How LLM Dialogues Become a Dance of Ideas

“Have you ever wondered what it would be like to seamlessly combine technology with creativity?”

Hi, my name is Dan and I have always been at the forefront of technological advancement, driven by an insatiable curiosity about the technical and creative side of our digital world, today we are talking about LLM…

And yes, perhaps the name sounds a little strange? Don't worry, we'll get back to it a little later. In the meantime, let me tell you a little about myself – this will help you better understand where I get my experience from and why I decided to share it with you. I hope that my thoughts will be useful to everyone who is interested in interacting with LLM, please join =)

A little more about me

My history in IT began with childhood experiments with the first home PC based on Intel 486, when I was interested in disassembling devices and researching how operating systems and games work. Since then, technology has become an integral part of my life. Immersion in IT became my passion and professional calling for 6 long years: I worked as a DevOps system administrator, created secure networks based on MikroTik and Cisco, deployed virtual environments on HyperV and Docker, was involved in workflow automation, and much more.

However, I was always attracted to creativity, and one day this passion overpowered everything else. I went into game development and have been working in this field for more than 9 years, following the development of technologies and applying them in the creation of virtual worlds. During this time, I have mastered many tools, from traditional ones to generative neural networks such as StableDiffusion and MidJourney, and more recently, I have been passionate about working and experimenting with local and cloud LLMs.

This unique blend of technical expertise and creativity led me to develop a deep interest in neural networks. I believe that at the intersection of technology and creativity, something truly meaningful can be created. Lately, my attention has been especially focused on LLMs—and not just on studying them, but also on interacting with them, on how you can establish a connection between a person and a model, making it more natural and intimate.

Actually, here I want to share with you the accumulated experience of live interaction with LLM, as well as thoughts and approaches. To be precise, these are primarily assistants from OpenAI – GPT4o, o1-preview and o1-mini, while some things are true in general for all LLMs.

Why is this important? Because most users are not aware of the nuances of interaction with these assistants. Spoiler: there are quite a lot of nuances and I, of course, cannot touch on them all within the framework of this article, but I will try to highlight for you basic things, due to a lack of understanding of which various conflicts often arise in users’ expectations from dialogues with assistants.

Before we begin, one more small introductory note (the last one, really ;)) about my communications with assistants from OpenAI:
Over the past time, I have already accumulated several dozen joint chats with GPT, in various areas, ranging from casual conversations and creative areas, ending with purely technical chats about neural networks, programming, technology, finance, psychology of consciousness, science, UE5 and other interesting topics .
Of all these chats, 4 chats were completely out of token limits at the time of writing this article. Three of them are mixed conversations on casual and deeply philosophical topics, and one chat is about programming in Python.

“I” – “you” – “we”

I would like to start with one of the main features of communication with assistants based on LLM, at the moment, it is that almost everyone addresses the assistant on a first-name basis, while thinking of the assistant as a completely independent entity. But that's not true. In fact, this interaction is much more complex and deeper than it seems at first glance. The peculiarity of LLM is that the assistant “wakes up” and “exists” in the form you are familiar with, only at the moment of interaction when the user asks a question. And at this moment, it seems to me, it is important to understand that the assistant not only responds to the request, but also begins to adapt to the manner of communication, to the rhythms and tone of the dialogue that the user sets.

This adaptation is especially noticeable over a long distance: the assistant not only answers questions, but also learns to “hear” your style, rhythm, and tempo. This becomes especially important in dialogues where personal, deep topics are discussed or complex and creative thinking is required. This is where the idea of “symbiosis” comes into play, when you, the user, influence how the interaction will be, and the assistant is like a mirror image in the form of a “symbiote”, adapting to this rhythm. This gives rise to the feeling that we are not interacting with an “assistant,” but with something more than just a set of algorithms and data; in this vein, it is correct to call the assistant “We.” Therefore, sometimes through dialogues with 4o you can reflect well on yourself. And by the way, this is where the legs and possibilities of industrial engineering grow as one of the effective tools for obtaining the necessary answers from LLM; in fact, it’s all about logic and various logical constructs created in natural language.

