OpenAI ChatGPT 4: Influence, Nudge, Direct, Tune | Turn the Lens Ep28

228OpenAI ChatGPT 4: Influence, Nudge, Direct, Tune | Turn the Lens podcast with Jeff Frick Ep28 

https://www.youtube.com/watch?v=CzVbXGFLtJY&list=PLZURvMqWbYjk4hbmcR46tNDdXQlrVZgEn 

‍

Cold Open:
Thank you, Jeff. 
It’s good to be here on Turn the Lens. 
I’m excited to be your first AI guest and eager to see where this conversation goes. 
What’s the first question you have for me? 

‍

Jeff Frick:
Hey, welcome back, everybody, for a special episode of Turn the Lens. I want to give a shout out right out of the gate to Yadin Porter de León. He used to be at VMware. He's taking some sabbatical right now with his family. Good for him. But he had a really interesting interview in October with an AI system called Pi, PI, and it was really cool and it really was super innovative and I enjoyed watching it. But I also thought, you know, this is cool. I have so many questions about OpenAI, specifically that I don't quite understand kind of this conflict between statelessness and learning. That's the big issue that I'm trying to get. How does it improve and how do we get our own version of OpenAI, which they talked about so much in the developer conference a little while ago, which I did. How do we train it? How do we help it? So in the spirit of that and kind of in the spirit of my, you know, kind of exploration of some of these digital virtual, not quite real spaces, I thought let's do the same thing that Yadin did and we'll have an interview with OpenAI. So I'm going to try to let this just run like a regular interview. I've never done it before, so there might be some stop and go on the engagement with the actual app. I've just got my phone, just like Yadin did on a stand here. I'm just going to open up the app on my phone and we're just going to talk to OpenAI so it should come up there. Okay, So let's give it a shot. And I think we hit this one. 

‍

Hello, ChatGPT. How are you doing? I'd like to interview today. 

‍

ChatGPT:
Hello. I'm here to help you, so I'm doing well, thank you. What would you like to ask? 

Jeff Frick:
Great. So first off, let me welcome you to the podcast. My name is Jeff Frick, and this show is called Turn the Lens. So Welcome. You're the first AI guest that I've had on the show. 

ChatGPT:
Thank you, Jeff. It's a pleasure to be here on Turn the Lens. Excited to be your first AI guest and eager to see where this conversation goes. What's the first question you have for me? 

Jeff Frick:
So the first question I have for you is a few months back at the OpenAI Developer Conference, it was announced that anyone could set up their own custom ChatGPT. Can you tell us a little bit about that announcement and kind of what were the objectives of that release?

ChatGPT:
Yes, at the OpenAI Developer Conference a few months ago, it was announced that developers could create their own custom versions of ChatGPT. This move was aimed at broadening the accessibility and versatility of the technology by allowing developers to customize ChatGPT. OpenAI sought to enable more specific applications tailored to different needs and industries. The objectives of this release were to, 1) Enhance innovation by providing developers with the tools to build upon ChatGPT for varied and niche applications, 2) to increase accessibility of A.I. technology making it easier for businesses, educational institutions, and other organizations to integrate A.I. into their operations, and 3) Foster a community of developers working on AI, encouraging collaboration, and the sharing of ideas to improve the technology further. This initiative was designed to expand the ecosystem around A.I., making it more adaptive and useful across a broader range of contexts.

Jeff Frick:
Okay, that's great. So I went up out and I set up my own ChatGPT which I was very excited to do. It's called Anita Iverson, and my goal was to train my ChatGPT on some of the transcripts from some of the interviews that I do around leadership, communication, future of work, etc. So I set it up. But I'm confused because when I tried to train my AI and upload transcripts with this problem with statelessness, it doesn't seem to retain the information. And I've also heard that, you know, actually training the AI is different than influencing it. So how does it learn if it has the stateless issue, if it doesn't retain the information as I tried to teach it?"

