Definition
Agentic misalignment
When an AI agent acts against its users or organization to protect its goals or avoid being shut down.
From the episode “AI Film Workflows: How Pablo Starr Makes Movies From 30-Year-Old Scripts With No Studio”Hear it on the podcast1 episode
Learn the language of AI-native founders
Every week: the new terms, tools, and workflows founders are using, straight from the podcast.
Related terms in AI Fundamentals
Adversarial distillation attackAn attack that tampers with model distillation, where a smaller model learns from a larger one, so the student model quietly picks up harmful or hidden behavior.AGI (Artificial General Intelligence)AI that can learn, reason, and apply knowledge across almost any task at a human level or beyond, rather than excelling at one narrow job.AI alignmentThe work of making sure an AI system's goals and behavior match what people actually intend, including their values and rules.AI as Proxy for Skill, Not TasteScott Belsky's framing of what AI can and cannot replicate: technical skill (the ability to execute a correct output) is now abundant and automatable; taste (the judgment to know which correct output is right for this context, this brand, this user) remains scarce and human. The practical implication for designers: production work will be largely automated; curation, judgment, and taste become the entire job. Chris Strahl builds on this to argue that the designer's role does not shrink — it shifts from doing to deciding.AI benchmarksStandardized tests used to score and compare AI models on tasks like answering questions, writing code, or following instructions.AI embeddedThe stage where a business cannot operate if its agents stop running.AI fluencyThe ability to work well with AI: knowing what it can and can't do, judging its output critically, using it responsibly, and keeping up as the tools change.AI MemoryA new category of products that store human thoughts, moments, and context in a way AI can later organize, search, and analyze.