It seems we can hardly go a day without hearing about some newfangled way that Artificial Intelligence is going to take all our jobs, then give us lives of all play and no work, then ultimately end humanity as we know it. Whether we like it or not, we now have new terms to keep track of like AI, GPT, LLM, Copilot, Gemini, Grok, and Claude.
When people talk about AI, what they’re really talking about are Large Language Models, or LLMs. Literally, large models of language. These models consist of a massive corpus of text (think all of Wikipedia), which are then distilled into “tokens” made up of each little variation of letters and symbols. Then count, sort, and categorize those tokens, and you’ve got yourself a Large Language Model.
How did we get here? In 2018, a little-known company called OpenAI released a Large Language Model called GPT1 to little fanfare, then followed that with GPT2 in 2019. These first two models were largely focused on text prediction, or as Google says, “they excel at generating coherent, human-like text from prompts”. Pay attention to that wording: generating; coherent; human-like. The world really took notice during Christmas break of 2022 when OpenAI graced us with GPT3. It was no longer just some math whizzes and programmers building a text-prediction and translation tool. LLMs would soon be shoved into every nook and cranny of every piece of software you’ve ever used, whether you asked for it or not.
How does it work? Imagine taking a batch of Wikipedia articles on the history of the United States, and cataloging how often each word appears, how likely each word is to appear next to other words, and in which order. Next, use those probabilities as weights to predict the next word in a given sentence. Now, do that over and over and eventually you can generate a new wiki, but for, say, the history of Canada instead. What you will end up with is a document that reads like a plausible and grammatically accurate article, but with gaping holes in accuracy. However, these gaping holes will only be recognizable to those who know their Canadian history. To the untrained eye, the article will initially feel like it has substance, texture, story, meaning; as if you’re learning so much about Canada without trying very hard! How convenient! Upon closer inspection, you may be disappointed to find out that many of the facts, dates, and stats in the History of Canada are not, in fact, real. ChatGPT even states this right up front: “ChatGPT can make mistakes.” It’s a black box that gives you what it thinks you want, rather than what you need. It breaks down our language into its base parts, and then reconstructs a plausible facsimile, the mere appearance, of the real thing.
How about another example? Let’s say you gave the entire Shakespear catalog to an LLM, then said, “generate another play”. It’s true that our new generated play would have similar words and structure of Shakespear, the plausible flow of a classic play, the recognizable roller coaster of emotions, and character development to boot. However, it would also lack intentionality. It would have all the literary aspects of a masterpiece, except the integrity, the substance, or the taste. It would appear delicious to the eyes and ears but would be bland to the heart and soul. A document with big words, but lacking nourishment. Masquerading as something that it’s not. Sounds a lot like our food labels.
Why do I tell you all this? Because we’ve lived this world already, but with our food system. We already know how this play ends. Our food system is designed to break down our food into its base nutrients, and then design, stack, and build back up the perfect ratio of what will look good, last long, transport safely, and store easily, and after all that is complete, it also needs the appearance of “food”. Just look at a label for coffee creamers, protein drinks, nut butters, or any snacks. This system, taken to its extreme, looks like food that has very little in common with what our body craves, food that is very expensive, food that we have no say in, and food we don’t know or even recognize. Eventually, we will be unable to recognize the difference between what is “real” food and what is “generated” food. Worse still, we may be unwilling to care.
But our story is not done being written, and we still have a choice. We can choose not to attend that play, and instead write our own play, with more taste. Eating local food is the best way to know your food and your makers and growers, and ensure the food is actually real.
~ Tobin




