Image creation process - experiment conducted by me - the AI identified icons related to the history of Marketing
Those who have set aside prejudice and experimented with image generation AI know that the subjective impact is instantaneous. Immediately upon experiencing something so unprecedented, you begin to assimilate and question how a machine could have created that drawing, image, or video based on what you imagined when describing it. Capturing linguistic nuances, making adjustments often aligned with the textual tone of what was demanded of it. Of course, some bizarre results also appear, and then you start to question the ethical and subjective impacts, the often unconscious biases. It is certain that these problems are being quickly addressed, in order to foster increasing adoption in the market. One of the most famous programs, MIdjourney, is making significant strides and releasing increasingly enhanced versions, having even launched a prototype for video generation just last week, a functionality that Google also intends to release with Imagen.
All these feelings that arise during experiments, but also taking into consideration the facts, such as the inherent exponentiality of this type of technology (unsupervised improvements), lead me to conclude that it is time to consider the "productization" of these technologies. For marketing, the possibilities are enormous, precisely because of the subjective capture that such layers of personalization, inherent to generative AI, bring about, given that their own training has resulted from human interactions on the internet over the past decades. My bet is that once image-generating AIs are condensed into conversational products, where and when imagistic artifacts increasingly personalize conversations, we will truly enter the era of conversational AIs, not just a cold, generalist ChatGPT with little relevance to specific users. Powerful and dangerous at the same time, as I have mentioned before, due to the potential for alienation. But after all, isn't marketing about that? About creating identifications?
Conversations that generate greater customer identification
For those who have been involved in Discovery for chatbots, they know very well that people dislike talking to machines. They openly complain, curse, and report being poorly served, that the conversation lacks identification. And for everyone who has delved into the discovery process, they also know that greater personalization is needed.
Now generative AI is avaiable. But we need to go beyond ChatGPT, which is shallow, non-factual, and generalist (as it should be) given its purpose to serve a broad audience. But as I have explained in other posts, it is possible to personalize a language model by incorporating more data through embedding techniques. And what do corporations have that is most valuable today? Data about their customers: past conversations, incident tickets, sociodemographic data about each of them, aspirations reported in persona mapping, and so on. Why not use this data as a starting point and create conversational products that adapt to the customer who is talking at that moment? We already have the tools for this. Nowadays, an orchestrator can recognize the customer in a matter of seconds by incorporating CRM data into a conversational chatbot plugged into a language model using the libraries I mentioned in the last post.
Moving away from the technical solution for a moment, imagine the marketing potential of conversations that are more "human-like," combined with a meaning that comes precisely from what is known about that specific customer or prospect. It is a possible solution to the problems that product teams have been facing for years. You see, with the emerging technologies, we will move away from chatbots that simply say, "Oh, I found your data, your invoice is outstanding," to conversations that subliminally understand if the customer has actually come to pay their bill or, for example, if they want to resolve another issue first and, above all, be creatively received regarding the resolution of that problem. This is something that a good language model (even GPT-3.5) associated with additional training will outperform the current chatbots provided by established companies and products, such as IBM's Watson.
Where to start? 🥸
The first step is to organize the data (an old corporate dilemma, right?) that you believe is necessary to personalize your conversational product based on a language model (LLM). If it's a customer service product, for example, past conversations and customer data can be a good starting point. If it's a sales product, data about your catalog, sales strategy, and database of your target audience can make your LLM personalized and equipped with the necessary information to better persuade your prospects. Once the data is organized, it's necessary to move on to an LLM/LangChain/Pinecone framework or ready-mate cloud structures that allow you to personalize and deploy this product. Then establish your service or conversion criteria and monitor in a restricted beta how the LLMs have been performing in achieving these objectives. Improvements can be made through retraining using embeddings, and enhancing the conversational objective can be achieved through prompt engineering and ultimately fine-tuning to adjust the tone of the conversation, response lengths, and so on.
It's essential to start with this beta mindset precisely because it's something new, but also to adapt the adjacent legacy structures during this transition. For example, the data saved from these conversations can be stored differently from traditional chatbots, using conversational event logs that provide feedback to the model rather than transaction records, for instance. Well, the path is long, but the journey has already begun. Building software is changing, and it's time to give more autonomy to AI and act ethically in this transition, observing the results and making new bets—it's a matter of survival for businesses. It's time to innovate. Count on me in these experiments. Let's have a dialogue together.
Bonus 😎: Midjourney has just released the Zoom-out functionality that allows you to add predictive elements to the original image, creating an "expanded" representation of the original creation. In a lighthearted experiment, I reconstructed a scene that people familiar with BH (Belo Horizonte) in Brazil would recognize, a gathering at the Market. The AI managed to incorporate elements from Mercado Novo, most likely because it was trained on old images of that location and the Mercado Municipal. Notice that the AI struggles with drawing hands—the mystery behind this can be found here. 😂. Until next time and please subscribe!
Great articles, congrats :)