NATURAL INTELLIGENCE

What are LLMs and why should I care

 If you’ve used ChatGPT in the past week, or more likely the past hour, you’ve used an LLM. One of the most relevant trends in modern technology, I’ll explain what LLMs are, and why they are important to understand.

By VICTORIA FRANK
(Alanna Jimenez / Daily Trojan)

Before forcing myself to enter the tech world, I found it very intimidating. Sure, my generation practically grew up with iPhones, giving us an advantage at intuitively navigating new interfaces — which is apparent when interacting with older generations, like when I increase the volume on my mom’s iPhone and she applauds my computer savvy. However, with the rapid pace that technology is evolving and the lack of hardware and software knowledge known in non-STEM populations, it’s easy for people like me to throw our hands up and surrender to the wave of ones and zeroes. 


Daily headlines, sent straight to your inbox.

Subscribe to our newsletter to keep up with the latest at and around USC.

Writing this column proved intimidating indeed, as I am not an expert on artificial intelligence. However, here I go anyway, wearing my naivete on my sleeve.

Serious discussions about AI have been in circulation since 1950 when Alan Turing published his seminal paper “Computing Machinery and Intelligence.” Turing, an English mathematician whose death occurred following a court-ordered chemical castration after being “convicted” of homosexuality, was ahead of his time. His research brilliantly explores theological approaches to a machine soul, the difference between human and machine intelligence, and the future of differentiating between the two — also known as the Turing Test, which is still used today. 

All of this is to say that artificial intelligence is not some new product of 21st-century advancement. Nor is artificial intelligence easily simplified into one category, and is sometimes considered an entire field of study rather than just a type of technology.

Large language models, or LLMs, are a specific application of artificial intelligence, unquestionably the most popular today. 

In almost all cases, an LLM works in two steps: step one, an entry made by the human user, and step two, a response given to the user by the AI. Both entry and answer can take form textually, auditorily or visually; think ChatGPT versus Alexa versus Dall·E. 

To understand how complicated the process is between steps one and two, consider some of the following questions AI programmers must reconcile: What is the relationship between a literal sentence and its figurative meaning; i.e., can a machine decipher the feelings it evokes, the sarcasm it contains and the societal context it ignores? What parts of life can and cannot be datafied — a process called datafication. For example, can you turn the feeling a parent has when they hold their child into a piece of codified data? 

Another issue is the raw amount of computing power required to scan, decode and categorize the entire internet … and how much money that hardware costs. In an anecdotal example of the consequent financial shifts, Microsoft, after disclosing its plan to lay off 10,000 employees in January 2023, recently announced a $10 billion investment in OpenAI, the creator of ChatGPT.

In short, through a complex inner system of training code, data collection and neural pathway categorizing, LLMs are very talented at predicting language (hence the name). This is how LLMs spit out long, in-depth, human-sounding responses so fast. However, the predicted language they deliver does not promise accuracy. 

Google’s search engine is so mainstream that it’s now a verb — “Let me Google that!” However, when we search the internet for information by googling, we must translate our inquiry into appropriate search terms, or as explained by my professor Morten Bay, an adjunct instructor of communication data science, we must understand “keywordese.” ChatGPT, the program associated with the recent mass distribution of AI, enables users to search the internet through the same verbiage they use when talking to their friends (known as natural language processing). There’s no “keywordese” necessary; this might explain its immediate popularity.

It is crucial to remember that LLMs like ChatGPT, as they exist now, are not interchangeable with search engines like Google. LLMs sometimes “hallucinate” false information, because they are primarily trained for convincing communication, not literal accuracy. 

As you may imagine, there is a near-infinite supply of unsatisfied ethical questions, technical complexities and economic impacts that weren’t touched upon in this limited window into the world of LLMs (one example; what happens when someone wants an LLM girlfriend?). Some of these topics will be thoughtfully explored in future “Natural Intelligence” columns.

With all that said, you should care about LLMs because chances are you’re either using one now or will have to use one very soon. As always, in making the case for a tech-optimist future, I advocate that every individual educate themselves on the technology that is exploding into our everyday lives. That way, we don’t have to rely on a handful of tech bros in San Francisco to make the decisions for us.

Victoria Frank is a junior writing about the inevitable AI future with a focus on ethics and well-being. Her column, “Natural Intelligence,” runs every other Friday.

© University of Southern California/Daily Trojan. All rights reserved.