Jev: 8 real use cases for the fastest, cheapest model I’ve ever used | John Lindquist

The episode kicks off with Claire Vell introducing her guest, John Lindquist, a repeat guest known from a previous popular episode. They are celebrating 'Jev Week,' a period focused on a new AI model called Jev. Claire emphasizes the excitement around this model, which is not like traditional large language models (LLMs). Instead, Jev is a decision engine that outputs decisions, scores, and probabilities rather than just text, making it significantly faster and cheaper in a production context. They highlight the vast number of demos John has prepared, illustrating Jev’s capabilities across various practical use cases.

Claire and John explore Jev's core features, emphasizing its decision-making architecture over open-ended generative output. They discuss how Jev allows for rapid, real-time interactions, such as dictation apps that classify and execute commands instantly while listening. John demonstrates a to-do list management system where voice commands are analyzed and matched with specific tasks and functions through multiple analysis layers, including confidence scoring. They cover how chaining multiple Jev calls can build complex workflows, akin to traditional programming conditionals, but with speed and simplicity. John also explains how Jev can classify unstructured data into structured formats, supporting use cases like data deduplication, record merging, and navigating complex apps through natural language.

The conversation shifts to how developers can architect around Jev, using it as a router and decision-making layer. John illustrates with examples like guiding users through apps or routing commands to the appropriate back-end functions. They discuss building multi-layer classification systems by running sequential Jev passes, which improves accuracy. Several demos are mentioned: a Wikipedia path mapper connecting topics to philosophy, multi-agent coordination to avoid collisions, and voice coaching tools that monitor speech for completeness and timing in real time. The episode emphasizes Jev’s role in making complex interactions more efficient—such as web navigation, game moves, or even image analysis—by reducing the need for expensive, slow generative models.

John compares Jev's performance in scenarios like chess, where it can reason through moves at high speed, to traditional LLMs, showing Jev is 10 times faster and four times cheaper in these tests. They highlight how speed unlocks real-time capabilities previously thought impossible. Examples include navigating websites by identifying actionable buttons, playing fast-paced games like Tetris, and simulating decision trees for user interface interactions. Discussions occur around how Jev handles constrained inputs efficiently, transforming seemingly infinite possibilities into manageable decision sets. Claire adds insights about image analysis and how Jev can complement LLMs by providing structured decision outputs, especially in data-heavy tasks.

The hosts acknowledge that Jev isn't perfect and discuss situations where multiple classification passes or additional layers are necessary for accuracy. John mentions testing different prompt formulations and the importance of thresholds like confidence scores. They note that while Jev is very fast and cost-efficient, it may not replace all generative tasks but excels in decision and routing applications. Claire asks about practical setup, including managing how input streams are chunked and processed, and they advise iterating prompt design and thresholds for better precision. The episode concludes with a reminder that Jev's current state is early in its adoption, with improvements expected over time.

In the closing segment, John and Claire reflect on how Jev unlocks possibilities once limited by cost and speed constraints. John shares personal anecdotes about managing local hardware setups—hard disks, disks space optimization, and remote Mac Minis—highlighting that real-world use often involves overcoming infrastructure challenges. They speculate on future trends, suggesting that AI decision models like Jev will increasingly replace traditional, slow, and costly AI solutions. The episode ends with a quick lightning round, where John shares interests beyond AI, including multimedia and game development, emphasizing the importance of fast, local processing. They joke about hardware issues and how AI might eventually be embedded directly into devices, reducing reliance on extensive infrastructure.