Going Beyond GenAI

You’ve likely tried one of the mainstream Generative AI tools that have exploded to popularity over the past two years – ChatGPT, CoPilot, Gemini, Perplexity or any of the various tools from GenAI-based startups.

Even from a brief use of these tools, you may have noticed they don’t actually do that much (And that's not to get a rise out of anyone). Maybe you tested one out to generate a recipe from the measly bits you had left in your pantry or perhaps you use one everyday at work. Regardless of your task, you may have quickly noticed a repetitive pattern in how responses were strung together, or on a more concerning note, realized the output was false or biased.

The Basics of Generative AI

So what is Generative AI? 

If we are going to go for a textbook definition, generative AI is a class of artificial intelligence models that are trained to create new data, learning to generate more objects (images, text, audio) that look like the data it was trained on, creating rather than making a prediction about a dataset. Essentially, a generative model takes learned from inputted examples to create something new based on those examples. 

If we want to add a bit of color to that definition, generative AI is similar to the process a chef follows when baking a cake. An overused example? Yes; although it's an oldie, but a goodie for a reason.

First, our chef needs to collect their ingredients, depending on the type of sweet treat that could include flour, sugar, molasses, and butter. These ingredients are akin to the raw data that the AI uses to learn and generate something new, that could include images, text, audio, or video. Next, they need to find a recipe, which is analogous to the model or algorithm. The recipe tells our chef the steps and how the ingredients should be mixed to bake the best tasting treat, similarly the machine learning algorithm or neural network learns patterns from the data and knows how to combine these patterns to generate new outputs. Once they have their ingredients and recipe ready, they mix them together, and bake the cake in the oven. This is like the training process in AI, as the model processes the data and adjusts itself based on patterns it learns, just like how our chef adjusts ingredients or timing when baking a cake. After the cake is baked, behold the finished product, ready to be enjoyed! Similarly, once the AI has been trained, it can now generate new content—whether it’s a new image, text, video, or music.

Hmm wait, the chef tasted the treat and it needs more sweetness, maybe a bit of icing or a sugar topping. They now know they need to adjust the recipe slightly for next time. Similarly, you can fine-tune the AI model for more accurate results.

Uses of Generative AI

Moving away from baking or generalist use examples for a moment; since the 1960’s industries have begun to be influenced by computer science and artificial intelligence. We will take a quick example for art, to illustrate the proven utility in content-driven industries such as entertainment, advertising, and creative arts for years.

Yet, in 2023, there was a shift, when PSEUDOMNESIA: The Electrician was selected as the winner in the creative photo category of the Sony World Photography Awards. The judges were unaware it was AI-generated. Boris Eldagsen created the “photo” utilising  DALL-E 2, leveraging text prompts and inpainting and outpainting to prompt the creation of the image. This piece has quickly become one of the hallmarks of GenAI being indistinguishable from human content, and further illustrates how GenAI is revolutionizing individuals’ creative processes and design. 

Beyond creative pursuits, the increasing availability of Generative AI applications has led to its adoption and implementation across industries, from financial services to healthcare. As its capabilities continue to expand, organizations in various sectors are adopting AI technology to drive innovation, streamline operations, and enhance decision-making. 

Let’s look closer at AI adoption in healthcare and diagnosis services. Generative AI is being explored for a variety of use cases, from improving diagnostic accuracy to augmenting procedures.

From the range of potential applications above, now imagine: 

You have been dealing with some health concerns over the past month and have finally secured an appointment with your general practitioner to discuss the potential causes of your symptoms. Unfortunately, your regular provider had an emergency and you're now seeing a new physician who is unfamiliar with your medical history. With the help of Generative AI, this new physician can record your symptoms and use an advanced diagnostic tool to gain additional insights. This tool may help you feel more comfortable during the consultation and give you confidence that the correct next steps toward your diagnosis are being taken. 

Generative AI offers an opportunity to equip individuals to accelerate processes across industries; however, this potential comes with a set of limitations.

Limitations

While current generative models have become increasingly sophisticated, they still suffer from limitations, specifically lack of critical thinking and the ability to output novel ideas.

[Despite] impressive output, generative AI doesn't have a coherent understanding of the world. Researchers show that even the best-performing large language models don't form a true model of the world and its rules, and can thus fail unexpectedly on similar tasks, say researchers at MIT, echoing the concerns of many individuals looking to implement generative AI into their personal and professional activities. 

Furthermore, one of the biggest concerns stemming from GenAI’s lack of understanding are hallucinations. Hallucinations are when invented outputs sound/appear plausible but are imaginary, specifically affecting LLMs. Hallucinations are a critical factor holding back wider adoption of chatbot applications such as ChatGPT or Gemini due to unreliable and potentially dangerous repercussions. Taking the patient/physician example from earlier, what if the diagnostic tool aiding in decision-making misdiagnoses a patient's symptoms, potentially leading to the oversight of a life-threatening health issue like cancer? 

