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

Episode Summary

Executive Summary: Flo Crevello presents Lindy as an AI assistant that connects to users’ apps and data, writes code to complete tasks, and can act like a virtual executive assistant. He argues that falling software-building costs will enable disposable, task-specific apps and conversational computing. The discussion covers Lindy’s workflows, model strategy, privacy and reliability safeguards, market positioning, and the potential effects of AI on work and society.

Main Topics: Lindy as an action-taking AI assistant: Lindy aims to turn natural-language requests into work across services such as calendars, email, Slack, Salesforce, and meeting tools. It can search, draft, summarize, recruit, prospect, and coordinate workflows, using APIs and generated code. Disposable, task-specific software: Crevello compares cheaper software creation to the falling cost of materials such as aluminum: as code becomes inexpensive to generate, an assistant can create a one-off app for a particular session and discard it afterward. Assistant experience and market strategy: Crevello sees conversational computing as a broad shift for both consumers and professionals, while Lindy initially targets busy professionals and executive-assistant tasks. He argues startups can move faster and focus more narrowly than large consumer platforms. Models, context, and personalization: Lindy primarily uses GPT-4 for quality while building model-agnostic infrastructure and fine-tuning its own models for specific tasks. It combines stored preferences, a context-injection system, and code that retrieves relevant information from connected services. Reliability, privacy, and safety: Lindy emphasizes privacy and reliability, and balances acting independently against the risk and stakes of mistakes. It distinguishes read-only research from consequential read-write actions, which require user confirmation. The company also uses stated principles to guide product decisions and model training. Employment and broader AI risks: Crevello expects AI to automate menial work and increase productivity, and is not especially concerned about near-term net job losses given economic adaptation. He identifies misinformation and longer-term AGI risks as concerns and supports AI regulation.

Key Arguments: An assistant that can access applications and data can turn a request into a multi-step workflow—for example, checking calendar availability, emailing options, and scheduling a meeting after a reply. As the cost of generating software falls, it becomes practical to build disposable apps tailored to individual tasks rather than relying only on prebuilt applications. Lindy’s initial professional focus gives it room to serve specialized needs that broad consumer assistants may not prioritize; startups may also iterate faster and tolerate rougher early releases. Model quality is the current priority over serving cost. Crevello says GPT-4 substantially improves performance and that Lindy is developing its own models and infrastructure to switch models as capabilities evolve. Personalization can be supported by placing a user’s preferences directly into the model context, while a context injector and runtime code retrieve relevant details from connected data sources. Autonomy must be constrained by reliability: Lindy considers both its confidence in an action and the consequences of an error, and asks for confirmation before read-write actions. Crevello expects automation to free people from repetitive work and argues economies can adapt as jobs change; he also acknowledges misinformation and non-zero long-term existential risks, advocating regulation.

Data Points: Funding raised: $50 million - The hosts describe Lindy as having raised this amount ahead of its announcement and limited beta. Estimated cost to build an app: About $10,000–$100,000 today, potentially about $0.01–$0.10 - Crevello’s comparison for how dramatically software creation costs may fall, enabling disposable apps. Model serving cost: Dozens of dollars per customer per month - Crevello says this was Lindy’s approximate current cost and not a near-term priority relative to quality. Context window discussed: 32,000 tokens - Crevello says GPT-4’s larger context window makes it easier to include user preferences without retrieving them from a vector database. Potential share of work affected: At least 25% - The host characterizes a study of anticipated labor-market impacts as suggesting language models could eventually take on at least this share of work, especially with tools. Change in annual work hours: About 30%–40% fewer hours than 100 years ago - Crevello uses this as evidence that working hours have declined over time and may continue to do so. Farm-worker share of the active population: About 90% in the early 1900s; about 5% today - Crevello cites this historical shift to argue that economies can adapt to major changes in work.

Pivotal Quotes: "We're rapidly approaching a world where an app costs of the order of one or 10 cents to build." — Flo Crevello: He argues that cheap code generation makes single-use, task-specific software practical. "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: His vision for assistants as a new interface layered over existing applications and data. "Given the choice between both of these principles, reliability always wins. We never sacrifice reliability." — Flo Crevello: He explains how Lindy balances initiative with caution, especially for consequential actions.

Implications: If assistants reliably connect models to users’ tools and data, computing may shift from navigating separate apps to delegating tasks conversationally. Disposable software could expand customization, while privacy, confirmation safeguards, reliability, labor transitions, and AI regulation become central concerns.