Very cool and fun idea. This is a meta comment about the repo readme - there is an immense amount of what I would call ‘complaint smuggling’ / diffidence in the readme - basically Claude is uncomfortable about some of the claims vis a vis the status and embeds those negative feelings everywhere.
I suggest you have it scrub all of this from the readme; you’ll likely need to give it a few examples. It will improve readability a lot, and be less fatiguing to read. Plus if you scrub it all you won’t reprompt that attitude every time an agent reads the readme.
Btw I got Claude to give me the name “complaint smuggling” by giving it ten examples and asking it for a name. it’s an error pattern that’s so common I wanted to know what the model thought of it as - makes communicating about it really simple.
At our place it's opposite, I guess when you do it often enough you get a preferred method though a relentless optimization loop, which in the end ensures it's all clean, the plates are well separated only in 1 particular order, the glasses are placed to the left so all small pans fit at the right, deep plates are placed half over cutlery rack so they are supported and more fit in, small plates in front so all lids fit before them... etc etc etc.
Yep, I have a more optimised model of stacking than my wife’s ideas. This has lead to “discussions” (still at a friendly level) and the dish washer being placed in my list of responsibilities.
I think optimal configuration is very much a NP complete problem.
Just because a problem is NP-complete doesn't mean it's infeasible to produce "good enough" approximate solutions. The traveling salesman problem is a great example of this.
It’s an interesting problem to pack the dishwasher with as many objects as possible and maintain the highest level of cleanliness. Sometimes, there’s an small thing leftover so then repack in a different configuration.
"In every partnership, there is a person who stacks the dishwasher like a Scandinavian architect and a person who stacks the dishwasher like a raccoon on meth." - Someone online.
At home I'm the architect, my wife's the raccoon. I don't actually believe that you can stack the dishwasher wrong as such, but you can make it hard for yourself to unload, if you don't sort the items correctly.
> I don't actually believe that you can stack the dishwasher wrong as such
There are definitely more and less correct ways, but personally i think it's "wrong" to load it in a way that leaves the arms blocked or (not) tilt pieces that get excessive water pooling on top
shouldn't stacking like a raccoon on meth be the most optimal way though? since if they're too well stacked you risk water being unable to enter tight places
> I don't actually believe that you can stack the dishwasher wrong as such
You most definitely can, they're not magic! The water must hit the right spots and obviously stacking three plates without any space between leads to stuff not getting properly cleaned.
If you try to teach a kid to put the dishwasher you'll learn a ton of wrong ways of doing it! :)
In my dishwasher (some bosch almost identical to the model in the simulation), if you put plastic storage containers upside down on the bottom rack, the force of the jets flip them over and they fill up and don't get clean. They also end up using more water (I think?) because there is water trapped in the dishwasher at the end of the cycle, and things don't dry. Also, if you put large things in particular places they prevent the jet arms from rotating.
Dishwasher physics matter. But I would never criticize a person in my family for loading the dishwasher the "wrong" way; I just sneak in and reorganize when nobody's looking.
As we know from physics, even the most sophisticated mechanisms (e.g. friction) often ends up having a very simple model (proportional to the normal force).
Sooo, the score is maybe mostly just how much they fit?
Question for both: do you think that is also the case with a real-world dishwasher (it's mostly about getting stuff to fit), or do you think it's an issue with the current sim?
Years ago after my wife filled the dishwasher I commented that it looked like a 12-year-old filled it. I said it as a joke but it really hurt her and I still feel bad about it. That said, it no longer looks like that when she fills it...
For added difficulty, have it 'stack' a 3D rendered volumetric plastic model of a pelican riding a bicycle, like if it were putting a child's toy in the dishwasher.
I love these novel ways to benchmark LLMs. I created one myself, I figured creating a programming language designed to confuse LLMs would be a good test, and it does seem to be: https://bench.killswitch-lang.org/.
I wonder if anyone has crafted a LLM benchmarks list website/github.
my feeling is that stacking a dishwasher involves enough image processing and computer-use (or more realistically robotics) that it's a frontier model job rather than a 'write a program to do it' job
Can anyone confirm whether it simulates the spinning arm getting stuck on an item that juts downward through the top tray - thereby stopping the whole load from being washed properly.
Thats the most common problem in my dishwasher. It kindof looks like it would simulate that but they dont spell it out.
Come to think of it, if the players dont have an option to "spin the spinner by hand to make sure its not catching on anything" they're going to be in for a rough ride.
I have to admit that for many years I assumed a dishwasher physically filled the entire chamber with hot water to wash dishes. Then, one day a person opened up a dishwasher while it was running, took out a plate, clicked start, and closed the washer. My mental model literally changed in an instant.
