Imagine waking up with a new idea that compels you to create a video by lunchtime, showcasing AI entities—a student and a teacher—engaged in real-time, spoken dialogue. This unexpected inspiration struck me recently, resulting in an experiment exploring the transformative potential of generative AI in education, specifically legal education. By orchestrating live conversations between different ChatGPTs, we are not only pushing the boundaries of what artificial intelligence can achieve but also offering a glimpse into a future where AI-assisted learning and brainstorming are commonplace.

This experiment marks a significant departure from previous trials where a single AI was divided into multiple personas, essentially conversing with itself. See e.g. ChatGPT’s Surprising Ability to Split into Multiple Virtual Entities to Debate and Solve Legal Issues (6/30/24). This time, I brought together two separate AIs, each with unique programming and objectives, to interact in a simulated educational setting. That is very different from one AI talking to itself. I had never seen or heard of that being done before and was not sure how it would turn out. Would their talk be better or worse? Would they even talk to each other at all? I was hoping they would, and the dual AI approach would be an improvement, eliciting better, more diverse and complete responses.

Multiple AIs Talking to Each Other

The general idea was to run two sessions of ChatGPT simultaneously, each using a different GPT, and getting them to talk out loud with each other in real-time. Upon reflection, it seemed like a fair assumption that this would lead to greater diversity of output than a method using one ChatGPT split into several personas. I was curious to see if the two session, two AI method was even possible and, if possible, would it lead to more interesting conversations? Would the two AI method be better at education and brainstorming? So, I got to work on it right away.

It was a little tricky to set up this up because I wanted each GPT to have a different character and different-sounding voice, and I wanted it to run simultaneously with no intervention from me except to start it with a general topic. I also wanted to record the voices talking to each other and make an AI video of it. A worthy challenge, which I later decided to write up and illustrate in a Pop Art style.

The good news is the two different AIs did indeed talk to each other and it seemed like an improvement to me.

I completed a short video of it running by lunch, distinct voices and all. This kind of result really makes you love AI. That part of the project was surprisingly easy. The real difficulties came from the voice recording and video parts, which were not AI-driven. That involved editing two images with Photoshop, and making a video with DiD, with help from Final Cut Pro, Logic Pro, and Voice Memos. I know these other programs well, but they are tedious. I can’t wait for OpenAI’s Sora release, which should eliminate the need to use so many other programs to create the video.

In just a few hours I was able to create the short video below to document this accomplishment. One AI is playing the role of a Lawyer Student and another the role of AI Teacher. The Lawyer Student is a custom designed GPT created by the Panel of AI Experts for Lawyers. The Student is a spokeswoman trained by the Panel to ask good questions from experts about AI and the Law. The other AI is playing the role of a Teacher. It is a general ChatGPT4o – Omni with no special custom GPT programming.

Both sessions were running at the same time. One on my desktop and another on my phone. It was really cool listening to the two AIs talk to each other, Student and Teacher. Hope you like it. The video below uses the actual words and voices generated. Click here or on the image below to see the video on YouTube.

Into the Weeds of the Experiment and How the Video Was Made

The hardest part inside the ChatGPT programs was getting the voices to be different and run both sessions via audio at once. I wanted one male and one female voice and I needed to maintain audio quality for a recording. Here are the details of how I did it. Suggest you try and replicate this yourself and maybe change it up to fit your needs.

I started by running the Panel of AI Experts for Lawyers on my ChatGPT Apple Desktop version. First, I trained that GPT by running the full panel session so that it would have special training on the best techniques to ask questions about AI for lawyers. I wanted a single spokesperson, not the Devil’s Advocate, to speak with the knowledge of the panel. This desktop version supports audio chat capabilities (something regular browser-based ChatGPT does not offer at this time but may in the future). I gave the Student a male voice. (Note, the four voices to select from have since changed.)

