musings
Welcome to the conversations we have as we work together at wks.
Musings range from imaginative sketches to provocative assertions. Collectively, they represent what-ifs and why-nots. It's what we are seeing and hearing in Generative AI right now.
Is it me or is search feeling long in the tooth? Statistica estimates that search advertising will grow at a CAGR of 8.3%, reaching $417 billion globally over the next four years. That looks pretty healthy.
But I dunno.
It’s not just that I look at my search results and wonder why it can’t find anything-- I know the info is out there. More concerning is the rapid growth in AI content. What happens if there’s so much AI content out there that the search results become basically useless? I thought it was just me until I heard Paul Roetzer utter the same on his AI marketing podcast this week (check him out, he's good).
If AI were to disrupt search, it could happen very quickly. We’re just now entering multi-modal phase of AI. Current large “language” models LLMs use architectures that can already work with images, video and sound when they are trained to do so. So, the AI content is going to become much more immersive. Similarly, AI in search presents challenges for the marketer. AI can simulate human browsing patterns so well So, how do I know if impressions and clicks on my dashboard are from a human at all? This creates a quality problem in terms of both content and audience.
This morning, Google launched its new LLM Gemini and CEO Sundar Pichai described Gemini in search as part of their "fundamental vision is that people come to search to experience the richness and diversity of the web and the content ecosystem." His examples centered on summarizing information and conducting multimodal searches. No mention was made of the growing and systemic problem of AI content and user authenticity. Even if Google can easily integrate AI into search and address these problems, AI inference is 50% more expensive to deliver than web search (Dylan Patel has a great analysis of these economics at Seminanalysis.com Feb 2023). This would cut Google's search income in half and it doesn't include the costs to train those models.
For now, marketers spend on media in order to influence **human** behavior. Search marketing is a huge chunk of that spend right now. So, any significant change to search effectiveness would have large consequences on the media landscape, not to mention the entire Google ecosystem. For consumers, there are no switching costs to abandoning Google. As a marketer, though, I'd want a backup plan for how to shift my spend, if necessary.
That's why I'm mulling the concept of “verifiable” media, being media with humans at either end. These are places like TV or Amazon where marketers have reasonable assurance that the audience is human and where the audience can be pretty sure that the content is sourced by humans. Hollywood is a vanguard in providing AI protections human creators. This could create some insulation from this trend and provide further incentives for a system that more resistant to the problems of audience and content quality. In verifiable media, AI can still make the experience better, just as in Sundar's vision, through content findability, presentation, modality-play, augmentation and even AI creation, but within the context of a human to human pipeline. Social media and creator platforms can be verified if they can authenticate user and content. We'll see. Alternatively, we could all start paying for search.
Modern day Generative AI applications are build on a statistical process that takes in text and represents it as numbers in such a way that similar words have similar numbers. This is done as a way of preserving some of the meaning of the underlying document. From there, all sorts of cool information processing can happen: text generation, information retrieval, classification and recommender systems, etc.
What started as text is spreading to other "modes" as they're called in machine learning. So now, this same approach work with images, sound, and video as well. Similar words, pictures, sound and video can be coded to have similar numbers, assuming you've got a good enough dataset to train the model. OpenAI claims it does. You can type the word fluffy dog and pull up a picture of one. Have the model write a Ken Burns-style narrative for a silent video and suggest accompanying music? Sure, no problem. It's here and this post will soon feel outdated as as result.
If we step back, what's the idea behind this technique? The idea is that the meaning something carries has a lot to do with the context the thing is in. The American linguist Zellig Harris noticed that word existing in similar contexts tend to have similar meaning. Soon after, the Scottish linguist John Firth arrived at his idea theory of relatedness: often summarized as: "you can know a lot about a thing by the company it keeps." It was formalized mathmatically as the Distributional Hypothesis.
Now, this theory is being applied to text to understand the relationship of words to each other. You can model these relationships using Machine Learning and apply it using Generative AI. You can take any dataset of words sufficient size to create probabilities and make predictions with them. But context in writing is about more than just the position of words to each other. It's about where the word appears in the document. A word in the headline often carries more meaning than one in the footnote.
Or not. The footnotes of an infomercial are quite interesting, actually. Tom Waits once crooned that "the big print giveth and the small print taketh away."
Meaning could also depends on what type of document we are dealing with (novel versus medical report). It could matter who wrote it. And on and on.
So, context is about much more than the words. Marketers know that you can drive human behavior by when you present the right message at the right time at the right place. Brand marketers go to great lengths to place their messages in contexts where other things might lend their brand some equity. Meanwhile, performance marketers optimize a context called media. Direct marketers made a science of aligning the message, format and offer.
Imagine if you could train a model with time and place data along with the message. What if you tied generative models with some data about performance and asked them to learn? You can know a lot about a thing from the company it keeps. Turns out we already do and fully expect the next generation of AI models will utilize this information in some interesting ways.
We work with a lot of creators to make stuff. As AI demonstrates growing capabilities, through, we hear from some creators that using AI can feel a bit like cheating. It's not just the ethical issue of attribution, but also ego, in a good sense. Many creators are driven by ego just as much as recognition. It's yours because you spent time creating it. Sort of, the you-get-out-what-you-put-in mantra in effect for many of them. What does that mean for a creator when Generative AI can deliver so much, and almost seem "creative"? Does the nature of creativity change or do we change what we choose to create?
This may hold true in a wider context for all of us. Tasks are continually made easier by technology. How many of us would continue to play golf if every shot were a hole-in-one for everyone? Challenge and mastery drives a lot of people. Winning motivates a few more. Yet each of these is relative. Technology can always come along and make it easy to do the next thing, thus leveling the playing field for all.
In the end, we decide what is valuable. As some things become easy, Perhaps, our definition of value will shift as a society. Different things become important to us and creators will have new and different ways to meet their enduring need to create.
AI's Premise: Every aspect of intelligence can be simulated? upcoming
When humans design, they are using creativity to make something. When Gen AI models produce, they are using statistics to predict something. This distinction will become immaterial as emergent forms of behavior, but for now it's important to keep these ideas distinct.
Background source: Mary Ellen Matthews, NBC