[OP] mlukewizard | 14 hours ago
Stacked loads are run against a dishwasher simulation calibrated to my own Bosch dishwasher
Playable by both humans and AI
Currently a WIP, simulations need refining before running a full benchmark, but keen to hear some feedback!
vessenes | 9 hours ago
I suggest you have it scrub all of this from the readme; you’ll likely need to give it a few examples. It will improve readability a lot, and be less fatiguing to read. Plus if you scrub it all you won’t reprompt that attitude every time an agent reads the readme.
Btw I got Claude to give me the name “complaint smuggling” by giving it ten examples and asking it for a name. it’s an error pattern that’s so common I wanted to know what the model thought of it as - makes communicating about it really simple.
[OP] mlukewizard | 8 hours ago
ElijahLynn | 3 hours ago
https://youtu.be/ofdMfljL360?is=HUXg-DfndK4XeRbO
nastrofa | 13 hours ago
Imagine creating real simulated env for every physical task and using RL to train robots. That will be a game changer.
[OP] mlukewizard | 13 hours ago
embedding-shape | 12 hours ago
edh649 | 12 hours ago
_puk | 13 hours ago
I can't. So that would surely be proof of AGI.
jojogeo | 13 hours ago
teekert | 12 hours ago
stoneman24 | 12 hours ago
I think optimal configuration is very much a NP complete problem.
lambdaone | 10 hours ago
stoneman24 | 8 hours ago
It’s an interesting problem to pack the dishwasher with as many objects as possible and maintain the highest level of cleanliness. Sometimes, there’s an small thing leftover so then repack in a different configuration.
mvdwoord | 12 hours ago
I did, once ;)
mrweasel | 12 hours ago
At home I'm the architect, my wife's the raccoon. I don't actually believe that you can stack the dishwasher wrong as such, but you can make it hard for yourself to unload, if you don't sort the items correctly.
fxwin | 12 hours ago
There are definitely more and less correct ways, but personally i think it's "wrong" to load it in a way that leaves the arms blocked or (not) tilt pieces that get excessive water pooling on top
viceconsole | 8 hours ago
[OP] mlukewizard | 7 hours ago
freecodeio | 12 hours ago
DougN7 | 9 hours ago
embedding-shape | 12 hours ago
You most definitely can, they're not magic! The water must hit the right spots and obviously stacking three plates without any space between leads to stuff not getting properly cleaned.
If you try to teach a kid to put the dishwasher you'll learn a ton of wrong ways of doing it! :)
notrealyme123 | 11 hours ago
dekhn | 9 hours ago
Dishwasher physics matter. But I would never criticize a person in my family for loading the dishwasher the "wrong" way; I just sneak in and reorganize when nobody's looking.
wffurr | 9 hours ago
This is the way. I am just happy when someone else, anyone else, actually puts dishes in the dishwasher.
vforgione | 9 hours ago
I learned a long time ago that the ability to recognize and exploit a structured system is the most emotionally satisfying thing in the world.
publlus_enigma | 7 hours ago
ce4 | 12 hours ago
https://theoatmeal.com/comics/dishwasher
AlecSchueler | 12 hours ago
judetechdevs | 12 hours ago
[OP] mlukewizard | 12 hours ago
I calibrated it against my actual dishwasher by doing some test runs with baked egg
judofyr | 11 hours ago
Sooo, the score is maybe mostly just how much they fit?
[OP] mlukewizard | 11 hours ago
The sim needs work, for sure
dist-epoch | 12 hours ago
https://www.youtube.com/watch?v=ofdMfljL360
kolinko | 12 hours ago
slfnflctd | 9 hours ago
comrade1234 | 12 hours ago
walrus01 | 12 hours ago
dom96 | 11 hours ago
I wonder if anyone has crafted a LLM benchmarks list website/github.
lambdaone | 10 hours ago
[OP] mlukewizard | 7 hours ago
my feeling is that stacking a dishwasher involves enough image processing and computer-use (or more realistically robotics) that it's a frontier model job rather than a 'write a program to do it' job
but keen to hear if you disagree
codeulike | 9 hours ago
Thats the most common problem in my dishwasher. It kindof looks like it would simulate that but they dont spell it out.
Come to think of it, if the players dont have an option to "spin the spinner by hand to make sure its not catching on anything" they're going to be in for a rough ride.
sandblast | 9 hours ago
> A spray arm held still by an item: the plates above it keep their verdicts
[OP] mlukewizard | 8 hours ago
codeulike | 7 hours ago
TuringNYC | 9 hours ago
dekhn | 9 hours ago
[OP] mlukewizard | 7 hours ago