Mystery Hunt 2026

Jan 24, 2026

This post is about the 2026 MIT Mystery Hunt. You can see the hunt website with most of the puzzles and solutions at puzzmon.world.

Thanks to Cardinality for writing an excellent hunt this year. I had to decide between participating in hunt for my 12th(!) time in person vs. attending the birthday of a very dear friend; I chose hunt in part because I was convinced that Cardinality would make it worth my while, and they did not disappoint. I experienced generally clean puzzles, novel round concepts, and a lot of fun. Below I’ll cover a few notable points.

An inflection point for AI in puzzlehunts

In early January 2024 I bet $30 of fake currency on this Manifold market, that LLMs would NOT be banned in the 2026 MIT Mystery Hunt.

My reasoning was that

• The ethos of Mystery Hunt is that you’re allowed to use whatever tools are at your disposal to solve the puzzles. LLMs were “just another tool”, no different than reverse image search, etc.
• LLMs would not improve enough by then to meaningfully impact the experience of solving puzzles.

In fact, LLMs were NOT banned, but I think the second point above was clearly wrong. For my first time in a puzzlehunt, I found myself reaching to LLMs to a significant degree and they definitely helped our team. Interestingly, the way in which AI helped the most was in generating ideas on what a puzzle was referencing, rather than in the execution of solving a puzzle once we knew what to do. This is by no means obvious, but I do think it makes sense: many puzzles have references in their flavortext to media that any given solver may not be familiar with, but LLMs, which are trained on all human knowledge, can catch these references with ease. For example:

• I used an LLM to tell me that Town of Terror was referencing the 1987 edition of Arkham Horror, a board game I had never heard of.
• An LLM told me that The Hexagon was a reference to the Tour de France. (I had no clue that France was nicknamed “The Hexagon”, or that the Tour de France was held in July/August.)
• It was an LLM that had the bright idea on Chemical X to spell out the clue answers as chemical elements.

I did use AI in a couple of instances for execution:

• In Spilled! I got it to identify a few of our missing Hamlet quotes as well as write code to solve the Hamle.
• In DROP * FROM Teams I got it to write code to create a joined table with all of the data.

But in many cases, it was faster to hand a puzzle to a human teammate who could either crank out a program or solve a logic puzzle faster than it would take me to painfully transcribe our data and feed it to an LLM. I do think that if we had some kind of agent that was armed with tools to annotate letters and numbers on an image, that would have made a meaningful difference in getting an LLM to solve a puzzle end-to-end, but without such tooling a lot of execution was just faster and easier to do with humans.

I created a market for the same question in next year’s hunt, and I’m curious to see how things shake out.

Hunt is too long

It is! How many times does Dan Katz have to repeat this before we listen to him? We are way, way overdue for a hunt that finishes before Sunday morning. I also 100% agree with Dan that the era of aggressive meta hinting has given folks an unrealistic sense of how tractable modern hunts really are. If all but one of the teams is unable to finish without a meta hint, things have gone quite wrong.

Favorite puzzles

• Devilish Devilries Incredibly fun and funny puzzle to solve with a group.
• Spilled! Clever idea, and perfectly calibrated to make standard tooling for trigrams difficult to use.
• Balancing Act A novel word puzzle that I really enjoyed.
• The Hexagon I have some quibbles with the fact that solution doesn’t use the individual cities passed through, but other than that this was a brilliant round and capstone.
• Hyperbolic Space It’s a little too structurally similar to 2024’s A Rift in Hades for my comfort, but I had a great time solving this round.
• Land of No Name I didn’t like the capstone very much, but the round as a whole was terrific.
• Ancestry I learned about C/M/P taxonomy from this video by Epicurious about trying every type of citrus. Incredibly fun moment to open this puzzle and have my brain make the mental leap to the puzzle’s conceit.
• All of Layers was a fun, easy, group solve.
• At the Close was the puzzle we solved on the final “runaround” (not a lot of running around, which I didn’t hate). It’s an incredible puzzle that combines the gimmicks of all six dimension rounds into one cohesive, well-written puzzle.