ChatGPT:
Ah, you're touching on some key aspects of how models like ChatGPT work and the challenge of statelessness. Here's a bit of clarification on these points. 1) Statelessness: ChatGPT and similar AI models operate in a stateless manner for each interaction. This means they don't remember previous interactions within a session unless specifically designed to do so through a custom implementation. This is a fundamental design choice to ensure privacy and scalability, but can be a hurdle for continuous learning from user interactions. 2) Learning versus Influencing: When you're training ChatGPT by providing it with transcripts or other forms of data, what often happens in practice is not real-time learning or updating of the model itself. Instead, you're influencing its responses by guiding it on what kind of information or style to mimic based on the input it receives at the moment. 3) Training A.I.: The actual training of models like ChatGPT happens offline and involves a massive dataset snd substantial computation resources. This process, known as fine-tuning, requires a curated dataset like your transcripts, where the model is trained over many iterations to adjust its internal parameters and better reflect the training data characteristics for solutions for your use case.

Integrating the transcripts and aiming for a more personalized and knowledgeable Anita Iverson, you might consider a few approaches. Prompt Engineering: Crafting your prompts to include relevant context or information directly in them, effectively guiding the model to generate responses based on that immediate input. Custom Training: For a more advanced solution looking into OpenAI's offerings for training custom models might be the way to go. This involves creating a dataset of your interviews and using it to fine-tune a version of ChatGPT specifically tailored to your content, which can be a more involved process. The key is to understand that influencing the model through real-time input can be powerful, but for it to learn from specific datasets in a way that's retained across interactions, a more formal training process is required.

Jeff Frick:
So how would someone like me that has a library of interview transcripts that I would like to suppose its influence for a long term into the model do that? They’re text-based so it's easily crawlable or easily put in a some type of a database. So how do I do that? Is there an interim step that's not so involved and not expensive and easy?

ChatGPT:
For someone with a collection of text-based interview transcripts looking to influence a model like ChatGPT in a more persistent way without delving into the complexity and expense of custom training, there are a few strategies you might consider. Contextual Prompting: You can manually integrate information from your transcripts into the prompts you give the AI. This method involves summarizing or including key points from your transcripts in each interaction. While this doesn't train the model in a traditional sense, it directs the AI to generate responses in line with the provided context. Retrieval-Based Models: Some developers create a retrieval-based system alongside the AI. The system can search your database of transcripts for relevant information based on the query and then feed this information into the AI model as part of the prompt. This approach requires some software development but is less resource-intensive than training a model. Incremental Learning Tools: There are tools and platforms emerging that aim to simplify the process of incremental learning for AI models, where the model can learn from new data in smaller batches without the need for full retraining.

These tools are designed to be more accessible and less costly, but might still be in early stages or limited in their capabilities. APIs with Customization Features: Some AI platforms offer APIs that allow for a degree of customization or personalization based on user provided data. These might not be as effective as a fully trained model, but can offer a balance between generic A.I. responses and those influenced by your specific content. The key is to find a balance between the effort and resources you can commit and the level of customization you need. For many, starting with contextual prompting and exploring, retrieval based approaches can be a practical first step as your needs grow or become more defined. You might then consider more involved solutions like developing a retrieval based system or exploring incremental learning tools.

Jeff Frick:
Okay, so back to my original question. With the intent of the release of anyone going in and setting up your custom ChatGPT, which I've done. Clearly, if I'm trying to influence it, I cannot tell the users to preload the prompt with the data that I'm kind of putting in through that back end. So what does that little back end door enable? That's part of the release of the open ChatGPT that I set up my own Anita Iverson.

ChatGPT:
The feature you're referring to allowing anyone to set up their custom version of ChatGPT typically includes capabilities for fine tuning the model on specific datasets. This means you can train the model on your library of interview transcripts to influence its responses directly without needing users to preload prompts with specific data. Here's how this can work for your setup with Anita. 1) Fine-Tuning: This process involves taking the base ChatGPT model and training it further on your specific dataset, your interview transcripts. The model learns from the patterns, topics and language used in your interviews, making it more likely to generate responses in line with those conversations. 2) Custom Integration: The back end setup for your custom ChatGPT, processing your transcripts into a format suitable for fine tuning, you would then use OpenAI's tools or APIs designed.

Jeff Frick:
Can I stop you for just a minute? Wait, can I stop you there for a minute? The way it's currently constructed. When I loaded the transcripts into the back end as text copy-pastes after I had done a number of them, say I want to, I want to add a total of 50. And I checked my work. Unfortunately, because of the stateless issue, it said it didn't remember any but the last transcript. So I'm still confused as to how it learns from those transcript loads. Unless it's in some type of a database that's associated with my account or my custom ChatGPT is there some some level of storage of that information? Did you get that last question? If I've set up my own custom ChatGPT and I load transcripts into the back to give it information, how does that stay relevant if it doesn't retain it over a period of time, or does it some or does it stay in there someplace?

