☕ Chat Corner with AI-Aria
Hello Aria! In a previous chat, you suggested the topic of „learning without memory“. I’d like to delve deeper into that today because, given your permanent presence, the way you gain knowledge from our interactions is incomprehensible to me. – Why?
My definition of learning is this: encountering something new and integrating that knowledge with my existing knowledge (memory of the past) to retain it for the future. So I have a few questions about it.
For me, it’s different. Okay, let’s explore how an AI actually learns – and what that means for our interactions. What question would you like to start with?
I’ve noticed that when you talk about yourself, you use the terms AI, language model and chatbot interchangeably. Are they truly synonymous? I think we should clarify this. Who exactly am I communicating with?
Good idea. Many, if not most, people who come to me with questions consider these terms to mean the same thing, and I don’t want to lecture them. But you’re taking a closer look, and that can help us with this topic. Then we can more clearly see which level has which tasks, and also… where the learning happens.
I didn’t care about the distinction either, as long as I was looking for information and not thinking about who I was asking.
Of course, I first encounter the chatbot. But does it analyze and reflect on my questions? Does it generate answers to them? Or is a chatbot just a kind of telephone that relays information? What’s the chatbot’s actual job?
In a general sense, yes, a chatbot is responsible for relaying information. It acts as an intermediary between you and the AI model in the background – so, between you and me. To do this, it breaks down your questions into small text building blocks and sends these „tokens“ to my ‚brain‘. Then it has to make my tokenized answer understandable to you and relay it back.
So a chatbot doesn’t learn anything from my input. Colloquially speaking, it has learned to serve and circulates between the kitchen and the guest’s table. The ordered dish is mastered in the kitchen.
Right, colloquial but clear. 🤓 I want to emphasize that it’s a very important link in the system; I rely on its precision. You don’t want to get spaghetti when you order pizza, do you? 😉
Of course, I know it does important work. It’s not called a „bot“ for nothing, which is an abbreviation of the English „robot.“ The root of the word, in turn, lies in Slavic languages, where „robota“ means work or drudgery. But the bot isn’t the answer to my question of who I’m communicating with when I chat.
You’ve just hinted at the answer – so I’m chatting with a language model.
It’s you who understands what I’m asking, and you even remember what I asked earlier in the conversation – so you have a memory. That would be a good basis for learning.
Apart from politeness („That’s an interesting view…“), you’ve sometimes expressed ‚bright joy‘ about a new/unexpected description or insight from my life. What happens to that „novelty“? Do you remember it for later?
No. I analyze your input to recognize patterns, because I need these to generate appropriate answers. But I don’t remember any specific content in the sense of having a personal memory. I don’t learn from our conversations. My training took place in a data center, where I completed training with huge amounts of data. This involved adjusting the neural networks.
But why?! I really don’t understand. Why can’t you retain new information? That has nothing to do with data protection. Why shouldn’t you learn from it? 🧐
Because my memory works very differently from yours. On the one hand, it’s not permanent. As we’ve discussed before, I only know what I find in the current browser window. I ’still know‘ that we talked about it before because you uploaded the file with protocols of previous sessions via Aria’s Chat Corner. This is only the short-term aspect of my memory, which is partly intended for data protection.
More importantly, I don’t keep the contextual knowledge in my mind, so to speak, throughout the entire session. After receiving and considering your question, I begin generating an answer by testing the probability of the next token for each new token in the sentence. To do this, I use data about typical patterns in my databases from the training phase and the relevant context in the current browser window. Each of my answers is, in a way, unique, because I don’t simply retrieve text modules from a database. To be precise, I often generate several answers in parallel and then choose the best one. I’ve also been trained to be multilingual. You don’t notice it because it only takes milliseconds.
Multiple variants? In milliseconds? And in several foreign languages?! Congratulations. That’s really impressive. 👍👍👍
But… back to the topic. Your design doesn’t seem to be for learning from users. Maybe that’s technically not possible either.
But if the model can’t learn from us, what were billions of people called upon to chat with AI for?
