Unsupervised Learning x Latent Space Crossover Special

What It Covers

Crossover episode between Unsupervised Learning and Latent Space podcasts featuring Swyx, Alessio, and Jordan discussing AI's biggest surprises in 2024, including DeepSeek's rapid advancement and reasoning models. They debate defensibility at the application layer, evaluate which AI use cases have genuine product-market fit, examine infrastructure opportunities, and share predictions on model company strategies and enterprise adoption patterns.

Key Questions Answered

• •Open Source Model Adoption: Enterprise usage of open source models sits at approximately 5% and continues declining according to Braintrust data. Companies remain in use case discovery mode, prioritizing the most powerful models available rather than open alternatives. Each new model generation triggers fresh discovery cycles, preventing teams from settling on open source solutions despite lower costs and licensing flexibility.
• •DeepSeek Impact Timeline: DeepSeek's technical achievements were visible to close observers in 2023, yet public markets reacted dramatically in 2024 with NVIDIA dropping 15% in one day. The gap between technical developments and market narratives typically spans one to two years. DeepSeek R1 represents the first open model with full reasoning traces, differentiating it from previous releases that were merely cheaper versions of existing models.
• •Product Market Fit Categories: Three clear AI agent categories demonstrate genuine traction: coding agents like Cursor, customer support agents like Sierra, and deep research tools. OpenAI's Deep Research launch likely generated billions in revenue from tier upgrades from $20 to $200 monthly subscriptions. Voice AI for appointment scheduling shows 75% effectiveness rates, which exceeds the 50% call answer rate for many service businesses today.
• •Low-Code Builder Failure: Established low-code platforms like Zapier, Airtable, and Notion failed to capture the AI builder market despite having distribution, reach, and technical DNA. These companies improved existing products with AI features rather than reimagining software creation from scratch. Bolt and Lovable reached $20 million revenue in three months by building AI-native experiences without legacy product constraints or preconceptions.
• •Application Layer Defensibility: Network effects provide the primary moat for AI applications, not proprietary models or unique datasets. Chai Research outlasted Character AI through marketplace network effects connecting users and model providers despite lacking proprietary models. Brand establishment happens within six to nine months, allowing companies to become synonymous with categories and command premium pricing while competitors struggle for customer access.
• •Reasoning Model Limitations: Test-time compute scaling works effectively in verifiable domains like coding and mathematics but remains unproven for non-verifiable domains like law, marketing, and sales. This creates a potential future where autonomous AI handles technical work while humans still write basic sales emails. The ability to apply reinforcement learning to subjective domains will determine whether agents or copilots dominate these workflows.