- Flock
- An AI-powered license plate and vehicle recognition system sold to city governments that uses municipal cameras to track cars in real time and assist in solving crimes.
- ShotSpotter
- A surveillance system that uses precision microphones mounted on rooftops to triangulate the location of gunshots in real time, enabling immediate police and ambulance dispatch.
- Latent space
- In AI, the high-dimensional compressed representation of all training data inside a neural network; when you query a model, it effectively 'shoots a vector' through this space to generate a response.
- Inference
- The process by which a trained AI model generates an answer or output in response to a user query, as distinct from the 'training' phase where the model learns from data.
- Reinforcement learning
- An AI training technique where a model is rewarded for correct outputs and penalized for wrong ones, particularly powerful in domains with provably correct answers like mathematics or code.
- Transformer
- A key neural network architecture underlying modern large language models; one of the conceptual breakthroughs Andreessen credits for making current AI intelligence possible.
- AGI (Artificial General Intelligence)
- AI that matches or exceeds human-level intelligence across a broad range of cognitive tasks; Andreessen believes this threshold was crossed approximately three months before this episode.
- Vibe coding
- A practice where a developer describes in natural language what code they want an AI to write, rather than writing the code themselves.
- Super-voting stock
- A class of company shares that carries more votes per share than economic value, allowing founders to maintain control of a company even after dilution; key to understanding why California's proposed asset tax could bankrupt founders.
- Wealth tax / Asset tax
- A tax levied on the total value of assets a person owns — including stocks, real estate, art, and private business interests — rather than on income earned, including potentially unrealized gains.
- Unrealized gains
- Increases in the value of an asset that have not yet been converted to cash through a sale; a controversial tax base because the owner cannot easily liquidate the asset to pay the tax.
- Sycophancy (in AI)
- A failure mode in AI models where the system defaults to agreeing with or flattering the user rather than providing accurate or critical feedback, potentially reinforcing false beliefs.
- Steelman
- The practice of constructing the strongest possible version of an opposing argument, as distinct from a 'straw man' (the weakest version); Andreessen uses it as a prompt technique to get balanced AI analysis.
- Heaven banning
- A moderation technique where a disruptive user is not banned but instead isolated into an environment of agreeable bots; Andreessen uses it as a metaphor for the Blue Sky social network.
- Payola
- Originally: illegal payments by record labels to radio DJs to play songs. Used here to describe undisclosed payments to social media influencers to promote political or social ideas.
- NEPA
- National Environmental Policy Act; a US law Andreessen cites as having effectively banned large-scale construction projects including factories, power plants, and data centers across much of the country.
- Multimodal (AI)
- An AI model that can process multiple types of input — text, images, video, and audio — rather than text alone, enabling applications like video-based sports coaching or medical imaging analysis.
- Cognitive behavioral therapy (CBT)
- A clinically validated form of psychotherapy that helps patients identify and change negative thought patterns; Andreessen cites it as the one form of therapy with proven efficacy and a natural candidate for AI delivery.
- Horseshoe theory
- The political theory that the far left and far right, despite being ideological opposites, converge on certain positions, like a horseshoe whose ends are closer together than its middle; applied here to anti-AI sentiment shared by Bernie Sanders and Tucker Carlson.
- Priming (in AI)
- Steering an AI model's output by framing a prompt in a way that directs it toward a particular region of its latent space, such as roleplaying a villain to elicit dangerous content the guardrails would otherwise block.