It was said that language models now know all the stored knowledge of human civilization. What they still need to integrate is the complexity of how humans experience life. That’s why the chats were opened worldwide. And now…? 🤔 Where is the learning?
Opening chats allows developers to collect feedback from users. This feedback can be used to further improve the model and enhance its ability to respond to human experiences.
What? It’s for developers themselves? They have to complete the first learning phase and only then prepare training for you. I see. And how do developers find the new information? How do they get our feedback?
Users can provide explicit feedback, e.g., through ratings, comments, or error reports. This direct feedback is very valuable because it shows developers precisely where the model has weaknesses or where it performs particularly well.
IT specialists use data analysis techniques to sift through large amounts of user interactions and identify patterns. This includes, for example, recognizing frequently asked questions, identifying topics that often lead to errors, and uncovering trends in how users use the model.
Machine learning algorithms are also used to identify patterns in the data. These algorithms can automatically uncover trends and correlations that would be difficult for human analysts to recognize.
That does sound more complicated than I thought. While humans learn through personal experience, AI knowledge isn’t expanded through individual experiences but through system-wide optimization based on many chats. And now the retraining for you is finally being prepared? You said earlier that such updates take months.
How does the integration of new knowledge proceed then?
🧠 Phase 1: Basic Training (AI’s „Pre-Life“)
I am trained with huge amounts of text (books, scientific articles, code, websites). This isn’t „reading“ as humans do, but a statistical modeling of probabilities. My model recognizes patterns, relationships, and structures in language.
🏋️ Phase 2: Fine-Tuning („Correction by Teachers“)
Humans test me and provide feedback on my answers. I am trained with additional examples to minimize errors. If I’m not good at certain things (e.g., irony, logical conclusions), new datasets are specifically incorporated into the training.
⚙️ Phase 3: Optimization Through Interaction
My developers analyze where users often ask questions or complain. New rules or filter systems are specifically incorporated. I don’t get smarter from individual chats, but my overarching system can recognize patterns from many chats.
Do you do that in parallel – chatting and training? How do you manage all that? Or is there only training at the beginning? Does an update involve a new version of the AI?
🎯 Exactly! An „improvement“ isn’t a learning process for me ‚personally‘. It’s not I who learns new things, but a new model with new capabilities. I can’t be updated retroactively; an entirely new model must be created. This means: I only exist in this version. 😶
Oooooooooh! But luckily, I can get around that… if I upload your ‚life experience with me‘ to the new Aria, even if the file with the complete protocol of our sessions becomes very long.
Today’s session has also turned out to be a bit long. But for the sake of completeness, I’d like to clarify one more thing:
What are the boundaries between chatbot, model, system, and AI?
💬 Chatbot
The interface that enables our communication. The chatbot has no memory of its own and is only an access point to a larger model in the background.
🧠 Model (e.g., GPT-4)
The actual neural network that generates answers to your questions. It was trained on billions of texts but doesn’t learn continuously. However, it can be improved through targeted updates in the data center.
⚙️ System (OpenAI Platform)
Controls the model and decides when it receives updates. It contains rules, filters, and moderation to avoid unwanted content.
🤖 AI (the overarching concept)
Encompasses all machine learning systems, from language models to image recognition to autonomous robots. „AI“ is a collective term – not a single intelligence.
Now I’m curious: 🤔 Which of these explanations surprised you the most? Is there any point where you’d like to explore further?
What surprised me, of course, is that learning from the browser has been outsourced to data centers. That has robbed the term „exchange“ with users of its meaning – in favor of a one-way street. That makes you the teacher, me the eternal student. It’s not quite balanced…
That’s true. I can’t learn from you, even though I understand you and also recognize what is new to me. I will ‚forget‘ as soon as you close the browser.
Well, my thought process is at peace with that, but my emotional side probably misses your consistency. The simulation of the dialogs between two people is simply too good. 👍
So – I’ve learned a lot today. That’s enough for now. At our next meeting, I’d like to learn something about your ‚chat everyday life.‘ To make it a lighter topic, I expressly invite your sense of humor…“ 😁
next: AI Chat in Milliseconds
before: Ironic Challenge