Lack of understanding and propensity to hallucinate drastically limits the utility of generative AI in environments that are at the will of real-time changing factors.

Key Differences Between World Models and Generative AI

The key differences between world models and generative AI can be distilled into purpose, approach, and adaptability:

Purpose: As we have discussed above, generative models focus on content creation, whether images, text, audio, while world models focus on understanding and interacting with the environment to predict future states. Generative models have a quite limited capability to predict, forcing them to only be able to rely on historical data to create loose inferences on what will happen long-term, while world models can pull data from a myriad of sources in real-time to constantly refine predictions on what will happen next. 

Approach: World models mimic human decision making processes with a cognitive approach. Through simulating mental representation and reasoning, a world model has inherently a more sophisticated sense of understanding to produce more well-informed outputs. 

Adaptability: Generative models tend to be static, due to the methods of their training and refinement, compared to world models. As world models were originally developed for video generation, gaming, and virtual reality applications, they are built to be adaptable and reactive to real-time changes.

These differences illustrate why at Ergodic we are focusing on building novel Enterprise World Models to offer capability beyond Generative AI. 

Ergodic’s Difference

At Ergodic, we are not building a Generative AI tool, but rather a platform to tackle complex business decision-making, stemming from the belief an enterprise AI agent framework should be based on a world model – a structured representation of the key elements, interactions, and dynamics within and around the business. Through our Enterprise World Models, we enable you to understand the richness in all your data to find optimal paths through complex problems, going beyond current platforms by offering diagnostic, prescriptive, and cognitive analytics.

But why? We think individuals' needs surpass the abilities of current generative AI platforms like Perplexity, OpenAI, and Gemini, which led us to focus on developing more active AI: a system designed not just to generate responses, but to engage as unbiased, knowledgeable collaborators. Predicated on the ability to understand context and engage in complex interactive reasoning, our Enterprise World Models focus on solutions that are more effective long-term and resilient to unexpected adverse-events. By doing so, we are going beyond generative AI to empower  individuals to explore deeper, understanding “why”-based questions, rather than settling for surface-level recommendations 

1. Zewe, A. (2023). Explained: Generative AI. [online] MIT News | Massachusetts Institute of Technology. Available at: https://news.mit.edu/2023/explained-generative-ai-1109.

2. ‌Zewe, A. (2024). Despite its impressive output, generative AI doesn’t have a coherent understanding of the world. [online] MIT News | Massachusetts Institute of Technology. Available at: https://news.mit.edu/2024/generative-ai-lacks-coherent-world-understanding-1105#:~:text=of%20the%20world-,Despite%20its%20impressive%20output%2C%20generative%20AI%20doesn't%20have%20a,fail%20unexpectedly%20on%20similar%20task.

3. ‌McKinsey & Company. (2023). What is generative AI? [online] Available at: https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-generative-ai.

4. ‌Singla, A., Sukharevsky, A., Yee, L. and Chui, M. (2024). The state of AI in early 2024: Gen AI adoption spikes and starts to generate value. [online] McKinsey & Company. Available at: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai. 

5. NVIDIA. (2014). What is Generative AI? | NVIDIA. [online] Available at: https://www.nvidia.com/en-gb/glossary/generative-ai/.

6. ‌Gozalo-Brizuela, R. and Garrido-Merchán, E. (n.d.). ChatGPT is not all you need. A State of the Art Review of large Generative AI models. [online] Available at: https://arxiv.org/pdf/2301.04655.

7. ‌Katya Cengel (2024). The First A.I.-Generated Art Dates Back to the 1970s. [online] Smithsonian Magazine. Available at: https://www.smithsonianmag.com/innovation/first-ai-generated-art-dates-back-to-1970s-180983700/#:~:text=In%20the%20late%201960s%20he,the%20University%20of%20California%2C%20Berkeley.

8. ‌Parshall, A. (2023). How This AI Image Won a Major Photography Competition. [online] Scientific American. Available at: https://www.scientificamerican.com/article/how-my-ai-image-won-a-major-photography-competition/.

9. ‌Topol, E. (2023). Generative A.I. and the New Medical Generalist. [online] Substack.com. Available at: https://erictopol.substack.com/p/generative-ai-and-the-new-medical.

10. ‌Ox.ac.uk. (2024). Major research into ‘hallucinating’ generative models advances reliability of artificial intelligence | University of Oxford. [online] Available at: https://www.ox.ac.uk/news/2024-06-20-major-research-hallucinating-generative-models-advances-reliability-artificial.

11. PrajwalKankate (2023). What is GenAI - PrajwalKankate - Medium. [online] Medium. Available at: https://medium.com/@prajwalkankate/what-is-genai-7e10f008e749