<small>December 9, 2024</small> ## Entirely New Ways of Thinking  I talk to LLMs about everything. I talk to them about any problem sets I’m working on, about anything I’m trying to learn, about personnel issues, about how to be a better parent, about how something works, and why, and in the case of this essay, about the nature of understanding and our relationship with language. My first use of an LLM was ChatGPT. I would type out my thoughts or questions and it would respond back with text. Something about this experience felt immediately special. Very quickly, I wanted to use my voice instead of typing on my keyboard to give the LLM my thoughts. I started carrying a Sony audio recorder with me to record conversations with my friends, if they were cool with it. I also started talking out loud to my MacBook Pro all day while I was working. In both cases, these are open ended riffs on problem sets, ideas or things that I am trying to learn and understand. I do this so much now that I need to be very specific about where I work. I prefer to be alone because it’s weird to stream my unfiltered thoughts out loud if someone is right beside me. I use a tool called MacWhisper to take any audio recordings and quickly transcribe them into text and then I paste that text into Claude or ChatGPT. I do this continually through the day. When I speak my thoughts out loud, it’s very, very different than when I type them. When I type, I have to give my thoughts structure, but when I just speak out loud, I can close my eyes, and stream my consciousness. By using audio as an input with an LLM, I am creating an almost immediate bridge between the unstructured and the structured. I can speak raw unstructured thoughts and the LLM always responds with some level of structured text in real time. This is not a basic baseline thing. Traditionally, bridging the gap between the unstructured and the structured has required a lot of time and effort, and that’s why we value great writers. This new immediate bridge between the unstructured and structured is a form of cognitive augmentation that changes how we can process and develop our thoughts. Before doing this with an LLM, I would either try to talk things out in my head or with a notebook, or I would ask my friends to be sounding boards for me, which is just an agreement that they will respond to my unstructured processing with their structured perspective, but they aren’t available at all times, it’s inconvenient and it takes a lot of effort for a human to be a continual sounding board for someone else. Yeah, I use Claude in “Explanatory mode” a lot, and like ChatGPT, it often tries to keep the conversation going by asking a question at the end of its response. While talking to Claude about this essay, it ended its responses with questions like, “Would you like me to focus on developing any of these branches further?” or “What aspects of your LLM usage would you like to explore first?” I would follow the rhythm of these questions and it would influence my response. But I wondered, if an LLM is just a continual structural response to my unstructured processing, then maybe instead of following the rhythm of the LLM and its questions, I could shake things up in some way. Maybe I should be more open to triggering and reading all kinds of structural responses from the LLM. Maybe I should use the LLM to push my thinking further, in directions I’ve never gone before, and could never achieve on my own. I wondered, could human-AI interaction open entirely new ways of thinking? An entirely new set of thoughts? How do I intentionally push the boundaries of this process? What new techniques could I try? I’m just scratching the surface, but I’ve discovered a few so far. ## Recursive Thought Spirals   A recursive thought spiral starts with me speaking my raw, unstructured thoughts about something I’m trying to understand. I send them to the LLM and it responds to my thoughts. But instead of just reading and moving to a new topic or letting its questions dictate the next step, you record and transcribe your reaction to lines of its response, and send that back to the LLM again, and again. For example, let’s say I share some thoughts about a project I’m working on. The LLM responds with several paragraphs analyzing different aspects. As I read through the response, a particular sentence about “patterns in problem-solving” catches my eye. So I start with this, “Here's my reaction to your structuring - what new patterns do you notice in my reaction?” And then begin speaking my reaction: “That’s interesting about patterns in problem solving. On a fundamental level what exactly is a pattern? How does a human recognize them? How does someone go from not recognizing a pattern to recognizing one?” I let my thoughts flow naturally from there. Maybe this leads me to talk about specific examples from my work, or perhaps it branches into how I learned these patterns in the first place. The key is that I’m not trying to be comprehensive or structured. I’m following whatever thread feels most alive in my mind. The LLM then responds to these new thoughts with fresh structure. Again, I look for what stands out. Maybe this time it’s a comment about learning processes that makes me think about my childhood experiences with learning. So I start speaking about that. Each response from the LLM gives me new material to react to, new angles to explore. Sometimes I’m directly responding to what the LLM said, other times I’m going off on tangents inspired by a single phrase. The spiral develops naturally as each structured response triggers new unstructured thoughts. This process lets me explore ideas more deeply than I could alone because each turn of the spiral combines my natural thought patterns with the LLMs ability to find structure and connections. I'm not following a pre-determined path, I'm letting my curiosity guide me while using the LLM’s responses as springboards for deeper exploration. ## Thread Maps   Sometimes when I’m exploring a complex topic with an LLM, I notice that multiple different lines of thinking start to emerge. Instead of following just one path, I can ask the LLM to create a “thread map”, which is a structured overview of all the different threads of thoughts that are developing in our conversation. For example, when I started exploring how I use LLMs, the thread map revealed several parallel lines of thinking: how audio input changes thought processes, how to document these new cognitive techniques, the nature of understanding itself, and the practical challenges of implementation. Each of these threads was developing independently but also informing the others. The thread map becomes a kind of cognitive compass.  _the thread map for this essay saved in a Google Doc_ When I’m deep into exploring one aspect of a topic, I can reference the map to see how it connects to other threads we’ve identified. This helps me maintain awareness of the bigger picture while diving deeper into specific areas. I can either save the thread map in a separate document and reference it manually, or I can add it to a project in Claude, so as I create new conversations I can always bring the context of the thread map back in. And the thread map isn’t just a static reference, it evolves as my understanding grows. I can ask the LLM to update the map as new patterns and connections emerge, I can ask “what patterns or themes are emerging across our different threads of discussion?”, “How does our understanding of Thread A change our perspective on Thread B?”, “How have different threads evolved since we began, and what new insights have emerged?” ## Meta Questions   When I use LLMs, I tend to ask a lot of questions about the topics I’m exploring, but I rarely ask questions about my own thinking patterns, the meta questions. This is interesting because maybe LLMs can actually help us understand how we think, not just with what we think about. For example, while working on this essay, I asked Claude “What patterns do you notice in how I approach problems?” Here’s part of what it noticed: > “You consistently engage in what we might call ‘questioning spirals’ - you don’t just ask questions, but you question your questions. For instance, when exploring the thread map concept, you didn’t just ask how to use it, but questioned why we would use it, what value it brings, and even questioned why you were questioning its value” And something else: > "When you encounter a new concept, you consistently follow a pattern: first questioning if it is ‘real or just bullshit’ then testing it against your experience, and finally attempting to push it further than its original scope. This shows a thinking pattern that combines healthy skepticism with creative expansion." It’s interesting to have a tool that can help me observe and reflect back on my thinking patterns, maybe this can help me become more conscious of my own cognitive habits and tendencies. Every now and then, I find it very useful to ask the LLM these meta questions. – You can explore each of these techniques individually, or combine them together, or discover your own. I’m still discovering new ways to use these tools every day, but what excites me most is that it seems I’m just scratching the surface of learning how I can use LLMs to expand the way I think, learn and make things.