A new kind of AI model from a ChatGPT inventor is thrilling developers

AI automation is now cheaper and more reliable, enabling new kinds of intelligent software.

Former OpenAI researcher and RLHF co-inventor Andrej Karpathy has launched TypeSafe AI with a new model named Jev, designed to address the limitations of current large language models (LLMs). Unlike LLMs that optimize for human language and can hallucinate, Jev produces probabilities or "calibrated decisions" and does not output text. This approach makes Jev significantly faster and cheaper, with a pricing model based on billions of input tokens rather than millions. Developers are reporting substantial improvements in speed and accuracy when using Jev for tasks like command review and email classification compared to existing LLMs. Jev's output includes confidence scores, making it ideal for automating workflows and potentially serving as a check on LLM behavior or for intelligent model routing. Karpathy believes this focus on "System One" intuition rather than language optimization will lead to more widespread, emergent, and distributed intelligence in software.

Jev's departure from text-based LLMs signals a potential shift in AI development, prioritizing calibrated decision-making over linguistic fluency. This could unlock new avenues for AI-driven automation in software, where precise, reliable outputs are critical. The model's speed and cost-effectiveness challenge the dominance of expensive LLMs for tasks requiring quantifiable confidence, potentially democratizing AI integration for a broader range of applications. Its ability to provide confidence scores makes it particularly attractive for automating complex workflows that previously required human oversight or expensive, less precise LLM intermediaries.

The introduction of Jev suggests a market bifurcation where LLMs continue to excel at generative tasks and human interaction, while specialized models like Jev will handle more deterministic, automation-focused AI needs. This could lead to hybrid AI architectures that leverage the strengths of both types of models, with Jev potentially acting as a crucial validation or routing layer. Developers and businesses should watch for the emergence of platforms that integrate Jev-like capabilities, offering more robust and cost-effective AI solutions for operational efficiency and complex decision automation.