Closing the AI Value Gap Part 4: Chat GPT & Adaptive Analytics in the Wild
The 4th installment of the AI Value Gap series covers how Chat-GPT represents the first consumer-facing Adaptive AI system
In previous posts, I’ve discussed the hype surrounding AI and how it needs to evolve into a more adaptive form in order to truly deliver the value that has been promised. While AI has made incredible progress in recent years, it’s essential that it continues to adapt and grow to meet the ever-changing demands of our world. Today, we’ll be exploring how the development of Chat-GPT represents the first public-facing instance of adaptive AI analytics, and take a look at the ways Open AI is following the Adaptive Analytics Roadmap as they roll out new features day by day.
To quickly summarize without making you read all 3 prior articles, here is a quick conversation with Chat-GPT, using the new GPT-4 model:
Chat-GPT is a powerful AI language model developed by OpenAI, capable of responding to user inputs and engaging in interactive conversations. The conversation above, in addition to providing an explanation of Adaptive Analytics, is a great example of Adaptive Analytics in action, with the model adapting to the my desire to use specific names for each loop. This technology embodies the AI Co-creation approach that is central to Adaptive Analytics — one of the crucial elements needed to close the AI value gap.
Chat-GPT covers all elements of Jobs to Be Done
Open AI has managed to cover all the components of the Jobs to Be Done (JTBD) framework introduced by Clayton Christensen in “Competing against Luck” and further expanded in other literature.
Functional Needs: Enhanced Utility and Versatility
Chat-GPT addresses the functional needs of knowledge workers by offering immense utility and versatility. It enables users to experience a dramatic acceleration in their ability to get the job done by providing rapid, on-demand information and insights. The interactive AI model also allows users to experiment with multiple approaches, giving them the flexibility to explore diverse solutions and find the best path forward. In an MIT study, these approaches provided a 37% boost to initial performance, and a 20% quality boost.
Personal Needs: Improved Outcomes and Productivity
The personal needs of users are met through the improved outcomes they experience when utilizing Chat-GPT. People using the tool have been found to be more productive than their peers who do not, creating a competitive edge that enhances their performance. In the MIT study, this productivity boost is especially true for those who were part of the least productive group before adopting the technology.
Additionally, Chat-GPT is an excellent tool for task-based concept learning. By enabling users to learn through interaction and iteration, it fosters a deeper understanding of the subjects they need to understand right now and accelerates the learning process with Just-in-time learning.
Social Needs: Riding the Hype Wave
The social needs of users are met by the hype surrounding Chat-GPT, which has made it a topic of daily conversation across numerous industries. The buzz around this AI tool has been fueled by its regular presence in the news, making it an exciting and relevant topic that users want to be a part of. As a result, Chat-GPT has become a prominent talking point, and its adoption signals a level of tech-savviness and forward-thinking that enhances the social status of its users.
Chat-GPT: the 4 Feedback Loops of Adaptive Analytics
Because the tool hits all of the components of people’s JTBD, it has been adopted rapidly. The pace of the rapid adoption has been influenced by the execution of the OpenAI team on the feedback loops of Adaptive Analytics (primarily the Interactive and Automated loops, with some expansion into Analytical). In this section, we’ll look at each loop and how it is implemented today, or how I believe it will be implemented over time.
The Interactive Loop: A Chatbot Interface
The interactive loop, as we’ve discussed in previous posts, is essential for creating an AI system that can learn from human interaction and adapt accordingly. In the case of Chat-GPT, this loop is implemented through its ability to engage in conversation and respond to users’ input. By providing human-like responses and learning from these interactions, Chat-GPT provides the back-and-forth interactions that are essential for the co-creation of results between the AI and a human collaborator.
In order to build this interactive loop, the initial experience for the human has to be acceptable. The AI research community calls this the Alignment problem. One of the key aspects of all of the Open AI GPT models is the vast amount of data they are built upon. While the specific details of the dataset used for GPT-4 remain undisclosed, it’s safe to assume that it encompasses a substantial portion of human published knowledge, potentially including books, articles, and other forms of written content in addition to the full internet. This extensive dataset enables the model to provide highly accurate and relevant initial responses, encouraging users to interact and collaborate with the AI system. The ability to build on such an impressive foundation of human knowledge allows Chat-GPT to adapt its responses to a pretty much any topic and/or context.
By jumpstarting with an extensive “database” of human knowledge from pre-training, the interactive loop is possible. The Chat-GPT interface builds on this approach by learning the context of the user’s asks during each chat session. The interface enables the user to change the prompt if the result was not what they were looking for, or to provide clarifying information to improve the responses. As users continue to interact with the system, it learns from these engagements and refines its understanding, further improving the alignment of its responses to the user’s desires.
The Automated Loop: Leveraging Massive Scale
One of the critical components of Adaptive Analytics is the automated loop, which enables retraining and improvement of AI models based on user interactions across the entire system. Chat-GPT takes advantage of this loop using Reinforcement Learning from Human Feedback (RLHF) to continually improve its performance.
RLHF is a technique that trains AI models by leveraging feedback from humans in response to the AI’s actions. That feedback trains another model that acts as the reward in a Reinforcment Learning model. The RL model then iterates on the prompts generated by the original language model to maximize the likelihood that a human would provide good feedback on the response. This approach allows the model to learn from real-world user interactions, enabling it to adapt and improve over time.
