E9: The AI Assistant Revolution with Flo Crivello of Lindy.AI

Episode Summary

Executive Summary: This podcast episode features an interview with Flo Crevello, founder of Lindy, an AI assistant that aims to automate tasks by connecting to users' applications and data. Crevello explains how Lindy uses large language models like GPT-4 to write code, use APIs, and perform actions such as emailing, calendaring, and prospecting. He draws parallels between the decreasing cost of AI code generation and historical drops in material costs, predicting a future of disposable apps. The discussion covers Lindy's technical underpinnings, its focus on reliability and privacy, the competitive landscape, and broader implications for labor, AI safety, and the future of computing.

Main Topics: Lindy's Product Vision and Functionality (Priority: 5/5): Lindy is an AI assistant designed to automate executive assistant tasks like email management, calendaring, travel booking, and prospecting. It connects to user accounts (Gmail, Google Calendar, etc.) and uses LLMs to write code and execute tasks. Technical Architecture: Models, Code Generation, and Data Handling (Priority: 5/5): Lindy uses GPT-4 for most operations, but is building its own fine-tuned models for quality and cost control. It generates single-use code for each task and handles user preferences via context windows and data through context injectors and runtime API calls. Cost Dynamics and the 'Disposable App' Future (Priority: 4/5): Crevello compares the plummeting cost of AI-generated code to the historical cheapening of aluminum, predicting a world where apps cost cents to build and are used for single sessions before being discarded. Competitive Landscape and Startup Advantages (Priority: 3/5): Crevello downplays competition, arguing that the market is huge and startups have structural advantages over incumbents like Google and Apple in speed, focus, and risk-taking, particularly in B2B use cases. Reliability, Guardrails, and User Trust (Priority: 4/5): Lindy prioritizes reliability over initiative, using a combination of perplexity models, stakes assessment, and confirmation prompts to avoid errors. Read-only actions are automated; read-write actions always require user confirmation. Privacy, Constitutional Principles, and AI Safety (Priority: 4/5): Lindy has seven constitutional principles, including privacy and reliability. Crevello advocates for AI regulation due to existential risks and misinformation, and emphasizes that Lindy will never sell user data. Broader Implications for Labor and Society (Priority: 3/5): Crevello is optimistic about AI's impact on labor, arguing that economies adapt and new jobs emerge. He sees AI freeing people from menial work and boosting productivity, with a gradual decline in working hours.

Key Arguments: AI code generation will become so cheap that disposable, single-use apps will be the norm, analogous to cheap aluminum replacing precious metals. Startups have a structural advantage over big tech in the current AI landscape due to speed, focus, and ability to serve specific B2B use cases. Reliability is Lindy's top priority, balanced against 'comfort with ambiguity' through perplexity and stakes assessment, always erring on the side of caution. The future of computing is conversational: users will talk to their computers, which will then execute tasks using a combination of APIs, web browsing, and custom code. OpenAI's removal of log probabilities from GPT-4 is a drawback for building guardrail models, motivating Lindy to develop its own fine-tuned models.

Data Points: Cost to build an app today: $10,000 to $100,000 - Crevello contrasts current app costs with a predicted future cost of one to ten cents. Cost reduction in LLM usage: 20x - Cost of models has divided by 20 over the last year, from six cents per thousand tokens to 0.2 cents. Lindy's current cost per customer: Dozens of dollars per month - Crevello accepts this high cost, expecting it to drop rapidly. GPT-4 context window size: 32,000 tokens - This large context window allows Lindy to stuff all user preferences directly into the prompt without a vector database. Historical reduction in US working hours: 30-40% fewer hours than 100 years ago - Used to argue that Keynes' prediction of reduced work is happening, albeit slowly. Percentage of US farmers in early 1900s vs. today: 90% then, 5% now - Illustrating economic adaptation to automation. GPT-4's speed (response time observed): Several seconds - Noted as a downside, but considered faster than a human assistant.

Pivotal Quotes: "It's an app today costs of the order of $10,000 to $100,000 to build. We're rapidly approaching a world where an app costs of the order of one or 10 cents to build. And when that happens, you basically. Start having disposable apps. You basically start building an app just for this one session that you have right now." — Flo Crevello: Explaining the core thesis of Lindy's 'disposable app' approach to code generation. "I think the computing experience of the future is not you doing work on a computer, it is you having a conversation with your computer." — Flo Crevello: Describing the paradigm shift from direct manipulation to conversational interfaces. "We don't care about cost, for example. So right now, that's the guidance. I've given as a team, it's like, ah, like GPT4 is so expensive. And I'm like, I don't want to have this conversation. We're not talking about costs in this room." — Flo Crevello: Emphasizing that output quality is the single criterion for model selection, overriding cost concerns.

Implications: Crevello's vision signals a shift toward highly capable, conversational AI assistants that integrate deeply with users' tools and data. For the industry, this means intense competition and a race to build reliable, safe systems. Listeners should expect a rapid proliferation of agentic AI that automates complex workflows, raising questions about privacy, job displacement, and the need for regulation.