We weren’t around when Socrates roamed Athens, yet we know that he feared the written word itself would weaken human memory. As a result, we now rely exclusively on the notes taken by those around him to know his thoughts. Imagine his impressions about breakthroughs like the printing press. Or the World Wide Web. Or… AI.
Team Worktank has been in video production long enough to witness plenty of new and impactful technologies to advance our work, and we bring a similar curiosity to AI. What’s different is the sheer velocity of updates coming at us where we’re learning about a new AI production or post-production tool every day. While we adapt to changes, we also dwell on the impact to our industry – specifically, the fear of how AI will impact livelihoods and creative work.
Before we get in our heads too much, it’s important to take a step back and unpack AI and understand a technologist’s point-of-view on where AI will grow over the next few years. For this I reached out to an old pal and tech executive/software engineer at enterprise level organizations Bojan Beran who leverages AI in his workflow and builds real world products using AI.
Here are excerpts from a conversation that lasted over two hours but could have easily gone longer.
Explain AI as if I’m 11.
AI analyzes enormous amounts of data and creates statistics around that data. It can then formulate “answers” to “questions” posed to it about that data. AI is not “actual” intelligence. Although their responses are comprehensive and often do give right answers, the models do not actually possess intelligence to understand the difference between a right and wrong answer. The technology works this way: sentences are broken up into what are called tokens. Tokens are just groupings of characters, but not full words. The model is then trained over and over to be a particularly good predictor of what tokens should follow next. This means when you ask it a question it doesn’t understand your question. Rather, from your question, it returns tokens that most likely follow the pattern.
This leads to particularly good generators of legible sentences, but with no underlying understanding of facts, physics, truth. The raw models are good at generating text but “hallucinate” a lot, which means responses that sound correct, but are not based in fact.
The public models that we use are heavily altered Large Language Model’s (LLMs) where a good amount of specific training is provided to return better and more relevant answers.
The biggest misunderstanding of the next generation of AI is that it is so close to how we communicate, people misunderstand and think that it is smart. It is just an extremely good data processor.
What are its strengths and weaknesses?
The strength is the sheer volume of work it can do very quickly. Its weakness is that it doesn’t “know” anything. It just knows what is likely based on the training it has received. An example is when you ask it to illustrate a hand, the results are quite comical to us. Because we just know what a hand is supposed to look like. But all the AI sees in the millions of images it consumes are hands in various positions from different angles with no delineation of the fact that the average hand has five fingers. A similar example is the often-comedic exchange we see when we use auto-translate and the results to actual native speakers are hilarious if not scandalous.
What do humans offer over AI currently?
A few things. We intuitively know some things to be true. We know when you drop a ball on Earth it falls to the ground. We know what a hand should typically look like, and we can right away spot one that doesn’t. We know how to relate to other human beings, we have empathy. And most importantly we can come up with completely unique and novel ways of viewing things or of spinning a story that engages. We are both introspective and can be held accountable. If a human messes up a task, we can talk to that person and understand why they made this mistake, mentor, teach or fire the person. AI is a black box. When it messes something up it cannot tell us why it generated what it generated, at least not in any way that would be useful to us as users of it. We also cannot easily hold it accountable. If it messes something up, we cannot fire one “agent”. We either need to hope to fix the model that interacts with every customer/employee/task or hope that the people training the model can fix it centrally.
Next we’ll dive into all the current and future AI apps making a huge impact – for better – in our creative work. You can learn about executive producer John McDonald’s weeklong experience with Microsoft Copilot and how AI created space and time in his work and creative.
