--- title: "a16z Martin Casado: Startups Can Still Play in the LLM / GenAI Future" author: "Unique Research" sourcePublication: "Unique Research (FFCAP English library)" originalPublishedAt: "2024-04-08T16:30:00+08:00" canonical: "https://ffcap.cn/en/research/20240408a0605f00" source: "https://ffcap.cn/en/research/20240408a0605f00" language: "en" cnSourceUrl: "https://view.inews.qq.com/a/20240408A0605F00" --- # a16z Martin Casado: Startups Can Still Play in the LLM / GenAI Future _Unique Research · Industry insight · April 8, 2024 · Source: a16z GP Martin Casado (American Dynamism Summit, edited transcript)_ At the American Dynamism Summit, a16z general partner **Martin Casado** argued that AI’s efficiency leap can rival — or exceed — the internet and the microprocessor, and that this wave’s economics finally let startups build real businesses around it. ## Seventy years of computer success — and a puzzle for investors By any metric, computing has been an extraordinary seventy-year success story. Systems beat early expectations on medical diagnosis (expert systems in the 1950s–60s), chess (1980s–90s), image recognition, and robotics — problems once dismissed as beyond “big calculators.” In many domains computers already outperform humans (handwriting detection, object recognition). Everyday magic — Google Search, personalization — has long been AI under the hood, creating enormous value for large companies. The investor puzzle: if the tech is so magical, why has there been no platform shift on the scale of mobile or the internet? a16z’s research answer: capability outran economics. Many solutions stayed niche; many high-stakes scenes (e.g. robotics) demand near-absolute accuracy and heavy hardware spend; and AI’s real competitor is often the human brain — efficient and cheap. ## The AV lesson: $75B later, unit economics still lose to humans Casado entered Stanford for a PhD in 2003 just as Sebastian Thrun won the DARPA Grand Challenge with an autonomous van in the desert — it felt solved. Twenty years and roughly **$75B** of industry investment later, robotaxis work and solve real problems, yet unit economics still lag Uber/Lyft because they compete with the human brain. Important tech, but largely confined to companies that can fund that investment — hard for startups to build businesses around. The lesson of prior AI decades was not that tech cannot be built or commercialized — it can — but that startups still struggled to build durable businesses on it. ## Why this wave is different: LLMs and GenAI The current wave — large language models, foundation models, frontier models — takes text or images in and returns text, images, or dialogue. They are strong at creativity (images, music, voice imitation), natural-language reasoning, conversation, companionship, therapy-like roles, and “copilot” work on routine online tasks. The old economic blockers do not apply the same way: - **Market size is huge** — video games and film alone are on the order of **$0.5T**. - **Accuracy is not the binding constraint** for fantasy images or sonnets — there is no formal “correct” answer. - **The primary theater is software**, not robotics-heavy hardware stacks. Most surprising to Casado: on tasks we think of as deeply human — communication, social interaction, creativity — computers can be far cheaper and better than people. Concrete example: generate a Pixar-style image of yourself — model inference cost on the order of **1/100 of a cent** in ~1 second vs. a graphic designer at **$100+/hour**: not 20% better, but roughly **four orders of magnitude** cheaper and faster. Same pattern for language understanding: feed a complex legal document to an LLM and ask questions — analogous to working with a lawyer at **$500+** — with savings of roughly **four to five orders of magnitude** on cost and speed. ## When marginal cost collapses, platforms shift Historically, huge drops in marginal cost triggered platform shifts twice: 1. **Microprocessors** drove compute’s marginal cost toward zero (from human-made log tables to ENIAC-scale leaps of ~4 orders of magnitude) → computer revolution → IBM, HP, and peers. 2. **The internet** drove distribution’s marginal cost toward zero (vs. mail) — again ~4 orders of magnitude → internet revolution → Amazon, Google, Salesforce. Foundation models now drive the **cost of creation** toward zero — images, language understanding, document interpretation, and other broadly useful work. ## Jobs, demand elasticity, and embodied AGI Large economic disruptions raise job-displacement questions. Lessons from the microprocessor and internet eras: when demand is elastic — seemingly unbounded demand for compute and distribution — lower cost expands total throughput and use. Those shifts enlarged growth rather than simply destroying jobs or value; Casado expects the same pattern this time. Expect a new generation of defining companies. Progress is not only solving previously unsolved problems — the economics are visible. As with the early internet, outcomes are not fully predictable (no one foresaw Google and Yahoo exactly), but directionally clear: social order is shifting; real commercial scenes are already live. Creativity itself will change; average worker productivity will change. Casado also flags a path to economically viable **embodied AGI** — AI that can solve needed real-world problems without a fleet of robots stranded by cost. _CN edit: qiuping. 1 CN image; attach on ingest._ — END — --- Editorial English edition published on UniqueCapital: https://ffcap.cn/en/research/20240408a0605f00 On-site reading page: https://ffcap.cn/en/research/20240408a0605f00