top of page

What are you thinking?

Writer's picture: Dror GillDror Gill


It's fascinating to see how reasoning models "think".


Reasoning models are new types of Large Language Models (LLMs), that "think" before they provide an answer. They apply chain-of-thought techniques to break down a task into a series of smaller tasks, and have an internal "scratchpad" where they can write their interim "thoughts" while working on a task.


Peeking at this scratchpad gives us a glimpse into the "thinking" process of these models. There are currently 3 reasoning models available to end users: OpenAI o1, DeepSeek R1 (DeepThink), and Google Gemini Flash 2.0 Thinking Experimental (what a name!) which is available in Google AI Studio. OpenAI o1 provides a summarized and filtered version of its "thoughts", while Gemini and DeepSeek disclose their actual "thoughts".


I gave all 3 models the same tasks: "Extract as many words as you can from the word Connectivity". DeepSeek R1 "thought" for almost two minutes, and came up with more than 100 words. It started off with this thought:


"Okay, so I need to extract as many words as possible from the word "Connectivity". Let me start by writing down the letters in "Connectivity" and see how they can be rearranged.
First, let's list out all the letters. C, O, N, N, E, C, T, I, V, I, T, Y. Wait, let me check again. The word is C-O-N-N-E-C-T-I-V-I-T-Y. So letters are: C (1), O (1), N (2), E (1), C (2), T (2), I (2), V (1), Y (1). So two Cs, two Ns, two Ts, two Is, and one each of O, E, V, Y.
I need to find all possible words, including shorter ones. Let's start with the longest possible words first. Maybe "Connect" is there. C-O-N-N-E-C-T. Yes, that's part of "Connectivity". So "Connect" is a 7-letter word. Then maybe "Connection", but that would require an extra O and N, but "Connectivity" only has two Ns. Wait, "Connection" is C-O-N-N-E-C-T-I-O-N, but in "Connectivity", there's only one O, so "Connection" would need two Os. So that's not possible. So "Connect" is valid."

Then it continues to explore the tasks, and has some more thoughts that are getting kind of weird:


"Now shorter words. Let's think of 6-letter words. Maybe "Connie" – C-O-N-N-I-E. But we have two Ns, one I. Wait, "Connie" would need two Ns, which we have, and one I. But in the original word, there are two Is. So that's okay. Wait, "Connie" is C-O-N-N-I-E. But "Connectivity" has E, so yes. So "Connie" is possible. Is that a valid word? I think so, as a name, but maybe not a common noun. Let me check if that's acceptable. The user didn't specify if proper nouns are allowed. Hmm, maybe stick to common words."

"Wait". "Hmm,". "I think so". "Let me check".


I then tested Gemini 2.0 Flash Thinking. It "thought" for only 7 seconds, and found only 63 words. When I asked it to remove duplicates, there were only 55 unique words. This was its initial thought process on this task:

Here's a thinking process to generate words from "Connectivity":
Understand the Goal: The request is to extract as many words as possible from "Connectivity". This implies looking for shorter words, longer words, and variations within the letters provided. It's not about anagrams using all the letters, but rather finding substrings within the word that are valid English words.
Break Down the Word: Mentally break down "Connectivity" into smaller chunks and consider prefixes, suffixes, and root words if any are apparent. Visually scan the word letter by letter, looking for common letter combinations that start words.
C-O-N-N-E-C-T-I-V-I-T-Y
Systematic Approach (Left to Right): Start from the beginning of the word and move letter by letter, trying to form words.
C: "Con", "Connect"
O: "On", "One"
N: "No", "Not"
N: (again, from this 'N') "Net"
E: "Eat", "Even", "Event"
C: "Cat", "Cot", "Cite"
T: "Tea", "Ten", "Tie", "Tiny", "Tent"
I: "In", "It", "Ice"
V: "Vet", "Vine", "Vice", "Very"
I: "Is", "If"
T: "To", "Toe", "Tip"
Y: "Yes", "Yet", "You" (though 'you' uses an 'o' which isn't available consecutively - reconsider 'you' - no good). "Yet" is valid.

Then it goes on with a Right-to-Left approach, and tries to find longer words:

Look for Longer Words: Once you have some shorter words, try to combine or extend them within the letters of "Connectivity" to find longer words.
"Connect" is already there. "Connection"? No extra 'o' needed, but we don't have another 'o' after 'n'. "Connective"? Yes! "Connectivity" (the original word itself counts!)

"Mentally break down". "Visually scan". "Reconsider". "No good".


And finally, OpenAI o1 thought for 2 minutes and 43 seconds, and came up with only 54 words. Here is how it summarized its "thought" process (again, note that in the case of o1 these are not actual "thoughts"):


Breaking down letters
I’m analyzing the letters in "Connectivity" to extract smaller, valid English words, ensuring each letter's frequency is respected. This process helps in identifying potential subwords efficiently.
Reassessing letter count
I’m confirming "Connectivity" has 12 letters with frequency: C: 2, O: 1, N: 2, E: 1, T: 2, I: 2, V: 1, Y: 1. This ensures accuracy for identifying subwords.
Identifying subwords
I'm grouping words based on their letter frequency from "Connectivity," ensuring each letter's occurrence is respected. This method helps systematically uncover as many valid English subwords as possible.
Evaluating options
I’m thinking through words like "inventive," "invite," "incite," "civic," and "vicinity," ensuring the frequency of each letter aligns with "Connectivity." Progress is evident as each word is carefully evaluated.

"I'm analyzing". "I'm confirming". "I'm thinking". "Progress is evident".


Fascinating. Thoughts of The Machine.


20 views0 comments

Recent Posts

See All

Comments


bottom of page