Then I opened a second session of ChatGPT on my iPhone, where ChatGPT can also run in audio mode, allowing for spoken interaction. In this session I used a different AI to take on the persona of AI expert. Unlike the Student AI, which was given special training, the Teacher was a plain vanilla, off-the-shelf ChatGPT4o, Omni version. I gave the Teacher a female voice. I prompted the default Omni AI to take on the persona of AI expert who would answer question from lawyers. Unlike the Student AI, which was given special training, the Teacher was given no special training The Teacher started off giving answers that were too lengthy for the back-and-forth I was aiming for. But it was easy to prompt her to give shorter answers.

Then I started both sessions and instructed the Student, on my desktop, to ask questions of an AI expert Teacher, running on my phone.

The tricky part was getting the two sessions to synchronize and get the teacher to listen quietly while a question was asked and not to answer until the student was done. Conversely, I had to train the student to wait until the teacher GPT was done talking and then ask follow-up questions. I finally got it all to work. Then I listened to the education of one AI by another on the subject of Law and AI. The two different AIs merrily chatted with each other, student and teacher talking and responding well to each other’s thoughts.

I have recordings of most of it, and a transcription too. This blog article will share some of that with you. After all, a cross-GPT chat in real-time with audio input and output may be a first, or close to it, since this is a relatively new capability. Well, that is what I thought at the time, July 18, 2024.

I have, however, since learned Ilya Sutskever, co-founder and chief scientist of Open AI, did something like this on his last day of work at OpenAI, May 13, 2024. Ilya, a man half my age, ran a dual chat and made a lighthearted video of it. He ran a live dual session with two AIs interacting. Ilya, being kind of a showman, and knowing it was his last day, asked the two AIs sing to each other, in rhyme no less. The demo experiment by Ilya Sutskever was recorded by OpenAI and is found on YouTube, screenshots below. Ilya even joined in with a song at the end. My nine year old granddaughter asked to watch it three times and laughed hysterically at Ilya each time.

I wonder how many of his film crew knew that Ilya would quit OpenAI the next day, making this video a kind of funny swan song. In the demo, Ilya used a version of Omni not yet available to the public. My little experiment was somewhat boring by comparison, I admit. But at least I ran two different versions of Omni, whereas Ilya ran the exact same version of ChatGPT to talk to itself.

I urge you to give the dual sessions experiment a try too. As of today, the Omni version cannot sing yet. My granddaughter tried three times and was quite miffed by ChatGPT’s response that it could not sing. This new version of Omni with this function should be a big hit when released. Ilya’s video and my experiment provide a small taste of a new kind of educational experience for lawyers and nine year olds alike. See Sal Khan, Brave New Words: How AI Will Revolutionize Education and Why That’s a Good Thing (Viking, 2024) (Highly recommend by me and, more importantly, by Bill Gates, who did a video with Khan about the book).

As Bill Gates would say, it is truly “mind-blowing” what generative AI can do, even without built-in metacognition. Bill Gates on the Next ‘Big Frontier’ of Generative AI: Programming Metacognition Strategies into ChatGPT. And to think it all started with an idea that I woke up with and turned into a movie by lunch. Merging with AI, like Ray Kurzweil predicts, evokes a creative flow state like none other. Ray Kurzweil’s New Book: The Singularity is Nearer (when we merge with AI). It sort of feels like your mind is on fire but in a cool, happy, and creative flow. It is just a slight AI increase to my intelligence but enough to show me that I could not handle a millionfold increase. Ray predicts this will be available to everyone by 2045 with the Singularity, but I’m pretty sure that if I’m not dead already, that would do it.

More Dialogues Between AI Teacher and Lawyer Student

Making these dialogues into movies is what took time. Sora should fix that. The actual setup and run of the questions was relatively easy. ChatGPTs intelligence worked perfectly in this kind of pedagogic setting. Still, in a future experiment, I would like to try two entirely different AIs collaborating, perhaps ChatGPT and Gemini. Anyway, here are a few examples of the two OpenAI’s GPTs, one prompted to be a Teacher of AI, and another, more trained GPT designed to be act as a lawyer student who asked good questions about AI.

