14 days ago, I was honored to speak at MAICON, my fifth opportunity to present at the annual spectacle produced by the Marketing Artificial Intelligence Institute.
As I shared then with the Institute’s Claire Prudhomme, MAICON is a spectacular forum where the most novice and the more experienced of practitioners gather to discuss how AI and machine learning are abruptly reshaping marketing workflows. In 2023, the conference was met by more empathy and a greater understanding of public statistical language models and prompt engineering with speakers such as Ethan Mollick and Cassie Kozyrkov providing wider views of what marketers can expect from automation, centralized intelligence, and machine decision-making, or “reasoning.”
Over the next week, I’m going to breakdown my MAICON 2024 talk with some prescriptive guidance for your business to take control of its future. Instead of transformation happening to you, I’m going to outline some fundamental steps for you to control your transformation.
As we navigate the weekly-evolving landscape of artificial intelligence, it’s crucial to understand the true drivers of AI success. While large language models like ChatGPT garner the majority of our attention, the real power lies in performance data management and specialized AI models tailored to specific business operations and needs. Not enough respect is paid to the limitations of public AI models, the importance of data as the foundation for effective AI, and the promising future of private language models and AI agents in revolutionizing business operations.
In this first of four articles on controlling your future, let’s look at public AI models such as OpenAI’s ChatGPT family and Microsoft’s Copilot.
The Limitations of Public AI Models
Many organizations are rushing to implement AI solutions based on public models like those from OpenAI.
However, this approach has significant limitations that businesses need to be aware of, including:
1. Lack of control over the underlying technology, and inability to custom configure model performance. Your business has a very unique combination of operations, culture, customers, data, products and services, employees, and market competitors. Leveraging technology that is brought to market “to do something for everyone” is a sure way to get lost in a sea of sameness. Remember the days of Twitter “fail whales” and those few times that Facebook went out of service for several hours? Having a single point of failure that is beyond your governance is not the basket in which to put all your eggs.
2. Potential for unexpected changes in model behavior, “hallucinations” and a rising threat from data poisoning of public models. Of key concern with ChatGPT is the high probability of bad actors and other nuisances uploading false memories that weaponize a current model and steal user data as long as it in service. “Garbage in, garbage out” applies to impressive public models, especially as they grow in sophistication and parameters.
3. Related, CTOs rightly have privacy concerns about employees doing bad things with good intentions, sharing sensitive data with public models. Customer-centricity is a competitive advantage, and it is what your customers expect, so safekeeping the precious intelligence shared with your company through everyday experiences and exchanges is paramount.
4. Inability to incorporate proprietary business knowledge. Brand-centricity is what employees, customers, and partners need to trust you and continue doing business with you. Protecting your IP and the “special sauce” that keeps customers returning goes far beyond a refreshed logo, automation, and home or curbside deliveries. You have to respect your secret formulas and earned empathy with employees and customers alike, and you stand to forfeit these qualities if you replace human reasoning (that can train an open source and privately secure model) with a public model or a wrapper product.
5. Limited customization options for specific industry needs.What works to sell hamburgers won’t apply to automotive sales and dentist promotions. Each industry has nuanced operations, regulations, and special situations. To align with all of this, build repeatable processes, and enjoy business continuity, you’ll want to establish business rules and workflows within and around the use of a custom language model. We’ll get deeper into how in part 3/4 of this series.
6. Potential for biased outputs based on training data. Again, you want to be customer/brand-centric, creating champions of anyone who buys from or follows you. If you use a third-party public language model, the data it is trained on may conflict with your order of operations or worse, have outputs that are contrary to employee needs and customer expectations.
Instead of relying solely on these public models, businesses should focus on leveraging their own data and developing custom AI solutions that address their unique challenges and opportunities.
In Part 2/4, I’ll provide a how-to list on unifying your data for training and operating a custom language model. Stay tuned!
