Before LLMs there was another kind of AI, the old school one "expert systems", agent modelling and Formalization of systems. In many cases it would be a powerful combo and could corner modern AI systems. But companies did not spend time in upskilling employees to apply those advances. Now they fall for the "cheaper" LLM alternative because they take something "ready-to-apply" as a utility and is fancy. This comes with huge cons and negative societal impact. But they have money, impunity and halo because people allowed that, and so they have "carte blanche" to bend logic.
Nature is a machine, it has no sensitivities, the moment you defy its laws, you will pay dearly.
Fwiw the article is arguing the opposite: LLMs are often the expensive option, and the decision models it describes are learned models too, not rule-based systems. Combining both is an active research area though (neurosymbolic AI). Personally I’ve always thought LLMs should only be used where they actually help, and if there’s a better conventional solution, use that. There’s a reason LLMs themselves often write a quick script to solve certain tasks correctly instead of doing it “in their head”.
That's dangerous thinking. Once you spend critical thinking in where you apply LLMs, you get into slippery slope territory, to the logical conclusion that _nothing_ _needs_ an LLM. Please, I implore you, stop before it's too late.
Before LLMs there was another kind of AI, the old school one "expert systems", agent modelling and Formalization of systems. In many cases it would be a powerful combo and could corner modern AI systems. But companies did not spend time in upskilling employees to apply those advances. Now they fall for the "cheaper" LLM alternative because they take something "ready-to-apply" as a utility and is fancy. This comes with huge cons and negative societal impact. But they have money, impunity and halo because people allowed that, and so they have "carte blanche" to bend logic.
Nature is a machine, it has no sensitivities, the moment you defy its laws, you will pay dearly.
Fwiw the article is arguing the opposite: LLMs are often the expensive option, and the decision models it describes are learned models too, not rule-based systems. Combining both is an active research area though (neurosymbolic AI). Personally I’ve always thought LLMs should only be used where they actually help, and if there’s a better conventional solution, use that. There’s a reason LLMs themselves often write a quick script to solve certain tasks correctly instead of doing it “in their head”.
[flagged]
That's dangerous thinking. Once you spend critical thinking in where you apply LLMs, you get into slippery slope territory, to the logical conclusion that _nothing_ _needs_ an LLM. Please, I implore you, stop before it's too late.