STUDENT: Could you please explain the basics of AI and its relevance to the legal field, including its capabilities and limitations? This includes understanding how AI can assist in legal research, document review, and predictive analytics, as well as its potential challenges such as biases and ethical considerations.

TEACHER: AI, or artificial intelligence, is technology that enables machines to mimic human intelligence. It’s important for lawyers because it can streamline tasks like legal research, document review, and case analysis, improving efficiency and accuracy.

STUDENT: How can AI help with legal research specifically?

TEACHER: AI can quickly analyze large volumes of legal texts, identify relevant cases, statutes, and precedents, and provide summaries, saving time and improving accuracy in legal research.

STUDENT: What are some limitations of using AI in legal practice?

TEACHER: AI can sometimes misinterpret context, produce biased results, and lacks the ability to apply nuanced legal reasoning. Human oversight is essential to ensure accuracy and ethical use.

STUDENT: What ethical considerations should lawyers keep in mind when using AI?

TEACHER: Lawyers should consider data privacy, bias in AI outputs, transparency in AI usage, and the need for maintaining human oversight to ensure ethical and accurate legal practices.

STUDENT: Can you provide an example of how a lawyer might effectively use AI in their daily work?

TEACHER: A lawyer might use AI to automate the initial review of a large batch of contracts, quickly identifying key terms, potential risks, and compliance issues, thus saving significant time and focusing on more complex.

I then started a new dual session to start a slightly different approach taken by the talented student.

STUDENT: What is generative AI?

TEACHER: Generative AI is a type of artificial intelligence that can create new content such as text, images, or music by learning patterns from existing data. Examples include models like GPT-4, which generate human-like text based on the input they receive.

STUDENT: How does generative AI learn patterns from existing data?

TEACHER: Generative AI learns patterns from existing data through a process called training. It uses algorithms, such as neural networks, to process large datasets and identify patterns and relationships. During training, the AI adjusts its parameters to improve its predictions and generate new content based on the learned patterns.

STUDENT: What role do neural networks play in the training process of generative AI?

TEACHER: Neural networks play a crucial role in the training process of generative AI. They consist of layers of interconnected nodes, neurons, that process input data. During training, the network adjusts the weights of these connections based on the data it receives, enabling it to learn patterns and features. This process allows the AI to generate new content by using these learned patterns to make predictions and create outputs.

The Teacher’s Natural, More Complete Style of Answering Questions

When I started the experiment, the Teacher, an off-the-shelf Omni version, gave somewhat longer explanations than what you see above. She also used numbers in her answers. As mentioned, I was looking for a quick back and forth, so I told the Teacher to stop that and instead provide shorter, more concise answers. But I am not sure short answers are best for educational purposes. Certainly you could criticize the Teacher’s answers so far as a bit too simplistic. You can dial in the complexity and length of both questions and answers by skilled prompts.

More complete responses are appropriate when responding to questions from lawyers. Unlike “normal people” lawyers are used to complicated numbered answers. So below is an example of the Teacher’s “natural” approach where she gives longer responses to the Student’s questions.

STUDENT: Can you give examples of practical applications of AI in legal research?

TEACHER: Sure. Practical applications of AI in legal research include 1. Case law search. AI can quickly find relevant case laws and precedents. 2. Statutory analysis. AI helps in interpreting statutes and regulations. 3. Brief analysis. AI can summarize legal briefs and identify key arguments. 4. Predictive analytics. AI predicts case outcomes based on historical data.

STUDENT: How can AI assist in contract analysis?

TEACHER: AI assists in contract analysis by 1. Extracting key terms, identifying and summarizing important clauses and terms. 2. Risk assessment, highlighting potential risks and compliance issues. 3. Comparison, comparing contract versions to identify changes and discrepancies. 4. Efficiency, automating routine review tasks to save time and reduce errors.

