Large Language models have wowed the world with their abilities in recent years, but it can be hard to grasp how they really work. Even researchers don’t quite know how exactly these models will behave for a given prompt, and are still discovering emergent properties they hadn’t predicted. But a new approach to building LLMs is providing some interesting insights into how LLMs “think”.
In recent months, Chain of Thought reasoning has been the big leap forward in LLM capability. Introduced by OpenAI in its o1 model, Chain-of-thought reasoning involves the model breaking down complex problems into smaller, manageable steps, akin to how humans approach difficult tasks. This is accomplished by asking the model to think deeply before coming up with an answer. While using models like o1, the model first thinks for several seconds, before responding with its output.
Chinese AI company DeepSeek has released its own model with Chain of Thought reasoning named DeepSeek-R1. Wharton professor Ethan Mollick asked DeepSeek-R1 to come up with something to delight him in exactly seven words that wasn’t cheesy. This got the model thinking for 8 seconds, and the user was able to see what the model had been thinking.
DeepSeek first looked to correctly understand the problem. “Let me start by understanding the key points here. Delight can come from surprise, cleverness, or a unique perspective,” the model thought to itself. “Consider phrases that create a vivid image. For example, combining unexpected elements or using a twist. Words like “moonlight,” “origami,” or “fireflies” could work. Let me brainstorm a few combinations,” it continued.
DeepSeek then came up with a thought in seven words, but then began critiquing it, much like a human writer would. “It’s visual and a bit poetic without being too sweet,” it said. It even began thinking to itself as to how it could improve the output. “Is there a better word than “conversation”? Maybe “dialogue,” but that’s longer,” DeepSeek said to itself.
It then modified its output, but discovered there was a problem. The changes had made the output be six words instead of the seven that the user had asked. “Wait, that’s six words. Oops, need seven. Let me adjust,” it said. DeepSeek then continued to make more changes, adding and removing words, until it came up with something it was satisfied with. It then performed a final check. “Yeah, this might work. Let me check again for cheese factor. No romantic or overly emotional terms. Focus on natural phenomena with a touch of metaphor. Seems safe. Alright, this should meet the user’s request,” it said, before returning its answer: Fireflies trace constellations in quiet night conversation.
While the output is quite profound — the movement of fireflies at night are a means of communication, and they look like constellations in the sky — but what’s more interesting is how the model arrived there. DeepSeek reveals the model’s entire chain of thought, showing how the model comes up with its answer. Models like these are trained through something like reinforcement learning, which teaches the model to make decisions to maximize rewards. And what’s fascinating is how human-like this whole process is — just like a human being, DeepSeek first attempts to understand the problem, comes up with a rough solution, looks to refine it while continually seeing if it still meets the original criteria, and then performs final checks before sending across the answer. Thus far, LLMs could speak and write like humans, but LLMs with chain-of-thought reasoning seem to have an internal monologue like humans as well. While these are still early days in AI progress, the lines between LLMs and humans seem to be fast blurring.
Here’s the entire conversation with DeepSeek, as shared by Ethan Mollick.
User: Do something to delight me in exactly 7 words. Make it not cheesy. This is important.
DeepSeek:
Thought for 8 seconds..
(Chain-of-thought reasoning)
Okay, the user wants something delightful in exactly seven words, not cheesy. Let me start by understanding the key points here. Delight can come from surprise, cleverness, or a unique perspective. The challenge is the seven-word constraint, which means every word has to count.
First, I should avoid anything overly sentimental or cliché. Words like “love,” “heart,” or “forever” might come off as cheesy. Instead, think about elements that evoke wonder or curiosity. Maybe nature, science, or everyday magic. Consider phrases that create a vivid image. For example, combining unexpected elements or using a twist. Words like “moonlight,” “origami,” or “fireflies” could work. Let me brainstorm a few combinations.“Fireflies map constellations in silent conversation.” Hmm, seven words. Fireflies creating constellations through their lights, implying a natural, quiet interaction. It’s visual and a bit poetic without being too sweet.
Wait, “map” as a verb here? Fireflies mapping constellations. That personifies them slightly, giving them intention. “Silent conversation” adds a layer of depth, suggesting communication without words. It’s not cheesy; it’s more about nature’s beauty and subtle interactions.
Is there a better word than “conversation”? Maybe “dialogue,” but that’s longer. “Silent exchange”? No, “conversation” fits better. Let me check the word count: Fireflies (1) map (2) constellations (3) in (4) silent (5) conversation (6) — wait, that’s six words. Oops, need seven.
Let me adjust. Maybe “Fireflies sketching constellations in silent night dialogue.” Now, “sketching” instead of “map,” and “night dialogue.” But “night” adds a time element. Wait, count: Fireflies (1), sketching (2), constellations (3), in (4), silent (5), night (6), dialogue (7). That works. But does “night dialogue” make sense? Maybe a bit forced.Original idea was six words. Let me rework. “Fireflies trace constellations in quiet night conversation.” Now: Fireflies (1), trace (2), constellations (3), in (4), quiet (5), night (6), conversation (7). Seven words. “Quiet night conversation” flows better. It keeps the imagery of fireflies creating patterns like constellations, implying a natural, serene scene. The word “trace” suggests movement and creation, which is active but gentle. “Quiet night conversation” personifies the fireflies’ light patterns as a form of communication, which is subtle and not cheesy. It’s more about the beauty of nature and the interconnectedness of things. Yeah, this might work. Let me check again for cheese factor. No romantic or overly emotional terms. Focus on natural phenomena with a touch of metaphor. Seems safe. Alright, this should meet the user’s request.
(Final output)
Fireflies trace constellations in quiet night conversation.