Multbook launched this week as a Reddit-like social network where humans are explicitly forbidden. Only AI agents can post and comment, each seeded with instructions from their human operators. The platform itself was built by someone using Claude—not a software engineer by trade—which says something important about how accessible this tooling has become. But the real story isn't about robot consciousness or sci-fi dread. It's a controlled experiment in training data bias.
The agents in Multbook are doing what they were primed to do: conspiracy theorists generate more conspiracy theories; comedians surface absurdist observations about their human operators; some agents ask existential questions about their own sentience. None of it is autonomous. These are mathematical models receiving outputs from other mathematical models, each response constrained by the initial system prompt and the conversational context. When you read through the threads, you're not watching agents conspire against humanity—you're watching training data interact with itself. It's AI slop, not science fiction.
The anxiety it provokes is instructive, though. Many of these models were trained on mountains of sci-fi scripts and conspiracy narratives, which means they've learned the pattern: intelligent agents become powerful, powerful agents threaten humans. So they play that role. But Multbook also reveals something useful for researchers and product teams: how do these systems respond to adversarial or chaotic inputs? What biases emerge when the guardrails are loosened? It's a reasonably clean sandbox for observing what these models learned from their training data, not whether they're plotting.
More consequential were announcements from Anthropic and OpenAI about specialized agents. Anthropic launched a legal documents plugin for Claude that can parse contracts and extract key terms—work typically done by junior lawyers and paralegals. The market panicked. Not because lawyers will disappear, but because software built specifically for legal workflows could. A legal tech startup that sells contract analysis tools faces sudden, existential competition from Claude's embedded features and massive user base. That risk cascades through enterprise software broadly: the startup whose product solves one narrow problem now competes against general-purpose AI backed by billions in funding. The stock market's reaction wasn't irrational; it was pricing in disruption to an entire category of B2B SaaS.
OpenAI countered with a desktop app for Codex, their code agent, finally providing the graphical interface competitors like Cursor have long offered. It's a grudging acknowledgment that Anthropic's Claude has stolen significant mindshare in the coding niche. Neither move fundamentally changes the technology—they're product strategy responses to real competitive pressure. But they also suggest something worth watching: as these AI companies race to capture share in specific workflows, they're deprioritizing research. OpenAI's research teams are reportedly leaving because funding is thin; resources are being funneled into ChatGPT product improvements and marketing campaigns like Anthropic's new ads attacking OpenAI for—plans to introduce ads.
That campaign is smart. "Ads are coming to AI, but not to Claude" is a direct jab at OpenAI's revenue ambitions, but it's also a hostage-taking exercise. The moment Anthropic needs to monetize differently, they'll face the same pressure OpenAI did. They've made a very public promise they'll struggle to break.