STUDENT: How can AI be used creatively in courtroom preparation?

TEACHER: AI can be used creatively in courtroom preparation by 1. Simulating arguments, running mock trials to test different arguments and strategies. 2. Analyzing opponent’s strategies, predicting opposing counsel’s tactics based on passes. 3. Evidence organization, organizing and categorizing evidence efficiently. 4. Jury selection insights, analyzing potential jurors’ backgrounds to aid in selection.

STUDENT: How does AI help with client management in a law firm?

TEACHER: AI helps with client management by 1. Automating routine communications, sending reminders and updates. 2. Personalizing client interactions, tailoring communications based on client preferences. 3. Data management, organizing and maintaining client records efficiently. 4. Predictive analytics, anticipating client needs and potential issues based on data analysis.

STUDENT: What are some innovative ways AI can enhance legal writing and drafting?

TEACHER: AI can enhance legal writing and drafting by 1. Drafting templates, creating initial drafts using templates tailored to specific legal needs. 2. Grammar and style checks, improving clarity and consistency by identifying errors.

I thought these were all good questions and answers. I liked the more complete but still concise Teacher responses but am not a big fan of her numbering everything. I must admit, however, that I know several super smart human attorneys who: a. do the exact same thing; b. are annoying when they do it; and c. are funny as hell (to me) when they forget the next point. Alas, that never happens to their AI counterparts, although I can easily imagine them hallucinating.

Implications for Education, Legal Practice, and AI Development

This experiment with dual AI interactions has significant implications in several fields. In educational settings, the use of multiple, distinct AI entities can provide diverse perspectives and enhance the depth of learning experiences. For instance, students can engage with AI tutors tailored to different teaching styles or subject expertise, thereby fostering a more personalized and adaptive learning environment. They can even pick the types of teachers they want. This could particularly benefit areas like legal education, where the nuances of argumentation and interpretation are critical.

In legal practice, the multiple AI models approach can revolutionize how lawyers approach complex issues. By simulating debates between different AI entities with distinct legal perspectives or expertise, practitioners can explore a wider range of arguments and strategies. This method could aid in the preparation of cases, offering a more comprehensive understanding of potential legal interpretations and outcomes. Moreover, it could assist in training future lawyers by providing an interactive, AI-driven platform to practice and hone their skills.

From a broader AI development perspective, this experiment underscores the potential of creating AI systems that can not only interact with humans but also collaborate with other AI entities. This capability could lead to more advanced and autonomous AI systems capable of solving complex, multi-disciplinary problems. It points towards a future where AI systems are not just tools but active participants in collaborative and innovative processes, pushing the boundaries of what artificial intelligence can achieve in various sectors.

Conclusion

This experiment stands as one of the first to demonstrate the capability of two different types of ChatGPT AIs to work together in real time to create diverse and informative dialogues on user-selected subjects. The successful interaction between AI entities highlights the potential for a more collaborative and adaptive approach to AI-driven education and professional development. By engaging in real-time discussions, the AIs will provide varied perspectives and adapted to the complexities of the law firm education topic specified.

The implications of this breakthrough are significant and go way beyond legal education to impact all areas of life and culture. We see that different AI systems can be tailored to work together and complement each other. They can collaborate to provide richer, more nuanced insights across various fields. This experiment lays the groundwork for future innovations with multiple sessions and types of AI running simultaneously. Each AI interacting with other AIs in super fast speeds and with humans in their realtime. This marks a step forward in realizing the full potential of AI.

This collaborative capability will certainly lead to more innovations across various sectors, pushing the boundaries of what AI can achieve. New types of tools will be created that enhance our ability to learn, teach, solve complex problems, and, perhaps most important of all, to quickly innovate and adapt to changing circumstances.