Grab a cuppa and let’s talk.
I really wish everyone would stop calling LLMs AI. Artificial Intelligence. What intelligence?
Terms matter.
I remember when the term “hacker” meant something completely different and it was the “crackers” we had to worry about. I also remember when algorithm had more to do with step-by-step instructions for performing a task and not a term to represent an opaque, qausi-mystical entity shaping how platforms strive to keep your attention. Don’t even get me started on the concept of “cloud”.
I am from the hacker culture. I am also a gamer, a reader, and a huge lover of technology. Though you’ll see in a future article, that has started to change. So when I think of the term “AI”, I am drawn to visions of Mass Effect’s Geth or EDI. The reality is that all the “AI” the world is over-hyping right now reminds me more of the VI Avina found on the Citadel, just less intelligent.
Understanding what’s happening behind the “AI” curtain
When you start typing a message on your phone, the keyboard software uses a small machine‑learning model to suggest likely next words based on what you have typed before and on patterns it has learned from text. It usually shows just a few options because the model is small and designed to be fast and lightweight on your device.
A large language model (LLM ) works on a similar principle but at a far bigger scale. It is trained on very large collections of text so it can learn statistical patterns in how tokens (pieces of words, not just whole words) tend to follow each other in different contexts. At generation time, it takes the text so far, predicts the next token, appends it, and then repeats this process many times to produce a full response.
Modern LLMs add extra training steps so they do not just produce the most probable raw continuation, but instead try to follow instructions, stay on topic, and be useful. So when thinking about how tools like ChatGPT, or Perplexity, or Gemini, or any other LLM out there works, looking at it like “next‑word prediction” picture is a good starting pont. These models are large, trained on vast datasets, operate on tokens rather than only words, and are further tuned to behave helpfully rather than just autocomplete text.
The Good, The Bad, and The Very Very Ugly
Give an LLM a very specific dataset and some interesting things can happen. Cambridge researchers are using machine learning to help figure out how well patients respond to specific treatments before giving it to them[1]. The Univerisity of Oxford and Google Cloud have done research to show that a LLM can help figure out cosmic events[2]. As with all technology, the power is in the hands of the user of that tech and that tech has power demands which are insane.
From individuals mourning lost love after a ChatGPT update[3] to individuals dying due to conversations had with ChatGPT[4], the results are painful to watch from the outside. I am nowhere near knowledgeable enough to discuss the personal and societal impact this is having. It feels off to me. It is almost like the large companies who run these LLMs and train them, just care about the big almighty dollar and try to keep individuals from understanding what an LLM is really doing under the hood.
There is also the negative environmental impacts the ubiquity of LLMs is having[5][6] and the astronomical cost increases on other technology, like graphics cards, memory, and other storage[7]. There are issues with artists having their work stolen, books being shoved onto shelves that were “written” by AI, and the concept of using a LLM as “AI” being plugged into everything, and I just want to throw my hands up and walk away regardless of the good that could come out of using a well trained LLM.
This is Not Mass Effect
At the start of this article I stated, “I really wish everyone would stop calling LLMs AI.” If an LLM is trained for a specific purpose, it’s machine learning. If an LLM is being used to write code (and I really do not think anyone should be using an “AI” to write code), then call it a programming tool. The anthropomorphism of predictive text leads to scary situations.
Footnotes
[1] https://www.cam.ac.uk/stories/cambridge-cancer-research-ai
[4] https://www.youtube.com/watch?v=hNBoULJkxoU
[6] https://docs.un.org/UNEP/EA.7/L.14
This article is inspired by the incredible course by Sabrina Goldfarb on FrontendMasters.