Chat-GPT first launched as a Research Preview that was explicitly set up to gain real-world feedback to train the model. Each time a user changes a prompt, votes on a response, or continues a chat, a signal is captured that helps determine the effectiveness of the results that were previously generated.
With the rapid adoption of Chat-GPT by over 100 million users, the model has access to a vast array of data from nearly every conceivable type of task. OpenAI has long been at the forefront of developing scaled machine learning processes to automate model training and roll-out, enabling them to take advantage of this scale. This continuous stream of feedback is invaluable for refining the model’s performance, as it allows Open AI to train both the reward model and RL model in parallel, leading to a more accurate and adaptive AI system.
In fact, the rapid release cycle for GPT-4 can likely be attributed to the success of Chat-GPT and the scale it has achieved. As more users interact with the AI, the automated loop is continuously fed with new information, allowing the model to evolve and become more effective over short time periods.
The Analytical Loop: Iterating on the Interface
The analytical loop aims to refine and enhance AI systems by making informed changes to the captured data or the way recommendations are presented to users.
In the context of Chat-GPT, the analytical loop will involve the following key aspects:
Application-Specific Fine-Tuning: As AI models like Chat-GPT continue to gain traction, the need for fine-tuning these models for specific applications becomes increasingly important. By tailoring AI models to cater to specific industries, use cases, or problems, we can harness their full potential and achieve better results. Labeling plays a crucial role in training AI models, allowing them to learn from vast datasets and make accurate predictions. Integrating adaptive AI analytics with custom-tailored labeling processes will further enhance the model’s ability to understand and interpret complex data in different domains, ultimately leading to better decision-making and optimization.
UI Overlays for Specific Industries and Problems: To maximize the effectiveness of adaptive AI analytics in various sectors, it’s essential to create UI overlays tailored to specific industries and problems. For instance, specialized overlays could be developed for sectors like:
- Marketing, which was (as usual) the first to move, with Persado gathering more than $100M of funding to combine generative AI with A/B testing to dynamically optimize marketing copy at the individual level.
- Pharmaceutical Drug Discovery, where LLMs can analyze the vast amount of scientific literature to identify potential new targets for drug discovery by extracting relationships between genes, proteins, and diseases. They can then predict the binding affinity of small molecules to a target protein, helping to narrow down the list of potential drug candidates. However, presenting the results in an appropriate manner, combining the text-based output with visuals, fine tuning the models and prompt templates, and other UI advances are likely to be required to unlock this space.
- Conversational programming allows developers to interact with AI models in a more intuitive and human-like manner, making it easier to refine and optimize the system based on user input and feedback. The current chatbot web interface doesn’t provide a truly useful experience for developing new applications, although it does better than tools like Github CoPilot or Amazon CodeWhisperer in generating tutorials and explanations to help the developer learn why a specific block of code was generated. I’m thinking that a new UI approach with a prompt for higher level tasks and the familiar comment-to-complete approach in each code file will be needed to unlock the true power.
For an example of the current state of conversational programming, check out the following (1 sentance of prompts generates a full end-to-end tutorial):
These Experience Design advances will allow AI models to better address the unique challenges faced in each domain.
Additionally, one of the most intriguing aspects of Large Language Models is the ability to reframe complex data structures as language. This includes genetic code, protein structures, and other intricate patterns that were once beyond the reach of traditional AI models. By leveraging the same techniques used to create LLMs, we can expand the applications of adaptive AI and unlock new frontiers of knowledge.
Expansion Loop: Building end-to-end Applications
The expansion loop is where the true potential of large language models (LLMs) like GPT-4 is unlocked, heralding a new era of innovation and possibilities. By combining LLMs with other sources of computation or knowledge, developers can create powerful applications that surpass the limitations of individual AI systems.
GPT-4, as a multi-modal model, also brings images into the mix, enhancing its capabilities and enabling a more comprehensive understanding of the world. With the integration of images, GPT-4 can better interpret and generate complex, contextually rich content across various domains. As of this writing, this functionality has yet to be released widely.
The Chat-GPT plug-in architecture represents another significant advancement in the field, enabling a dramatic expansion in capabilities. These plug-ins, designed with safety as a core principle, allow Chat-GPT to access up-to-date information, run computations, and use third-party services. This evolution transforms AI systems from passive tools to proactive, intelligent assistants like Jarvis from Iron Man, a far cry from today’s more limited voice assistants like Alexa.
The Langchain library acts as the glue that binds these various elements together, providing a multi-step processing and reasoning engine to solve complex problems. By integrating LLMs with other computational resources and knowledge sources, Langchain facilitates the development of sophisticated applications that can tackle challenges previously considered beyond the reach of AI. This is accomplished by having the input prompt for later steps in the chain generated by the earlier steps. Microsoft Research has already started this progress with their MM-REACT paper and library.
Looking forward
Together, these advancements mark the beginning of a new wave in AI development that harnesses the combined knowledge of humanity. As LLMs evolve and become more sophisticated, they will enable us to solve problems at a scale never before imagined. This synergy of AI systems, computational resources, and human knowledge will propel us into a future where technology could serve as a powerful force for good, amplifying our capabilities and empowering us to overcome the greatest challenges of our time. This rapid advancement is not without its risks, such as the ability to inject prompts to circumvent safegaurds, the often-noted proneness to generate non-factual responses, and the unclear legal landscape.
In this exciting new era, we stand at the precipice of limitless possibilities, ready to embrace a world where AI and human ingenuity work together to unlock our full potential and create a brighter, more prosperous future for all.