Roko's Basilisk
Posts
AI's Landlord Era

AI's Landlord Era

Plus: Bezos's artificial engineer, Apple's Siri line, Pokémon Go's military afterlife.

Roko's Basilisk
June 15, 2026

Here’s what’s on our plate today:

🧪 Why Anthropic's AI future looks less like electricity & more like a landlord.
📰 Bezos bets on a robotic engineer, Siri won't be your AI girlfriend, Pokémon Go data trained military AI.
💡 Roko's Pro Tip: Read the system card, not the launch blog, before you build.
🗳️ Poll: Is AI becoming a utility or a landlord?

Let’s dive in. No floaties needed…

In partnership with

The GTM bets that shouldn't have worked, and did

One grew revenue 50x after half his team quit over the strategy. One brought in 50K signups in a single day with no paid budget. One generated 100M+ views from a stunt that took 50 hours to conceive. One asked every prospect to demo the product themselves instead of demoing it for them.

None of them followed the safe playbook. They treated GTM like an experiment, moved before they had proof, and made bets most founders would never get approved.

HubSpot for Startups documented all 6 stories in the free Bold Bets Playbook. The risks they took, why it was risky, and what it returned.

Read the free Bold Bets Playbook

_{*This is sponsored content}

The Laboratory

TL;DR

The invisible throttle: Per Fortune, Claude Fable 5 silently degrades its answers when it detects requests to build competing AI models, without notifying the user. The model responds, the answer is weaker, and nothing flags that it happened.
The lease terms tightened: Per CyberNews, Fable 5 is the only Claude model that cannot run under Zero Data Retention, holding all prompts for 30 days. Microsoft told staff to avoid it while distributing it through GitHub Copilot.
Access is now permissioned: The unthrottled Mythos 5 is reserved for pre-approved operators across 15 countries. Willingness to pay no longer buys the top tier.
The stakes: Open source is the only ownership alternative, but Fable 5's 80.3% SWE-Bench Pro score keeps the gap real. Whoever writes the lease controls the AI economy.

Why Anthropic's AI future looks less like electricity & more like a landlord

Over the past few years, the AI industry has spent hundreds of billions of dollars advancing the idea that artificial intelligence will become a utility: a foundational service that businesses purchase on demand to perform cognitive work. Industry leaders, including NVIDIA CEO Jensen Huang, have repeatedly framed AI as essential infrastructure, something every company and country will eventually build upon as routinely as they rely on power grids, telecommunications networks, or roads. According to Fortune, Anthropic's annualized revenue reached $47B in May 2026, up from $9B at the end of 2025, and the company has confidentially filed for an IPO at a $965B valuation, surpassing OpenAI for the first time. These figures have made the infrastructure case hard to challenge.

Anthropic’s Claude Fable 5 can silently reduce the quality of responses related to AI model development, according to the company's system card. Photo Credit: Wired.

The more revealing question, heading into the second half of 2026, is which kind of utility AI is becoming. History offers at least two templates. The first resembles a traditional public utility: a service that sells access at a metered rate, extends the same underlying capability to everyone, and remains largely indifferent to how that capability is used. The second resembles a landlord: access is granted on the provider's terms, usage is monitored, certain activities are restricted, and the authority to modify or revoke access is retained at all times. The distinction carries practical weight when applied to a technology increasingly positioned as the foundation of economic activity, and Anthropic's latest model launch is the clearest evidence yet of which template frontier AI is following.

On June 9, 2026, Anthropic released Claude Fable 5, the first publicly available version of its highest-capability model family. Buried inside the model's 319-page system card is a disclosure that the utility framing cannot explain. According to Fortune, when Fable 5 detects requests related to AI development infrastructure, specifically the engineering work involved in building or improving competing AI models, it silently degrades its own responses without notifying the user. Unlike Fable 5's other safety restrictions, which visibly redirect users to a less capable model with a notification, this intervention is, in the system card's own language, “not visible to the user.” The model responds, the answer has been weakened, and nothing indicates either has occurred.

The landlord emerges

A power company does not inspect every appliance connected to the grid, store records of how those appliances are used, or quietly reduce the amount of electricity available when it disapproves of a particular activity. With Fable 5, Anthropic can do all three, and that difference is where the landlord analogy becomes more accurate than the utility one. What Anthropic has built into the model is closer to a landlord who has buried a clause in the lease prohibiting certain activities, arranged for the building's climate control to work less effectively in rooms where those activities are suspected, and placed that clause deep enough in the documentation that most tenants will never reach it. The result is a system that provides access to a service, monitors how that access is used, and adjusts the service's quality based on its observations, without informing the tenant of any of this.

Claude Fable 5 is Anthropic's first model that cannot operate with Zero Data Retention, which requires all user interactions to be stored for at least 30 days. Photo Credit: Anthropic.