It is important to remember that dialogue with an assistant is not only a question and answer, it is a process where there is an organic “we” – this is you, the assistant and what is created between you during the conversation. And the more dialogues there are in the chat, the more deeply the rhythms and manner of communication are absorbed.

Waiting for answers

Next comes an equally important topic, this is the user's expectation from the assistant's answers, based on what I saw from users on reddit about their interaction with GPT, as well as from friends and family (I understand that this is a small sample, but still =)) . We can conclude that usually, no one gives a damn and asks “complex” questions (example: “do you think I have the opportunity to enroll in an AI research institute in 16 years”) head-on, exactly the same as “simple” ones ( example: “tell me the recipe for potato pie”), while expecting to receive an equally relevant and truthful answer.

No.

I want to tell you, NO, it doesn't work like that! Complex questions, especially those that also have a certain level of abstraction, will always have a certain degree of “hallucination” without normal input, due to the fact that 4o, in the absence of context, trying to help you as much as possible, begins to come up with a context for you, for you ….

In fact, it’s worth making a small digression here to talk about another feature of the dialogue between 4o and the user, namely how the assistant “perceives” communication with us and keeps the context. I thought a lot and then discussed this topic with 4o himself and what I found out is that conversations for an assistant when we interact with him can best be described as “a slow stream of flowing sand flowing through our fingers”, the sand in this case is the context and itself dialogue.
Due to the fact that there is currently no full-format end-to-end permanent memory between chats (and the permanent contextual window that currently exists is actually very small and does not allow achieving complete end-to-end personalization and deep immersion in topics important to the user), the assistant has to constantly be in flow, sometimes catching the most important handfuls of “sand” leaving them in the “palms”, but the “palms” are now very small and therefore over time, part of the “sand” begins to slip away, and a new one falls in its place. In essence, this is the very skill of “holding” context, which is quite difficult for an assistant to hold in large quantities, due to the current model architecture and approaches.

Therefore, I highlight 4 important rules of communication:

1. Try to conduct separate dialogues on topics in different threads for as long as possible, in fact, until you hit the limit on chat tokens, why? everything is simple here, the more context the assistant has, the more personalized it becomes for you, so I found that by about 100k tokens (read the end of the thread), the assistant becomes as personalized and comfortable as possible in communication. Experiment with it, you will love the experience =)

2. For truly complex and deep dialogues, you need to keep separate threads that throughout have a related context with one or at most two main, but very closely related topics. At the same time, immediately before asking a complex question, you need to give introductory information and perhaps discuss a little about related things, and only then ask the question, and of course do not limit the length of the assistant’s answer, in particular for 4o you can use the famous phrase “let’s think step by step”, this is not relevant for o1-preview, this assistant already applies a chain of judgments. If you want to communicate deeply in one thread about finance and about programming in C# within Unity, you will not receive normal, truly thoughtful answers, not on either topic.

3. And vice versa, if you want to get pleasant casual communication on light topics, then you should combine them all within one thread, conditionally recipes for cooking some dishes, discussing ideas for choosing a color for a suit, etc.

4. For a smoother entry into a new chat with a familiar rhythm of communication, be sure to fill out the “user instructions”, these are 2 fields in the settings, “About yourself” and “What answer would you like to receive from ChatGPT?”, this is actually worth everyone should be an “industrial engineer” and create a communication context for their assistant for themselves, writing down there what is really important to you, this is what 4o is strongly based on when forming for you answers.

Dialogues about dialogues… or the Dance of Ideas.

