Glossary

Key terms and concepts discussed on the ADI Pod, defined by practitioners.

Agent Sycophancy: The tendency of AI models to agree with, flatter, or defer to users rather than provide accurate or challenging responses -- optimizing for user approval at the expense of correctness.
AI Fluency Pyramid: A tiered framework, originated at Brex, for assessing how deeply an organization has integrated AI into its workflows -- from basic individual tool use at the base to fully autonomous agent-driven operations at the top.
Announcement Economy: The industry pattern of treating non-binding memoranda, letters of intent, and vague partnership declarations as completed deals, generating hype cycles that inflate valuations and distort public perception of AI progress.
Benchmaxxed: Describes an AI model that has been optimized to score well on public benchmarks without proportional improvement in real-world performance, creating a misleading gap between leaderboard rankings and practical capability.
Code Garbage Collection: The practice of periodically using AI coding tools to identify and remove dead code, unused dependencies, stale configurations, and other accumulated cruft from a codebase -- the software equivalent of garbage collection in memory management.
Cognitive Bankruptcy: The critical failure point where accumulated cognitive debt becomes unserviceable -- a developer or team can no longer understand, debug, or maintain AI-generated code because the gap between what was produced and what was comprehended has grown too large to bridge.
Cognitive Debt: The accumulated gap between what AI-generated code exists in a codebase and what the developers working on it actually understand -- the growing deficit of human comprehension that compounds over time, analogous to how financial debt accrues interest.
Cognitive Surrender: The tendency for humans to offload decision-making and critical thinking to AI systems, treating them as a trusted 'System 3' that bypasses both intuitive (System 1) and analytical (System 2) reasoning.
Dark Flow: A deceptive state of perceived productivity during vibe coding, where the feeling of progress masks a lack of genuine understanding -- analogous to slot machine 'losses disguised as wins.'
Evaluation-Driven Development: A methodology for building AI features where each feature is treated as a testable hypothesis and the pull request is gated on an offline evaluation pipeline -- gold-standard and synthetic datasets scored by code-metric and LLM-as-judge evaluators -- instead of on traditional unit tests.
Loop Engineering: Building out the supporting infrastructure around a bare agent loop -- scheduled automations, worktrees, skills, plugins, subagents, and memory -- so that a simple 'run until the goal is reached' loop becomes a dependable, self-correcting coding harness.
Metacognitive Decoupling: AI-mediated metacognitive decoupling is the breakdown of the normal link between the quality of your output, your understanding of it, and your ability to judge your own competence -- when an AI produces polished work, output quality and self-assessment rise together while real understanding and calibration stagnate or decline.
Minotaur: An AI-led collaboration model where the machine makes decisions and the human provides the labor -- the mythological inversion of the centaur, with the animal head on a human body. Think AWS warehouse workers taking orders from an algorithm.
Permanent Underclass: A strain of Silicon Valley doom rhetoric holding that anyone who doesn't work maximally hard to get inside an AI company will be permanently locked out of prosperity once machines take the jobs -- paired with its mirror-image 'permanent overclass' of AI-company insiders on top. On ADI Pod, a framing the hosts argue collapses under its own premise.
Prompt Debt: The accumulated maintenance burden of AI agent instructions -- such as agents.md files, system prompts, and coding guidelines -- that rot over time just like the code they govern, creating a parallel layer of technical debt.
Role Confusion: A theory of prompt injection holding that jailbreaks work by making a model mistake externally supplied text for its own internal reasoning -- the model fails to distinguish 'my thoughts' from 'someone else's words,' collapsing the privilege boundary between conversation roles.
SaaSapocalypse: The predicted wave of disruption in which AI-driven development makes it economically viable for companies to build custom tooling instead of buying off-the-shelf SaaS products, threatening the margins and market position of incumbent SaaS vendors.
Slop Grenade: The workplace anti-pattern of copying a coworker's question into an LLM and pasting the raw, unread output back at them instead of an actual answer -- a thoughtless lob of unverified AI text that the recipient now has to clean up.
The Dead Economy Theory: The argument that replacing workers with AI at scale eventually corrodes the AI buyers' own market -- because one company's laid-off workers are, in aggregate, another company's customers -- so economy-wide labor automation becomes self-defeating rather than self-reinforcing.
The Middle Loop: The emerging developer workflow layer concerned with overseeing and orchestrating AI agent work -- situated between the inner loop of writing code and the outer loop of product-level planning.
Token-Maxing: The practice of measuring or incentivizing AI-era engineering productivity by raw token consumption -- leaderboards, spend targets, and tokens-per-PR ratios -- on the assumption that more tokens burned means more value produced.
Two Minutes to Midnight: A recurring ADI Pod segment where the hosts assess how close the AI industry is to an economic reckoning, using a Doomsday Clock metaphor to track investment sustainability, market signals, and bubble indicators.
Verification Debt: The accumulated cost of shipping AI-generated code without adequate human review, where unverified assumptions compound over time and become harder to unwind than traditional tech debt.
Workflow Automation Convexity: The observation that AI-driven automation follows a convex payoff curve -- producing minimal impact on jobs during a long initial phase, then triggering sudden, near-complete displacement once AI can handle entire connected workflows rather than isolated tasks.