Something quietly uncanny happened this week. While Mirage pulled in $75 million from General Catalyst to keep building the AI models behind its video editing app Captions, researchers at arXiv dropped a paper asking a more unsettling question: do these systems even know what they are? The paper, "Me, Myself, and Pi: Evaluating and Explaining LLM Introspection" (2026, arXiv CS.AI), found that LLM self-assessment is patchy at best, confidently wrong at worst. Capital flows toward capability. Philosophy asks if capability is coherent.
AI Video Editing Funding and the Capability Race
Mirage's raise is not a fluke. It is a bet that AI-native video creation tools will eat the post-production stack, frame by frame. General Catalyst's Customer Value Fund structure is itself a signal: this is growth-stage money tied to revenue metrics, not vibes. Meanwhile, Snapchat launched AI Clips, a Lens format that turns a single photo prompt into a five-second video, democratizing what Mirage is monetizing upstream. The platform layer and the tooling layer are racing each other. If you are building in this space, the AI seed investor landscape has never been more crowded or more specific about what it wants: multimodal, real-time, distribution-native.
What LLM Self-Knowledge Actually Means for AI Creative Tools
Here is the rub. Anthropic's Claude Code and Cowork update now lets the model autonomously perform tasks on your computer, asking permission as it goes. That permission structure assumes the model has an accurate internal model of what it is about to do. The introspection paper suggests that assumption is shaky. A 2026 paper in arXiv CS.AI by the ProMAS team on proactive error forecasting in multi-agent systems found that Markov transition dynamics can predict failure states before they happen, which is a fancy way of saying the system's future errors are legible if you know where to look. The question is whether the system itself is doing that looking. For AI video tools, where a bad edit is annoying, this is fine. For agentic computer control, it is a different stakes conversation entirely.