2025: LLMs should prove reliability

Hi, I’m Sergey Osipov @s_osipov, cofounder of Placy (AI real estate agent). Launched about 20 startups and had some exits, including a unicorn IPO. Sharing here my observations and tips for AI founders. Join! #AI #Startup #PropTech

llmsreliabilityOpenAI

2024: LLMs proved their ability to do everything we love about AI. But it was running smoothly in <90% of cases

2025: LLMs should prove their reliability. The task is to create 99.9% great service. Quality, not quantity

Quite useful OpenAI report: AI in the Enterprise. Lessons from seven frontier companies. Start reading from page 5

#OpenAI #LLMs

Дискуссия

Bunny K
90%? That's a solid C!
Photo
😄😘
Andrey Kozik
Moving from 90% to 97% will require at least twice as much effort and computing power as it uses right now. Moving from 97% from just 99% will require to double the efforts once again. So moving from 90 to 99.9 will require at least x8 (or even x16) resource usage. And the question is, will it be still affordable and even possible. So either we need to learn how to efficiently use it probabilistic nature, or wait for pretty huge breakthrough on the conceptual level of how AI algos actually work.
Andrey Kozik
Just asked AI about this. And based on its calculations it says that x16 is kinda optimistic. It will be closer to x100
Присоединиться к обсуждению →