Discussions at the dayjob have included the both the risks of AI and the risks of not implementing AI.
I mean, the problems of AI as an accelerator for cybercrime were already apparent and are now becoming iridescent (metaphorically).
But my guess is that most of you are dealing with workplaces where there’s a cohort of people wanting to implement AI to do everything. They want a silver bullet. They really need to understand some of the points in last week’s post and the MIT Iceberg Index. Simply put: AI can often perform or accelerate some tasks of a job, but not the whole job. But odds are, if you started saying that , they’d stop listening. They don’t want a tool that can help in a variety of use cases, they want a silver bullet/thinking machine worthy of sparking the Butlerian Jihad. (And if you really want their eyes to glaze over, start talking about the Butlerian Jihad).
But what if there’s another consideration? What if the silver bullet simply can’t kill all the werewolves of an efficiency expert’s fever dream? It’s just not that powerful and never can be. Or, put less colorfully, what if there’s a limit to what an LLM can do, just computationally?
I first learned of this possibility when I came across this video by Caleb Ulku, someone who is both using AI and gives every indication that he wants to keep doing so. And then I went and read the paper he mentioned, “Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models.” And then I read it again, both because it’s a short paper and there is a decent amount of math referenced.
Basically, the authors explore the limitations to the actions LLMs can perform and have a hypothesis that there is a ceiling to the quality output before LLMs hit a computational barrier. And yes, I’m sure any plucky marketer would note how the sound barrier was broken, but they’re not the poor soul trying to deal with an increasing amount of LLM hallucinations.
So, to my mind, that’s another reason to really look into how to to chop up AI deployment into logical use cases.
I also might want to incorporate “Come on, Fhqwhgads” as part of an AI Prompt. Y’know, for science.