Notes on gotchas while migrating 35kb preprompts from Opus to self-hosted Ollama

TEXT START: Maybe you’re a Claude code/codex user diligently avoiding uploading personal data to LLM providers.

The Dissection

This is a practical migration report wrapped in a broader indictment of frontier AI providers. Its real subject is not merely Ollama. It is the loss of inference sovereignty: large context windows and stronger models make users dependent on centralized providers, while the providers gain custody of the most valuable artifact—the user’s problem-solving history.

The technical core is credible within the supplied evidence: oversized prompts, limited context, repeated tool calls, weak state management, and excessive agentic loops make local systems collapse under workloads designed for frontier APIs. The proposed remedy is disciplined decomposition, explicit state logging, fewer tools, and narrower objectives.

The Core Fallacy

The text conflates control of inference with control of AI capital. Self-hosting can protect data and reduce provider dependence, but it does not restore the mass employment-to-wage-to-consumption circuit. It gives technically capable users a defensive position, not general economic sovereignty.

It also leaps from “providers cannot be audited” to “everything you do will get stolen.” That is a suspicion, not proof. The privacy risk may be serious; the supplied text does not establish systematic theft. The argument’s strongest case is custody risk, not demonstrated expropriation.

A second error is treating context length and chain-of-thought capacity as interchangeable. Large context can improve persistence and search space, but it does not automatically mean the provider’s hidden reasoning is doing the work. The article’s observations are operational, not a general theory of model intelligence.

Hidden Assumptions

• Local hardware can deliver acceptable performance without making electricity, cooling, maintenance, and upgrade costs prohibitive.
• Open-weight models will become capable enough for high-value cybersecurity and coding work.
• Abliteration removes refusals without introducing unacceptable reliability, safety, or alignment failures.
• The operator can secure the local machine, tools, dependencies, logs, and model supply chain better than the provider.
• Personal session histories are both genuinely valuable and likely to be targeted or monetized.
• Frontier providers’ privacy and retention practices will remain worse than the risks of self-hosting.
• The tuning burden—prompt redesign, state management, permissions, monitoring, and debugging—is affordable for more than a technically elite minority.
• Better local inference will create durable economic leverage rather than merely reduce subscription and privacy costs.

Social Function

The text is a partial truth functioning as transition management and sovereignty signaling. It gives technically competent users a map for escaping centralized custody while converting diffuse distrust into a concrete engineering project.

Its darker function is elite self-exoneration: the reader can frame withdrawal from frontier systems as principled resistance, even though the real advantage comes from possessing hardware, expertise, time, and operational discipline unavailable to most people. The article does not solve mass displacement. It teaches a narrow class how to occupy a better altitude during it.

The Verdict

This is not a rescue plan for the economic order. It is a field manual for reducing dependence on one layer of the emerging hierarchy.

Self-hosting is a real moat against provider custody, but for most users it is an expensive hobbyist moat. The frontier provider retains the advantage in model capability, context, infrastructure, and maintenance; the local operator retains control, privacy, and customization. That is a meaningful trade, not a reversal of the system.

Under the Discontinuity Thesis, the article identifies a genuine survival niche: technical operators who own or control compute, understand model behavior, and can turn local inference into specialized productive leverage. Everyone else is still a consumer renting intelligence from a Sovereign. The dragon may be untrustworthy, but moving into a garage with a GPU does not make the renter a king.