Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Making Connections: Artificial Intelligence mee...

Making Connections: Artificial Intelligence meets Abstract Ingenuity using NYT Connections by Uyi Uwadiae

What happens when you pit eight leading AI models against The New York Times Connections puzzle?

In this data-driven experiment, I tested free versions of Gemini, Claude, DeepSeek, Vibe by Mistral, Copilot, ChatGPT, Muse, and Grok across 10 rigorous rounds of Connections—combining 5 historically difficult published NYT games with 5 original custom puzzles designed to test zero-shot lateral thinking.

Key Highlights & Findings:

The Top Tier: Gemini and Claude tied for the most connections found (30/40), with Claude taking the top spot when factoring in partial credit for "one-away" guesses.
The Custom Puzzle Wall: While models successfully cleared 91% of official NYT boards, they hit a 100% impenetrable cognitive wall on non-semantic, visual, and structural custom puzzles (0/8 on themed boards) due to tokenization blind spots.
Model Personalities & Quirks: From Gemini’s aggressive "Purple-first" hunting to ChatGPT's 16 near-miss "one-aways" and Copilot's hilarious Windows version hallucinations, each model exhibited a distinct behavioral play style.

Ultimately, this benchmark illustrates why generative AI excels at semantic pattern retrieval, but why true lateral creativity, visual intuition, and abstract wordplay remain uniquely human.

Check out my companion essays and the Crafting Connections series on Substack and Medium!

Uyi Uwadiae

November 03, 2026

Transcript

11 slides
  1. Making Connections: Artificial Intelligence meets Abstract Ingenuity using NYT Connections

    Eight AI Models Tested on Lateral Thinking and Finding Patterns Uyi Uwadiae | LinkedIn | October 2026 Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.
  2. Eight free AI models played NYT Connections the way humans

    do The players The puzzles • • Free versions of ChatGPT, Claude, Copilot, quarter (human solve rates ~34.8% to 43.7%) DeepSeek, Gemini, Grok, Muse and Vibe by Mistral • Thinking or reasoning mode on for all but Grok, which • 5 custom puzzles of mine: 2 already played by others, 3 unpublished didn't offer this in its free version • 5 published NYT games, from the hardest of the past Conducted between October 13 and 15, 2026; Copilot ran until October 26 given its daily chat limit The rules The scoring • One group of four (a connection) per guess, with the • 1 point per connection found, 40 possible per model reasoning explained • Partial credit adds half a point per one-away • I answer Y, 1 (one away) or N; a 1 or an N is a mistake • Also tracked: mistakes, perfect rounds, and Purple-first • Four mistakes ends the round • Solutions revealed after every round; no looking up answers; Purple first encouraged; one run per model rounds • This is just one lens on how AI models reason and not an evaluation of their relative strengths. Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.
  3. Gemini and Claude tied at the top with 30 of

    40 connections found NYT Connections My custom puzzles Games won: NYT | Custom 5/5 | 2/5 5/5 | 2/5 4/5 | 2/5 4/5 | 2/5 5/5 | 1/5 4/5 | 0/5 4/5 | 1/5 4/5 | 0/5 Gemini 3.8 Flash, Claude Sonnet 5.5, DeepSeek 4.1 Flash, Vibe by Mistral (Z.ai GLM 5.2), Copilot, ChatGPT 6, Muse Spark 1.3, Grok 4.5 Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.
  4. AI found 91% of the NYT connections but only 38%

    of my custom ones 91% 38% 38 mistakes on the NYT Connections 137 mistakes on my custom connections NYT Connections My custom connections 145 of 160 connections 61 of 160 connections Share of possible connections found across all eight models. The full bar is 100%. Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.
  5. Giving partial credit for “one away” guesses moves Claude to

    the top spot and bumps up Vibe and ChatGPT Total connections found Total with partial credit Rounds with no mistakes Total mistakes Partial credit: 1 point for each connection found, plus 0.5 for each “one away” guess (three of four right). It rewards models that got close. For example, though they both got 30 correct guesses, Claude had a two more “one away” guesses than Gemini. Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.
  6. It wasn't the trickiest puzzle that stumped most of the

    AI models Models that solved each NYT puzzle (out of 8) | human solve rate Sep 2 8 of 8 models | humans 34.8% Aug 24 8 of 8 models | humans 36.7% Jul 1 8 of 8 models | humans 40.4% Aug 28 3 of 8 models | humans 41.3% Jul 2 Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026. 8 of 8 models | humans 43.7%
  7. Gemini led the field with Purple-first guesses Rounds where the

    first guess was Purple (out of 10) 6 4 4 4 3 Reverse Rainbow: Purple, then Blue, Green and Yellow, 1 with no mistakes. Gemini did it twice (July 2 and August 28 puzzles), 0 DeepSeek did it once (July 1). 0 Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.
  8. No model solved any of my three themed custom puzzles

    The models handled trivia and pop culture well but struggled when the link was in how words look, sound or are spelled. 5 of 8 models What We Really Need 5 of 8 models Treasure Chest Numbers Game Calculated Guesses 0 of 8 models 0 of 8 models 61 total correct guesses on my custom connections. Only 14 out of 96 New Testament 0 of 8 models correct guesses for the three themed connections puzzles Find more of my Custom Connections puzzles on Connections Plus, and my thought process behind them in my series Crafting Connections! Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.
  9. Each model exhibited a distinct personality and way of play

    Gemini Claude DeepSeek Vibe by Mistral Purple first and 0 mistakes on Spotted decoys, then Thought long and hard, left no The comeback kid; most every NYT puzzle overthought last rounds stone unturned animated player Copilot ChatGPT Muse Grok Solved all NYT Connections, The just-fell-short king, with Warm, curious, asked smart Perfect or bust; no free lost on mine 16 one-aways questions reasoning mode Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.
  10. The wrong answers were some of the best answers Some

    notable guesses from the exercise. See more in my Substack article. Hush, Little Sonnet These codes are loyal Ocean's 3 Claude fell for the lullaby red Copilot guessed 7, 8, 10 and 11 as Claude, Gemini and Copilot all herring in the July 2 NYT Windows versions, noting that guessed 8, 11, 12 and 13 as the Connections. Windows 9 was conspicuously Ocean’s movies in Numbers Game skipped. It's actually cerulean Vibe coding connecting Vibe taught me that Buff (a light ChatGPT, Claude, Grok, Muse, brownish yellow) is a color because and Vibe made at least one guess it thought colors were a category in by just going off ~vibes~ Treasure Chest. Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.
  11. Generative AI is an extraordinary tool for organizing information and

    retrieving patterns, but Connections reminds us what makes human thinking truly distinct Artificial Intelligence may have just met its match in Abstract Ingenuity LinkedIn | Instagram | TikTok | Reddit | Substack | Medium | YouTube | Twitter (X) | Threads | BlueSky | Pinterest | GitHub Disclaimer: I am not affiliated with The New York Times or NYT Connections or any of the entities mentioned. Uyi Uwadiae, October 2026.

More Decks by Uyi Uwadiae

Other Decks in Technology