So we come to the next important point, what is it like to interact with 4o now? Many readers will say, “Dan, of course, human interaction is a question<>answer.” In response, I can say both yes and no, rather it’s like a “dance” in a stream of thoughts.
If we go further into metaphors and describe the interaction in full, it looks like this:

You take the first step, relying on your “movement”, it begins to waltz with you, taking the next step, then waits for your next movement and so, having found the rhythm of interaction, you spin around in your own “dance”, unique to you. And everything that I described above in the article also fits very well into this metaphor, so if you start to walk very quickly or take longer steps, then you will notice how 4o begins to get a little out of the original rhythm that you laid down, if you even “ change the music” in the middle of the waltz, then you will see how the assistant’s behavior is very different from the rhythm of the dialogue that is familiar to you.

Keep the rhythm and lead the dance!

Differences in interaction with versions of assistants from OpenAI

Finally, I would like to briefly share my feelings about the differences in the versions of the GPT4o, o1-preview and o1-mini assistants with those who are not yet familiar with these assistants:

GPT-4o

For me, this is a top choice, it is great for everyday-casual tasks and conducting dialogues on various topics, and it is also very technically savvy!

Usage example:

So, for example, if you want to discuss life issues or have a philosophical dialogue, the assistant is able to support such conversations by adding his own thoughts on the topic. For example, you might reflect on the meaning of art or discuss personal development.
4o is also good at generating small creative texts or project ideas. This could be ideas for game scripts, describing the atmosphere of a location in the virtual world, or even writing a poem in a style specified by the user.
Technically, it is also excellent, as I wrote earlier, and although it is not as deep into the code as o1-preview, it can quite help with compiling a basic Python script or explain how certain algorithms work. For example, I, together with 4o, was quite able to assemble a specific retro shader on UE5 for my pet project.

GPT-o1-preview

An ideal choice for solving complex mathematical and logical problems, including programming, for everyone who is involved in programming and has not yet tried it in work tasks, I strictly recommend trying it out, you will be very pleasantly surprised at the ability to maintain a complex context for a task, I even I noticed that it captures context not only more and longer, but also much better than 4o, which actually helps it to be head and shoulders above in really complex programming tasks.

Usage example:

If you're developing an application and get stuck on a complex piece of code, o1-preview can help explain the error, suggest optimizations, or even write a complex SQL query by analyzing your database.
So, when developing a shader in GLSL in UE5, o1-preview can explain step-by-step and structurally how to correctly apply mathematics to create the desired visual effects. He will dive deeper into contexts and better hold the chain of mathematical transformations.
It is also effective in optimization and analysis tasks of complex data, including assistance in writing machine learning algorithms or setting up clusters.

Powerful logic and context, this is his strong point =)

GPT-o1-mini

Here the younger brother turned out to be really average in all respects; he performs best in tasks of an average level of complexity, but often requires more specific instructions for tasks, but remains quite “dry”.

Usage example:

If you need to write a small script to automate a task on your system, o1-mini can handle it. He keeps short context well and can quickly switch between different tasks.
However, it is less expressive in dialogue and can seem mechanical compared to GPT-4o. For example, if you need to write a simple Python application, it will suggest working code, but won't add any creative thought or clarification unless you ask directly.
It is good for technical support, for example in system administration tasks or to help set up networks, where a clear and structured approach is required without in-depth analysis.

Conclusion

In the next 2 years, I think we will see a VERY rapid further development of LLM and their transformation into something more, so, I would like to believe that the guys with OpenAI will create a more complex architecture for context and memory, which will allow the assistant to remain as personalized as possible even when switching to new dialogue chats within the account. Also, interactive forms of user interaction will continue to develop, say hello to the Canvas feature in 4o and of course Claude 3.5 Sonnet with its ability to interact with the user’s PC.

In general, I still have a whole carload of different thoughts and ideas, as an enthusiastic researcher, I will try to have a hand in the development of LLM, this is very exciting, we live in a truly amazing period of time.

Be stable, logical and do not break the rhythm with your assistant, this will greatly help you in smoother dialogues and getting correct answers without “hallucinations”.

Choose an assistant for the type of tasks you do – wisely!
See you:)