AI seems like a boundless source of knowledge, but it doesn't truly understand what it's saying. Here's why that difference matters.

✒️ Paul Rigden

Knowledge and intelligence, while similar on the surface, are very different things.  Knowledge is what you know, while intelligence is your aptitude for learning and complex thought.  That difference is crucial to really understanding what to expect from language models, and what makes AI so confusing for a lot of people.  Usually, if we meet someone with encyclopedic knowledge on a huge array of topics, it’s probably a good bet they are an intelligent person.  Knowledge and intelligence don’t exactly correlate, even in humans, but they tend to go hand in hand.  This is a connection we assume by instinct, and it’s the driving factor in people’s misestimation of AI’s capabilities when it comes to creativity and complex thought.

While hallucination remains a problem (we’ll cover that later), today’s large language models far exceed any human in the realm of general knowledge.  Anyone who doubts that should try asking a jeopardy contestant to walk them through debugging code in obscure programming languages, or to list the comorbidities of rare medical conditions.  New language models are benchmarked as doing those things with surprising accuracy.  Unfortunately, these knowledge-based skills aren’t paired with the raw intelligence or creativity we’d expect from humans.  Even the best LLMs will struggle when asked to come up with insightful solutions when presented with complex, real life scenarios, perform other tasks like predicting which trends will resonate with an audience, or what kind of content people will want to read.  Unfortunately for content marketers everywhere, the reason for this is deeply connected to how language models are trained, and why they know so much in the first place.

     "

The focus of this series is not computer science, so we won’t focus too much on the technical details (read more about those here, if you dare), but in very simple terms, language models are trained by using the same kind of technology that drives the keyword prediction you see in Google’s search bar, but applied to predicting meanings and concepts as well as just words.  It’s not too hard to understand how a keyword prediction algorithm might predict the word “thank” is most likely to be followed by “you”.  But what if that same idea was applied to concepts?  What if all the text on the internet was analyzed not just for what words were most likely to follow others, but which ideas or concepts?

For example, if you were to say “My room is dirty”, a language model might be able to predict that sentences relating to the concept of “cleaning my room” are more likely to follow than ones relating to the concept of “diving to the store”.  By combining word as well as concept prediction, language models can, if given a starting prompt, begin to formulate responses that make sense.  They don’t just predict what word to say next, they predict what concept should be communicated next as well, and they do that based on the endless volumes of text data that they’ve been trained on.  That is their strength as well as their weakness.

What does it mean to be smart, or creative?  It comes down to making better-than-average predictions.  Solving problems other people can’t solve, knowing what joke will make people laugh, or how a complex event might impact a brand’s audience.  Language models are better at predicting the most common following concept, not the best following concept.  The same endless dataset that gives them boundless knowledge binds them to a law of averages when it comes to intelligence or creativity.  The sound of thousands of voices singing at a concert is rarely out of tune, but it’s never that good either.  AIs trained on massive datasets are impacted by a similar law of large numbers.  It makes them knowledgeable, but also bland, uninspired, and without any strong underlying point of view.  When it comes to using them to generate insightful, original, and engaging content, this underlying genericness can be a real problem.

It’s not just the event you’re reacting to, it’s also the medium.  Web articles, videos, and social posts are all different, and your reactions to them should be different as well.  Hone your Reactive Marketing skills by developing channel-specific strategies for reacting to video, web, and social content will take your Reactive Content to the next level.

Learn the best way to react to events in different media formats in the next edition of The Content Marketer’s Guide to AI: Different Types of Reactive Content and How To Make The Most of Them