“There are things known and there are things unknown and in between are the doors of perception.” — Aldous Huxley

I’m Junior Huxley Westemeier and welcome to The Sift, a weekly The Rubicon opinions column focused on the impacts and implications of new technologies.

______________________________________________________________

I’ve been a skeptic for so-called ‘Artificial Intelligence’ since OpenAI publicly released GPT-3 in November 2022. As an avid programmer and someone interested in applied mathematics, I understand the math behind these models. OpenAI used massive amounts of textual data- estimated to be around 45 terabytes for GPT-4 (or the rough equivalent of taking every SPA senior’s 512gb Lenovo laptop and filling the storage with random books and articles). But the actual ‘logic’ of how ChatGPT ‘responds’ to the user is an entirely different process. ChatGPT is looking at patterns in existing text to attempt to answer your query. There’s no actual thought processes happening. A complex arrangement of mathematical transformations mimic human-generated text- nothing more than a smartphone autocorrect algorithm on steroids.

Reliance on existing data means that it’s challenging for ChatGPT to generate anything original, since it’s taking bits and pieces from existing content. But it also leads to ‘hallucinations’, or instances where the algorithm up-front lies about something. ChatGPT is not malicious, it simply just doesn’t understand what is factual or misleading since the mathematical modeling functions don’t think, they just copy existing patterns. These moments are where the facade begins to fall apart, as they show that the ‘AI’ you’re talking to isn’t as intelligent as it seems. I asked ChatGPT-4o a question that’s been endlessly shared on social media: how many r’s are in the word strawberry? On ten separate occasions, it successfully identified the three r’s only four times. Any human could do better.

On Sept. 12, OpenAI announced its latest model: o1-preview or Strawberry (the internal code name is ironic considering its past mistakes with miscounting r’s). They ran Strawberry through the qualifying exam for the International Mathematics Olympiad, and it solved 83% of the problems accurately compared to the older 4o model which only solved 13%. Strawberry is meant to mimic a human’s thought process more than ever before by trying different strategies and outlining the steps it takes before reaching a conclusion.

I ask every student to question why such innovation is necessary at all. Why do we need a chatbot that could (successfully) cheat on various exams and write essays for us? What is the benefit to society for a website that can give misleading information about a product, potentially give wrong medical advice, or write code with security errors? How long before these hallucinations and mistakes are accepted as reality because they are delivered by a convincing AI?

Perhaps in a few years, there will only be two r’s in the word strawberry. After all, the AI said it so it must be right.