LLMs are multi-layered user interfaces

Software engineer and blogger argues that large language models (LLMs) should be viewed as multi-layered user interfaces rather than anthropomorphized AI, drawing parallels to touchscreens and the mouse. The post contends that the true innovation lies not in the underlying math but in the surrounding software and integration, which lowers the barrier to human-machine interaction.

LLMs are multi-layered user interfaces Why are we here why-are-we-here I have this hypothesis: the LLMs should be treated like user interfaces, similar to a touch screen is to many physical devices this day’s. In this post I will try to think out loud debating in pro of the above hypothesis, in a way that we stop the anthropomorphization of all what is AI today. I had a daydream i-had-a-daydream I like to think and daydream, more so, it is automatic. Specially when doing tasks that don’t require much mental effort. And you know, our life is filled with such tasks, so I will have a lot of ideas that I would like to just express somewhere. Given the nature of my job as Software Engineer or “Dev”, sometimes human interaction is a luxury that can’t be spent rambling or babbling about something that only makes sense in your head and has no end. I once tried to argue to my wife: that all the pedestrian crossing should be legally labeled as “bike crossing” and cars should always stop for anything that is not another car. Instead of having a sign for just pedestrians in some spots and for both bike riders and people in others. There should be universally reflected with a sign or two that crossings are for people even in bikes and that people in cars are not people . That way we will have a consistent interface and mental model for all car drivers, pedestrians and bikers. Pedestrian crossing: car stops all other can go or ride through. Of coarse she didn’t see the point, when she was in the back seat with the kids. Many people around the world, through time, have been dreaming of having human like interaction with machines. Science fiction is filed with robots and AI stories, so when we had our ChatGPT moment, dreams became a reality… right ? Means to an end means-to-an-end LLMs are starting point, where the goal is AGI. Remember AGI ? Wasn’t the whole point of OpenAI to create safe AGI ? As stated by wikipedia 1 user-content-fn-1 Artificial general intelligence AGI is a hypothetical type of artificial intelligence that matches or surpasses human capabilities across virtually all cognitive tasks You can see how something that a lot of people have been dreaming about, looked like it was right around the corner when ChatGPT was released. As things have settled down, we have forgotten the dream, lost the sense of wonder and excitement in regard to what this technology could be. Mainly because everybody realized that LLMs are just math. Or instead we have now focused on jobs, bubbles, IPOs and money. I see some sort of global collecting thinking from the majority: LLMs are just models, like a model used to predict the weather and that it is the end of the road for this arc. And that AGI is still a dream that has been lost or is just in a mind of a few. You can’t win unless you learn how to lose yourself you-cant-win-unless-you-learn-how-to-lose-yourself LLMs are a complex system of mathematical prediction, the same way the software in touch screen device tries to predict your intent. In this analogy, what is the actual hardware that resembles the actual screen? I don’t know? From where I see things, even if I sound lost and incoherent, the point is that a touch screen wasn’t the innovation, it was where it was put and everything around. This sum of all parts, is what created the iPhone. Same for the mouse, the innovation wasn’t the track ball, but that fact that now we lowered the bar to using a computer system. No need to type fast. People just use the mouse to point, as an extension of their body for interacting with a machine, similar how we do with car wheels. It gave us leverage and affordance towards efficiency and effectiveness when using a machine like a computer. Now LLMs and their harnesses and all the software machinery around it , enables users to leverage our mind and describe actions in words, which itself, is a chaotic representation of our thoughts. We finally don’t need to think linear, or deal with minimal and bad design. Of coarse the challenge is speed, they feel like mainframe with a terminal in the basement. You don’t go downstairs, you use the terminal. You write in your computer the instructions, sometimes incoherent, they go to some cloud provider and then back with some text that is either used to interact with the harness or as output. Fin fin Many think, as hardware innovation continues, computers will have enough power to run the models locally and get our version of the “300 millisecond feedback loop”. Honestly, it doesn’t matter, because like touch screen they will be and basically are going to be everywhere. We will learn that we over did it. As a result and hopefully, maybe… spend time in thinking about the user, the people, because LLMs are not people, they are means to an end.