Agents are just small software that's multitudes easier/faster to build, but multitudes more expensive to run, and less accurate ( therefore require building extra systems that allow human supervision, and hiring people ).
I think most businesses are going to chose the >= 99% accurate route, that has the higher upfront cost, after trying today's agents. Actual software is waaaaaaaaaaaaay cheaper in the long run after a company makes the investment.
It's not something you'd want to run at high scale due to the costs.
Business owners have been FOMO'd into dumber stuff before, for example, 'no code' systems.
A bit ago, OpenAI demonstrated an extended-thinking o3 model that used 1000x the compute as usual to solve a basic visual puzzle. >1000x the compute was required to gain an additional ~15% accuracy. What was not said was the electricity cost.
This demonstrated to me that additional accuracy is going to be super expensive right now. Humans are expected to be >98% accurate at their job, and also handle novel situations they haven't been trained on. That's the true benchmark for AI, and it's waaaay off the mark.
I'm a lot less impressed by it than the guys that are drunk on their kool-aid ( Jensen, Altman, et al )
I think most businesses are going to chose the >= 99% accurate route, that has the higher upfront cost, after trying today's agents. Actual software is waaaaaaaaaaaaay cheaper in the long run after a company makes the investment.
It's not something you'd want to run at high scale due to the costs.
Business owners have been FOMO'd into dumber stuff before, for example, 'no code' systems.
A bit ago, OpenAI demonstrated an extended-thinking o3 model that used 1000x the compute as usual to solve a basic visual puzzle. >1000x the compute was required to gain an additional ~15% accuracy. What was not said was the electricity cost.
This demonstrated to me that additional accuracy is going to be super expensive right now. Humans are expected to be >98% accurate at their job, and also handle novel situations they haven't been trained on. That's the true benchmark for AI, and it's waaaay off the mark.
I'm a lot less impressed by it than the guys that are drunk on their kool-aid ( Jensen, Altman, et al )