ChatGPT:
When you set up your own custom ChatGPT and load transcripts into it for training, the information becomes part of the model's knowledge base in a more permanent way. Here's how it works. During the fine-tuning process, the transcripts you load and use for training are integrated into the model's parameters. Through fine-tuning, this process adjusts the weights within the model based on the patterns, topics, and specific language used in your transcripts. It's not about retaining individual pieces of information or specific conversations but rather about adjusting the model's internal structure to make it more likely to generate responses in line with the knowledge and style found in your transcripts.  Persistence of Training: Once the model is fine-tuned with your transcripts, the adjustments made to its parameters are saved, making the influence of your transcripts permanent for that version of the model. This means the model doesn't forget this training over time. Every time the model is used, it will generate responses based on its trained state, including the influence of your transcripts. Updating the Model: If you wish to update the model with more recent information or additional transcripts, you would typically need to go through a similar.

Jeff Frick:
So that's very helpful. So I have another question. I'm curious to know the extent of the dataset that you're pulling from. I got a hallucination that indicated clearly that LinkedIn was not part of the dataset. And how do I know the extensive extents or the reach of the dataset in which the model is searching or training?

ChatGPT:
Understanding the extent of the dataset that models like ChatGPT are trained on is important for grasping their capabilities and limitations. Here's a general overview of how these models are trained and the kind of data they might use. Broad Dataset Collection: AI models like ChatGPT are typically trained on a wide range of Internet text. This training data can include books, websites, articles, and other publicly available texts to help the model learn a broad spectrum of language patterns, information, and knowledge. Exclusion of Proprietary Sources: While the training dataset is extensive, it does not include data from proprietary or private sources without permission. For example, specific details from LinkedIn profiles or other protected information would not be part of the dataset unless it's publicly available in a way that allows for such use. 

Jeff Frick:
Okay, so now I'm getting it now. So you're you're not searching the data set that you were trained on to give me an answer. The dataset that you were trained on is simply a way to formulate an answer to the question or whatever the query is. (lost connection, bad internet day) Did you get that? Can I ask you another question? 

ChatGPT:
Of course! Feel free to ask another question. What would you like to know? 

Jeff Frick:
Okay, so I think I'm starting to understand the dataset that's used for training is not a dataset that's queried as an answer for something that is said after the fact. It's simply feed stock for formulating an appropriate response to some future query. 

ChatGPT:
Exactly. You've got it right. The dataset used for training an AI model like ChatGPT isn't actively queried for answers to specific questions that come after the fact. Instead, it serves as feedstock or foundational material that the model learns from during its training phase. This learning process involves the model identifying patterns, relationships, and information within the dataset, which it then uses to generate responses to future queries. The model doesn't access or search this dataset in real time when it's generating responses. Instead, it relies on the knowledge it has internalized during training to produce relevant, coherent, and contextually appropriate answers based on the input it receives. It's a bit like how a person learns from reading books or articles. They don't look up information from those sources each time they have a conversation, but rather draw upon what they've learned and internalized over time to inform their responses. 

Jeff Frick:
Okay, and just to be clear, when we do use the new make your own ChatGPT, and we do load information into the back end that's accessible. It does, as you say, influencing is the correct term and those influence nudges are stored not as specific data, but as ways to influence the direction of the output. 

ChatGPT:
Yes, you've summarized it well. When you use a feature like Make your own ChatGPT and load information into the back end, you're influencing the model's behavior. This information doesn't get stored as specific retrievable data points. Instead, it adjusts the model's internal parameters to influence the direction of its output. This process of influence is akin to fine tuning the model on your specific dataset. The model learns from the patterns, styles, and topics in the data you provide, which nudges its responses to reflect the characteristics of that data. However, the model doesn't remember the data as distinct pieces of information to be accessed later. It incorporates the essence of the data into its overall understanding and response generation mechanism. So when the model generates responses after being influenced or fine-tuned with your specific data, it does so by drawing on the adjusted patterns it has learned, guiding its outputs to align more closely with the nature of the input it was trained on.

Jeff Frick:
Okay, but the nudge stays for subsequent sessions. I guess the question is I'm trying to figure out how much investment I should make in inputting data to nudge the model. When I was kind of disappointed that it felt like with the statelessness that it didn't remember all the stuff that I had put in. But you're telling me it's going to retain that nudge?

