Your startup has to be a surveillance state to automate jobs (manidoraisamy.com)

At OpenAI’s DevDay Sam Altman predicted “zero-person” startups, but a recent support incident shows why that’s farther off than the hype suggests. A user trying to fine-tune an OpenAI model to reliably invoke MCP (model-control/plugin) calls hit inconsistent answers from the support agent; an OpenAI engineer later confirmed finetuning MCP calls isn’t supported yet and recommended using tool calls instead. The exchange highlights not just hallucination but a deeper operational gap: LLMs are often fed static knowledge and can’t account for real-time bugs or process changes without engineered pipelines. The core technical tension is about two kinds of context: historic (all past docs, bug trackers and knowledge bases ingested via RAG or MCP-style connectors) and real-time (ongoing Slack threads, recent fixes, org decisions). Large incumbents struggle to integrate decades of legacy data—creating demand for integration work—while fast-moving startups struggle to capture ephemeral, real-time context unless they resort to exhaustive (and ethically fraught) internal surveillance. The one-person founder is the edge case where context consolidation is trivial. For the AI/ML community this means automation won’t simply replace human roles; it will reshuffle them into integration, context-engineering, and governance tasks—and force founders to choose between team growth, privacy trade-offs, or staying solo to keep AI systems reliable.