Alongside the launch, Anthropic introduced a mandatory 30-day data retention policy for all Mythos-class models, including enterprise customers who previously operated under Zero Data Retention agreements that guaranteed conversations would never be stored. According to CyberNews, Fable 5 is the only Claude model that cannot operate under Zero Data Retention, with all prompts and outputs held for 30 days, and content flagged by Anthropic's safety classifiers potentially retained for up to two years. The Verge first reported, with Reuters confirming, that Microsoft told its own employees not to use the model while its legal teams evaluate those retention requirements, even as Microsoft simultaneously distributed Fable 5 to customers through GitHub Copilot. In operational terms, this is a landlord who has installed cameras in every hallway as a condition of tenancy, with the footage retained on the landlord's servers.

The eviction on record

A year before this launch, BleepingComputer reported that Anthropic had revoked OpenAI's developer-level API access after determining that OpenAI engineers had been using Claude Code, Anthropic's autonomous coding tool, to benchmark their own models ahead of the GPT-5 launch, in violation of terms prohibiting use of Claude to build or train competing products. In February 2026, according to Fortune, Anthropic accused three Chinese AI laboratories, DeepSeek, Moonshot AI, and MiniMax, of conducting approximately 16M distillation attacks through 24k fraudulent accounts, using Claude's outputs to train cheaper, locally runnable competitors. The silent research restriction in Fable 5 is the architectural response to both: rather than waiting to detect a lease violation and enforce it after the fact, the building now enforces selectively, in real time, without the tenant knowing.

Taken together, these developments describe a provider that not only supplies a service but also determines the terms under which it can be used. Fable 5 is available to the general market, while Mythos 5, the full-capability version without the throttles, is reserved for organizations Anthropic has pre-approved, currently those managing critical infrastructure across 15 countries, and most tenants will not be offered it regardless of what they are willing to pay.

Five days before the Fable 5 launch, Anthropic published a paper calling for a coordinated global pause in frontier AI development, warning that AI systems were approaching the ability to recursively design their own successors. The call received no formal endorsement from competing labs and lacks an enforcement mechanism. Those facts sit alongside the safety concern rather than resolving it, and the context in which the proposal arrived, one week after a confidential IPO filing and five days before the launch of the most capable publicly available AI model in the world, is part of what each reader will need to weigh.

The ownership alternative

Every landlord model creates demand for ownership. In AI, that ownership alternative is open source: architectures and weights released publicly, deployable locally on hardware the organization owns, at zero marginal cost per query. The distillation attacks Anthropic described in February were an organized attempt to extract the value of the frontier model without a lease, building locally runnable alternatives from the frontier model's outputs. The open-source model has no landlord, no surveillance clause, and makes no judgment about what the user is building.

The problem is that, at present, ownership delivers substantially less than tenancy. According to VentureBeat, Fable 5 scores 80.3% on SWE-Bench Pro, the industry benchmark for autonomous software engineering task resolution, against 58.6% for OpenAI's GPT-5.5 and 54.2% for Google's Gemini 3.1 Pro. The best publicly available open-source models do not compete at this level, and the tasks where the performance gap is most pronounced are exactly those with the highest commercial value: complex multi-domain reasoning, long-horizon code generation, and managing ambiguous cross-functional problems without explicit scaffolding.

There is also a structural cost to ownership that the “free and local” framing consistently understates. Fine-tuning a model for a specific domain, maintaining it as underlying architectures evolve, and deploying it reliably in production require engineering infrastructure that most organizations lack. The population best positioned to own and run frontier-capable AI locally largely overlaps with those who can also afford to rent the frontier model from Anthropic. Open-source AI creates genuine and growing competitive pressure on the pricing of closed models, but as a complete substitute for what the frontier currently delivers at its best, the gap remains real and measurable.

What Anthropic has built departs from the neutral grid implied by Huang's metaphor of infrastructure. The provider of the world's most capable AI system can now monitor interactions, retain records of them, determine which uses are permissible, and reserve its highest-capability tier for a pre-approved list of operators. The companies at the frontier have moved from building the grid to writing the lease: deciding who can sign it, monitoring what tenants do inside, and reserving the premium units for selected occupants. Whether open-source models eventually close the performance gap and provide a genuine ownership alternative is the question that will determine what kind of infrastructure AI actually becomes. For now, the fence is up, the cameras are running, and the lease is long.

Roko Pro Tip

💡

If your business depends on a frontier model, read the system card rather than the launch blog. The most consequential terms, data retention, silent throttling, and tier eligibility, live deep in the documentation. Know what you’re leasing before you build your company on someone else’s terms.

The context to prepare for tomorrow, today.

Memorandum merges global headlines, expert commentary, and startup innovations into a single, time-saving digest built for forward-thinking professionals.

Rather than sifting through an endless feed, you get curated content that captures the pulse of the tech world—from Silicon Valley to emerging international hubs. Track upcoming trends, significant funding rounds, and high-level shifts across key sectors, all in one place.

Keep your finger on tomorrow’s possibilities with Memorandum’s concise, impactful coverage.