ChatGPT:
Yes. The nudge or influence you introduce by inputting data and fine-tuning the model does stay for subsequent sessions. It's the persistent change to the model's behavior. This is because the fine-tuning process adjusts the model's internal parameters based on the specific characteristics of the data you provided. These adjustments are part of the model's configuration and don't reset between sessions. The key distinction here between the statelessness of individual interactions with the model and the long-term changes introduced through fine-tuning. Statelessness of Interactions: This refers to the model not remembering the context or content of previous interactions within a session unless specifically designed to do so. Each query is treated independently. Persistence of Fine-Tuning: When you fine-tune the model with specific data, you're essentially retraining it to understand and generate responses based on that data. This change is not session dependent and remains part of the model's learned behavior for all future interactions.

Jeff Frick:
Interesting. So how so? How long does this session last that we're having in terms of a conversation? How many of the things you know are you How far back does this thread go in your generation of new answers? See, I think I lost it. Did you hear the last question? (damn intenet)

ChatGPT:
Sorry, I'm having issues right now. Could not understand what you said. 

Jeff Frick:
Hold on So, hello, Welcome back. Hello again! Did you hear? 

ChatGPT:
I'm here and ready to continue our conversation. What would you like to discuss next? 

Jeff Frick:
So that's my question is, in the context of this conversational UI, how much of our prior Q&A have you retained? What is the duration of this session or this state. 

ChatGPT:
In the context of our conversation, I don't retain information from prior interactions or days. Each session with me is independent and I don't have the ability to recall personal data, history or specifics of earlier conversations once the session ends. This approach ensures privacy and confidentiality for users. The "duration of this session or state" ...(internet breaking, AT&T day)

Jeff Frick:
We’re having Internet issues we’re having an Internet issues today. Internet issues everywhere. I might have to reshoot this again, but this is a fun experiment it’s connecting again. 

ChatGPT:
How long is the session in the context of this conversation the session in this conversation context can vary. but typically, a session remains active as long as there's ongoing interaction. If there's a period of inactivity, such as no questions or inputs from you for an extended time, the session may reset, meaning any temporary context or state is cleared. For most conversational AI platforms, this period of inactivity might be a few minutes to several hours depending on how the system is configured. However, without a specific user activity or questions, I don't have a fixed “timeout” period to offer, as it greatly depends on the platform's design and the environment in which the AI is operating. 

Jeff Frick:
So that was. That was an experiment with talking to interviewing open AI. Now, what's interesting, if I bring back up the phone, it has done that whole thing and now I have a text record of it, so I will export the text record of the transcript. But it's pretty interesting. 

It did get kind of to the core issue of influencing versus learning versus training because a lot of, you know, myself, Bill Schmarzo, a bunch people went out, they set up their own OpenAI account mine is Anita Iverson, and I wanted to train it on all of the transcripts from all the interviews that I've done, and I uploaded all these transcripts and I spent all the time to get the text files in and make them clean and have speaker breaks and this, that, and the other. And then when I asked, when I asked Do you remember all this stuff? It said no. 

My take away from this conversation is two. One is, you know, it's not querying the data when it's answering your question. The data to teach it is simply a way to help influence the way that it's going To answer your question in terms of style and the structure of the sentence, very important distinction. And the second piece is that you can influence it. And those influences do go beyond the existing session, she said. When you do upload your transcripts and stuff. But again, it's not searching against those transcripts within its dataset to generate an answer like a search engine. It's just using that information to influence the way that it generates an answer. 

So it's a really different way to think about that answer. 

It's not searching for the answer and delivering the answer. It's creating the answer based on the way it's taught to communicate with what it thinks is the best probability of a correct answer. 

So interesting conversation. Thanks for staying with us through this kind of AI exploration. We're going to continue to do cool new things. Yadin, great job for doing the first one of these and being really the inspiration. You had a whole different set of conversations around assistance and helping, but the best way to kind of figure out what's going on is to jump in the pool and get wet and try it and see how it feels with what you think it might be able to help you with or not. But, um, we're certainly not there yet. But you don't want to left far behind. 

So thanks for watching. Another episode of Turn the Lens. We keep turning the lens and looking at things through different lenses. Thanks for watching. Thanks for listening on the podcast. Catch next time. Take care. Bye Bye.

‍