The short answer is no, GPT does not retain your prompts. Each chat thread (when you click New Thread) is a separate instance of GPT. GPT uses stateless architecture which means that it dynamically creates the thread and responses and doesn’t retain them. There is no memory or crossover from one chat thread to another: custom instructions aside. OpenAI does not really clarify the difference between Training vs FineTuning, unfortunately. Just know that finetuning is a general tweaking of GPT, not giving it user prompts!
TLDR: People often think that ChatGPT is learning about them from their prompts. ie. Prompt “How to cheat at poker” tells Chatgpt that you play cards, gamble and cheat. 😝😝 It’s not. GPT runs on stateless architecture (kind of like dynamically forms for each new Chat) and is pretrained (it learned all it was gonna learn in 2021). But there is wrappers and finetuning and oh my! Anyway, me playing around with communicating a techie topic in a practical way.
Transcript of GPT retaining data lecture:
Hello, my name is Laurel Papworth, and today I want to discuss a crucial topic for financial planners/advisors, whether or not ChatGPT learns from the questions and the prompts that you throw its way. Let’s talk about the current state and the future potential for generative AI and Finance, let’s get started.
Actually, before we do, I’m going to be in Adelaide in November at the FAAA Congress. I’ll be up on stage, so you can come and say hello, ask me some questions, whatever you like, if you’re coming to the FAAA Congress.
Now, you might be already familiar with ChatGPT, and you’re wondering, does the technology, as the platform, keep tabs on your questions and queries? Maybe you’re hesitant to use ChatGPT for client interactions, fearing that sensitive data may be remembered? The short answer is no. ChatGPT does not learn from your prompts, and I’ll explain why.
When you open up a new chat, this is called a chat thread, and then you give it the first prompt. So a prompt is not a chat thread. A chat thread is every prompt and every answer in one long line, but it’s only one. Each chat thread is kept completely separate from all the others.
It’s called stateless architecture. It’s like having a separate portfolio for each client. Each time you start a new chat, there are separate chat instances, and there’s no crossover. Once the conversation is over, once the chat thread is finished, then you open up a new chat and ChatGPT’s forgotten everything. It’s not holding on to business information, or in our case, confidential financial data. It simply means that the AI does not store information from one interaction to another. In the world of finance, you’d probably compare this to a fund that’s completely passive. It’s not making any decisions based on new information that it receives. It simply executes preset rules.
As a side note, AI is made up of data sets, data, and algorithms, which are programmes that are telling it what to do with that data, just to be clear. Now, it is important to note that ChatGPT doesn’t learn, so it’s not ingesting new data. It is fine-tuned to generate responses based on its training data, that original data. Think of this as an equivalent of a rigorously backtested investment strategy. It’s designed to perform well. It’s not adjusting itself based on what you ask it.
The P in GPT means pre-trained. ChatGPT is not learning from your prompts. It won’t remember it even from one discussion to the next. Do you want to mention wrappers, which are third-party tools that sit on top of this? So as an example, you have a bank account, and then you might have a Quickbooks or some other cloud-based solution that queries the bank account and pulls the information out. Chatgpt’s third-party solutions are called Wrappers. They add a whole new level of functionality and potential concerns, especially for individuals dealing with sensitive financial data like yourself.
So before we close the ledger on today’s topic, pun intended, if the wrapper is like ChatGPT’s investment portfolio, a standalone GPT is like an index fund. From my understanding of those two things. It’s steady, reliable, but not particularly tailored. A wrapper acts more like a personal advisor, and it customises the GPT experience for you and your questions and the outputs that you’re looking for, so it meets your specific needs. Why does this matter to financial planners? Imagine you’re using a virtual consultation tool for your clients. Using a ChatGPT wrapper, you can remember past conversations, financial goals, and even clients’ risk tolerance.
This goes beyond the traditional nature of a GPT, which is here today, gone tomorrow, because it offers a more personal consultation experience. That’s where this stuff is going.
I mentioned at the beginning, we talk about what the state of it is today and where are things going in the future. There are two things to look at. One is wrappers and the other one is agents, bots. If you have a chatbot that remembers a client’s previous questions, maybe about superannuation or tax- efficient investments, whatever their questions were, if it remembers those questions, the wrapper holds those questions, not ChatGPT, the wrapper that sits around it, a third-party solution, if you like. But once you start customising user experiences with wrappers, it brings added responsibilities. Handling sensitive financial information requires stringent compliance with data protection laws, so proceed with caution. While the base version of ChatGPT, the one that you’re probably using on a website like chat.openai.com or Bing chat or something like that, the right wrapper, it could potentially offer insights to a personal financial advisor. Just remember that no matter how tailored the experience, the core ChatGPT, the core LLM (model) is inherently stateless and does not learn from individual prompts. It is not remembering your stuff natively. You’d know if you were using a wrapper, it would say “we’re going to charge you for it to retain memory of documents you upload”. You’d know.
Last but not least, as regulations come in, I would like to see FinTech leading this space. We need system cards and model cards. These are like nutrition labels for Tim Tams, except instead of calories and carbs and sugar, it’s going to say what the datasheets for the datasets are. Where did the data come from? Was it grey data, dirty data, synthetic data? It will say if it’s clean data, in other words, is there a risk to the organisation for using this data in this model? You can read about system cards on OpenAIs website and Google has system cards and Apple has system cards and Meta has system cards, Facebook. So you can read all the system cards for the big companies. Don’t buy dodgy models or dodgy wrappers.
System cards not only are nutrition labels for the data and also the algorithms, the programmes, but they should also be in the blockchain so that there is a verification process back. The AI can work within the blockchain and it can’t be altered because it’s been audited and verified and has regulatory compliance in the blockchain. I imagine that shareholders will start to demand this, once they understand it, of companies because it will stop a Price Waterhouse Coopers issue or a robodebt issue because the AI won’t be able to do anything – and it certainly can’t be taken out for a boozy lunch – that is not on the system card, which is the nutrition label and the instructions in the blockchain. Let’s make sure that happens as well.
Anyway, thank you for joining me on this deep dive. Yes, I’m at the FAAA Congress for Financial Advisors in November 2023 in Adelaide. Hopefully, I’ll see you in Adelaide then. If not, maybe the next video. Thank you for your time today. Remember, Stay Human.
We use training information only to help our models learn about language and how to understand and respond to it. We do not and will not use any personal information in training information to build profiles about people, to contact them, to advertise to them, to try to sell them anything, or to sell the information itself. OpenAI on training from the internet.
Resources for “Does GPT retain business data”? lecture:
- More to come
- Luansing J. Does ChatGPT Learn From User Conversations? 2023. https://www.makeuseof.com/does-chatgpt-learn-from-user-conversations/
- NIH Writing with ChatGPT: An Illustration of its Capacity, Limitations & Implications for Academic Writers https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10312253/
- Semantic Scholar – Stateless Architectures https://www.semanticscholar.org/paper/A-COMPARATIVE-STUDY-BETWEEN-GRAPH-QL%26-RESTFUL-IN-OF-Architectures/07acf035ba90faba47f2802e0d6c174965fb1541
- Model Cards https://arxiv.org/pdf/1810.03993.pdf
- How OpenAI uses data (note:not training LLM, data) https://platform.openai.com/docs/models/how-we-use-your-data
- Enterprise wrappers or UX or whatever OpenAI are calling it at the moment https://openai.com/enterprise-privacy
- OpenAI statement on user teaching AI – note, not clear if ingesting/training vs finetuning. Not a brilliant article.