is no economic law requiring its developer to release the full capability to everyone. - **A public subscription is not a window into private development.** Without meaningful independent assessment, customers cannot reliably establish how their available models compare with the strongest systems running internally. - **The promise of widely shared benefits has no automatic enforcement mechanism.** Scientific progress, market competition, safety policy and the distribution of economic benefits are separate questions. ## Sense check / The numbers 1. **Anthropic disclosed an unreleased internal model in August 2026.** Its risk report described "Model 2" as somewhat more capable overall than Claude Mythos 5, with noticeable improvements on several tasks relevant to internal use. Anthropic stated that it had no plans at the time to release the model externally. The report also acknowledged that Model 2 was weaker in some areas and had undergone less extensive assessment. This is documented evidence of a stronger internal system being withheld, not proof of secret AGI. \[Anthropic Risk Report\] 1. **More than 40 additional organisations received restricted access to Mythos Preview.** In April 2026, Anthropic announced Project Glasswing, providing selected partners and more than 40 additional organisations access to a powerful unreleased model for defensive cybersecurity work. Anthropic committed up to $100 million in usage credits and $4 million in donations to open-source security organisations. The model was not generally released, although Anthropic stated an eventual goal of making similarly capable models safely available at scale. \[Anthropic\] 1. **Four leading AI developers were independently assessed.** METR examined internal models from Anthropic, Google, Meta and OpenAI during February and March 2026. Its assessment estimated that the internal frontier was approximately 66 days ahead of the public frontier on one capability measure, but it found no evidence of a dramatically more capable hidden model. The estimate was uncertain and based on limited testing. This challenges the idea that a vast secret AGI advantage has already been demonstrated. \[METR\] 1. **OpenAI's first model to reach its Critical cybersecurity threshold received restricted capability access.** In September 2026, OpenAI identified GPT-6 Astra as reaching that threshold and announced that its most advanced cybersecurity capabilities would initially be available to selected testers. This shows that public access to some powerful capabilities can be deliberately restricted, while leaving open the broader question of how future general intelligence would be distributed. \[OpenAI\] 1. **A reported $1.25 trillion valuation.** On 6 November 2026, Reuters reported that OpenAI had secured $11 billion in new funding at a valuation of approximately $1.25 trillion, citing a source familiar with the matter. The valuation represents investor expectations, not profit or cash earned. It nevertheless illustrates how much financial value investors attach to control of advanced AI capabilities. \[Reuters\] --- ## The sketch Scene 1: **The private engine** *Inside a corporate research facility, engineers operate an enormous glowing AI machine marked INTERNAL. It produces diagrams, inventions and new machine designs. A smaller machine marked PUBLIC stands beside a delivery conveyor.* Dialogue: Engineer: "Another breakthrough." Executive: "Keep it inside." Customer: "What ships out?" Scene 2: **The public edition** *The same customer stands at a payment counter beside the small public machine. Through a glass partition, the enormous internal machine produces valuable designs for the company.* Dialogue: Customer: "Is this your best AI?" Clerk: "It's the public version." Customer: "That wasn't my question." Scene 3: **The great AGI announcement** *The customer watches an executive celebrate beneath a banner reading AGI ACHIEVED. A locked corporate gate separates the executive from the small public robot and its subscription terminal.* Dialogue: Customer: "Can we use AGI?" Executive: "We're using it already." Robot: "Subscription renewed." --- ![Three comic panels show engineers using a large internal AI machine, a customer receiving a smaller public model and an executive celebrating behind a locked gate.](https://themodernhermit.blog/wp-content/uploads/2026/11/themodernhermit-private-agi-public-scraps-who-really-gets-the-future-sketch-1300x546-1-1024x430.jpg) First the private breakthrough, then the public subscription, then the announcement nobody can access. --- ## What to watch, not the show - **The internal-public capability gap.** Independent evaluations should compare the strongest models used inside AI companies with those available to customers, including how that difference changes over time. - **Internal AI research acceleration.** Watch whether companies begin using private AI systems to accelerate model training, scientific research, engineering and product development faster than external users can. - **Models deliberately withheld from public release.** Examine how frequently developers acknowledge internal-only models, which capabilities those models possess and why access remains restricted. - **The difference between safety and commercial secrecy.** Restrictions protecting against serious misuse can be justified. They should still be subject to credible oversight, particularly when they confer substantial competitive advantages. - **Who receives the economic benefits.** Watch whether the most valuable AI-generated discoveries and services remain concentrated among model developers, cloud providers, investors and a small group of selected customers. - **AGI claims and independent verification.** A company declaring AGI should not be sufficient evidence that it exists. Capability claims need independent testing and clear definitions, particularly when the relevant models remain inaccessible. - **The role of open-weight models and independent research.** Competition from models that can be operated outside the largest AI companies could limit the ability of any single provider to control access indefinitely. - **The business model after AGI.** If internal AI systems can produce valuable inventions, software and services, the largest profits might come from selling their outputs rather than access to the underlying intelligence. - **Public accountability.** Governments and independent researchers need enough information to distinguish genuine safety restrictions from unnecessary barriers that concentrate economic power. ## The Hermit take The most dangerous AI monopoly would own the intelligence that changes the world while selling everyone else a restricted assistant. Humanity could celebrate the arrival of AGI without ever being allowed through the door. ## Keep or toss ***Toss*.** Keep AI research, genuine safety protections, independent evaluation and fair competition. Toss the possibility of permanent private control over general intelligence while ordinary customers are sold a carefully limited version of the future. ***Disclaimer**: This article combines verified reporting, personal speculation and satire. The existence of restricted internal models and selective access is documented. There is no verified evidence that any company has secretly achieved AGI or intends permanently to withhold it from the public. The three comic scenes depict a hypothetical future, not actual events.* #### Sources - Anthropic, August 2026 Risk Report, internal Model 2: [https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf](https://www-cdn.anthropic.com/f61d49fa5596956a5dec75fea0e973bf6a6a8378/Redacted%20Risk%20Report%20August%202026%20.pdf) - Anthropic, Project Glasswing and Mythos Preview access: [https://www.anthropic.com/glasswing](https://www.anthropic.com/glasswing) - METR, Frontier Risk Report, May 2026: [https://evals.alignment.org/blog/2026-05-19-frontier-risk-report/](https://evals.alignment.org/blog/2026-05-19-frontier-risk-report/) - OpenAI, Path to Astra, September 2026: [https://openai.com/index/path-to-astra/](https://openai.com/index/path-to-astra/) - Reuters, OpenAI funding and valuation, 6 November 2026: [https://www.reuters.com/business/openai-secures-11-billion-funding-125-trillion-valuation-bloomberg-news-reports-2026-11-06/](https://www.reuters.com/business/openai-secures-11-billion-funding-125-trillion-valuation-bloomberg-news-reports-2026-11-06/)