Enough with Prompt and Pray

Yoav dislikes the term AGI. He believes intelligence is just a combination of multiple capabilities, and the narrative that we'll suddenly reach AGI at some point is an exaggerated fantasy.

Companies champion AI-first and pilot hundreds of use cases, but few projects actually make it to deployment. There are two obstacles.

• Cost: High LLM serving costs undermine the economics of business software.
• Reliability: Even if most outputs are correct, occasional off-the-mark outputs can be fatal in investment memos or customer service.

The cost problem can be eased by smaller Small Language Models (SLMs) in the sub-7B, sub-3B range, and by more efficient architectures that aren't transformer-based, such as Jamba. The reliability problem, on the other hand, stems from the probabilistic nature of LLMs, so it won't disappear through alignment or guardrails alone.

So he argues we should abandon prompt and pray and move toward AI systems that combine LLMs, retrieval, tools, and traditional code. If a small LLM acts like a router that splits up tasks and inserts intermediate verification, it can boost controllability, efficiency, and reliability together—especially for tasks requiring non-trivial reasoning.

The challenge for startups is product-algo fit. The message is that a product must make the most of AI's strengths, with design compensating for the remaining imperfections, in order to actually reach deployment.