3K at 45%

Lmao is this entire market just @AdamK against every puzzlehunter?

I wonder if they solve hunt puzzles

@Soni Keep the mana flowing!

@AdamK Let’s see some more of those 50% limit orders!

Potentially of relevance to folks: https://scale.com/leaderboard/enigma_eval

@MoyaChen Ha! Thanks for sharing! I’m literally the author of the first example puzzle listed in this data set

@MoyaChen Funny you’d mention this eval…

As an aside - if this market has not been resolved as "YES" and the last MIT Mystery Hunts (ie 2029 and 2030 ones) are not run, this will be resolved as "N/A", assuming an alternate hunt with the same general spirit as Mystery Hunt cannot be found.

Putting aside whether or not ML models will be capable of solving at a top level (which I have severe doubts about), I think the 5-human limit here is going to be an incredibly limiting factor due to the physical nature of the MIT Mystery Hunt — a physicality which, I might add, I think is only going to ramp up further over the next 5 years. Events, physical puzzles, interactions, runarounds, even something like this year's open HQ are going to make it exceptionally difficult for a team of 5 bodies.

@Marnix A model giving directions to a human team seems allowed for a YES resolution. The model can just tell people what to do. It seems like the question author is fine with this:
"If it's 2030 and it seems like ML capabilities are getting close to doing this, I will field a team to do so :) "

I'm talking purely logistically here - I don't think a 5-person team is feasible, no matter how capable the AI behind it is. As someone who was on a relatively small onsite team (roughly 15 people onsite, 30-40ish offsite), I think 5 people in person is simply not enough bodies to engage with the physical aspects of this puzzle hunt on the ML model's behalf. MIT has a big campus.

This is especially true if you consider that:

• The 1 AM on-site closing time still applies, giving those humans a time limit for hunt-critical onsite puzzles every day,

• Wifi on the MIT campus is kind of Ass (introducing fascinating roadblocks for communication with the ML model whenever people are outside of team HQ),

• and that the human people on this team will eventually have to sleep (meaning you will either have no one getting the AI to work on puzzles overnight OR fewer than 5 people capable of doing onsite work during the day).

You can have a whole set of super-puzzle-capable ML models, but one of the reasons the top teams are also big is because puzzle hunts like this aren't just based on puzzle capability, but puzzle capacity. There are logistical challenges here that i think a 5-body team would struggle to get past (and that's before we get into the fact that I don't think we'll have a super-puzzle-capable ML model by 2030).

@Marnix Would be interesting if 5 people are physically incapable of doing the puzzles (maybe 6 buttons spaced far apart that have to be pressed simultaneously?), but I’ll take my chances on ML capabilities reaching the point where they could puppeteer a team to completion in principle. Feel free to fill my limit orders if you think otherwise.

This is a good conversation worth having. I'll put a think towards exactly how to update the "5-people involved" part to capture the intent here. My current thoughts are either:

1. If there are puzzles that necessitate having more than 5 people to complete (ie the "6 buttons" case) those are not counted in the denominator for the "80%" calculation

2. To allow some sort of "more than 5" as long as it's even that much more clear cut that it's the model doing all direction setting .

3. Lowering the 80% to 70%, as a hack-y patch that gets the gist.

> Events, physical puzzles, interactions, runarounds, even something like this year's open HQ are going to make it exceptionally difficult for a team of 5 bodies.

For what it's worth, having been on the writing side in the past (and also having seen things like Caltech Ditch Day), anything in person requires a lot of extra logistics + coordination, and very real time + $$ to set up. To this end, I do imagine that the bandwidth of the organizing team will limit how many physical interactions they can do (at least, without harmfully impacting other parts of their hunt logistics) which is why "80%" seemed like a reasonable bound when I first wrote this.

Still worth tightening up the wording here nonetheless.

(Among other things, I doubt it'll be possible for an agent to do a physical scavenger hunt.)

> Wifi on the MIT campus is kind of Ass
Don't underestimate the ability of a random corporate entity (as many of the big labs are lol) to put Wifi repeaters on campus if they want to hit some goal.... (Also on-device models lol?)

@MoyaChen Must the team use a single model? Or may the humans pick and choose among several models? E.g. audio puzzles are given to the SOTA audio system, visual puzzles are given to the SOTA visual system, text puzzles given to the SOTA text system, etc.

Or would that decision of what model to use need to be made by the model?

@JimHays : Decision of that would should made by the model, unless it's a form of interaction where a human needs to help with input.

(Ex. If a human needs to turn on camera stream or a microphone and presses some buttons to configure the ML model to accept it, that's fine. A human deciding "oh we're going to feed this web based puzzle to the grid-logics module because it's a grid logic!" would not be ok though.)

Must the hunt be solved during the normal time parameters of the hunt, concurrent with human teams and be acknowledged by the running team as a winning team at wrap up (assuming such a list of winners is given)?

@MoyaChen

I'm actually going to say "no" to "acknowledged by running team" in specific. That said, will say "yes" to "acknowledged by some combination of Puzzle Club, MIT admin, or traditional media".

@MoyaChen The time parameters bit is important too, imo - we're discussing an MIT Mystery Hunt being done during the normal timeframe of Noon on Friday to ~8:00 Sunday on hunt weekend, right? An old hunt wouldn't count.

@Marnix : Yeah intent is that this is a new, out-side-of-previously-seen-training data hunt.

How do you plan to resolve this if AI are banned?

@resf : Resolves as "NO".