Artificial intelligence is big news in 2023. Businesses are rushing to use it for a competitive advantage. But can AI really help? Or does it merely generate a lot of subpar blog posts and meta descriptions?
ChatGPT, Bard, and other language models will undoubtedly create a ton of inferior blog posts. Yet AI is entering a new phase that could produce many new opportunities. IBM described the advances in 2023 as a “step change in AI performance and its potential to drive enterprise value.”
Understanding the developments that have enabled those advances may help managers and owners at retail, ecommerce, and direct-to-consumer businesses employ AI to their benefit.
Ask someone how ChatGPT works. You might hear phrases like “large language model,” “generative AI,” or “vectors.” All describe aspects of ChatGPT and similar platforms. Another answer is to say ChatGPT is a foundation model.
An AI to predict the best-selling price for a product on an ecommerce site once required training that model on thousands or even millions of transactions. It would get the job done, but would take time.
A foundation model takes the process back a step. It is trained in an unsupervised way on a much larger set of information — the entire internet.
This generalist approach differs from traditional AI models trained for a singular, specialist task and is analogous to a virtual jack-of-all-trades. It leverages a broad knowledge base to perform an array of tasks, from producing human-like text to recognizing patterns in complex data sets.
Such a model excels in its flexibility. Its initial training in comprehensive and diverse data equips it with a foundational understanding of many topics.
The foundation can be fine-tuned for specific applications — such as predicting the best-selling price for a product on an ecommerce site — in a fraction of the time, data, and resources as previously required, making it potentially transformative.
Generating code, copy, or anything is the next concept to consider. Take the real-world example of GitHub Copilot. It uses OpenAI to process natural language into entire sections of code.
Platforms such as GitHub Copilot and ChatGPT are said to be “generative” when they create something that did not exist in that form.
Google is evolving to generative AI search results. The search engine will no longer present only links to web pages or even snippets from those pages. Instead, it will generate answers — albeit ones containing links.
In the context of ecommerce, a merchant could share a slew of product specifications with a generative AI tool and ask it to produce descriptions of those items.
Compare that to my article in 2016, “How to ‘Manufacture’ Product Descriptions for Ecommerce,” which described a five-step (manual) process.
My article used a soup spoon as its example product and ultimately developed this product description.
Hungry for some hearty chicken noodle or creamy clam chowder? This soup spoon has a large bowl meant to haul bisques and broths to your mouth. In fact, this soup spoon can hold about three times as much soupy goodness as your standard table spoon. You could take three times as many bites, or you could buy this soup spoon and slurp large.
I recently pasted that article into ChatGPT and asked it to learn the process and write a product description for a hammer. Here is what it generated.
Ready to take on your next DIY project? This hammer is built with a high-grade steel head, ensuring lasting sturdiness. It’s designed to deliver powerful, impactful hits every time you swing. Get the job done right and add this reliable hammer to your toolkit today.
The AI created a new hammer description. It combined its foundation model knowledge with the specific process I gave it.
Generating something new is part of what makes a foundation model so flexible.
Large Language Model
AI generates many things with images, audio, and videos. But text is among the most important forms of generative AI for business.
Tools such as ChatGPT, Google’s Bard, and Jasper introduce us to another concept: large language models.
An LLM uses its foundational knowledge to predict what word should follow another.
Last week I watched the engineering director of a software company give a short presentation about AI to his company. He cleverly described LLMs.
“I want to talk about how these models work,” he began. “I don’t really know what I am talking about, so think of this less as a factual presentation and more of me stringing together a set of words, each word making sense based on the words preceding it, but without 100% correctness being my primary goal.”
When you provide it with “don’t cry over spilled…,” an LLM will likely come up with the word “milk.” It can guess that word because of its foundation model.
Understanding foundation models, generative AI, and LLMs helps us contemplate how artificial intelligence creates business opportunities. Thus we wouldn’t typically ask ChatGPT to develop a product. But we could ask it to analyze market gaps for potential product opportunities.