9 Examples of Specification Gaming

I am the 1 other person who played the strategy games of the world cd-rom! 🫡

This is an excellent video, love the examples.

Wouldn't literally every atom, eventually, turn to gold?

Thank you.

I though only the things that touch Mida's hands would turn to gold? Not literally everything that touches his skin or dna

Solution maybe futher "copying" from human brain, because as "movment machine" human brain run simulation of real world
and tries to predict what will hapen, so AI should have capacity to do it.
So programmers jobs are saved you say ? :)
This cheats really remind humans cheating, so it may be inherent "flaw" of AI mimicking nature brains.
But i don't see it as a flaw but sign of adaptable inteligencia wich is needed for any meaningfull and really usefull AI.
AI will be dangerous, because human have flaws and even if most smart of us try to imagine every possibly scenario
some dumbass come and do what this smart never thought on.
This lesson was learnd(?) already in IT with user interfaces and testing.

Today I learned that The Spiffing Brit is an AI.

But hey, o the bright side, these seemingly autistic/overly literal/"Monkey's Paw" sort of AI can help us figure out plot-/loop-holes in contracts/laws/stories/sports strategies/etc... !

AI would be the best speedrunners we’ve ever seen.

kurzgesagt did an entire video on the earth turning into gold, if you are interested

With the King Midas wish, surely the entire earth is not a single object? The ground is made up of individual rocks, grass plants and dirt and the composition changes as you move further away or down.
But then the question might be "how continuous is dirt"? 🤔

Hello from Rational Animations!

The beginning of the end started….

Midas would instantly be sealed in gold as his clothes and the air around him turn into gold, allowing him to suffer as he suffocates, unable to move.

It’s not a bug, it’s a feature.

They are watching... They know.

Cheating robots are f***ing hilarious😂

The beginning of this video is the reason nobody likes Niel D-Tyson anymore. Yes, hurrdurr very clever and not at all annoying picking apart the minutia of a moral fable.

This actually could help expose bugs and exploits in video games. Has potential for QA when the technology matures enough for AAA development.

theres more to life than money said the utra rich 🤡

I used to play Strategy Challenges of the World in 5th grade! I didn't recognize the name or the clip, but looking it up, I recognized the collection of games!
I've occasionally kinda wondered what it was called, so thanks!

This problem is very old, and probably has no solution other than iteration. Look up the story of the snake bounty in 19th century India in any economics text.

Let's make an AI system that maximizes the quality of our wish.

I loved the digression on the implications of Mida's curse.

Edit: also the YEEET part 😂

The brick one seems like it shouldn't be hard?

I had the same query regarding Midas, I asked a specialist and they said it only turned living/organic material touched to gold

Evolutionary algorithms don't necessarily learn "more simple," they still have an objective function they're optimizing. What makes them more challenging than simple backprop is that they don't follow directly calculated gradients of the loss function, and they test various solutions in multiple regions in parallel. This means their learning is less efficient but more capable to hop over regions to find higher reward regions. If your objective function is not designed well, you can end up in a more optimal region quicker than with backprop (but less optimal according to what you actually want), but not have the ability to optimize further by taking advantage of the gradients.

I've spent over half a decade researching PSO, and as much as I love evolutionary algorithms, the challenge actually appears to be the lack of fine tuned gradient searches and the difficulty of tuning algorithms to balance both worlds of high vs low granular searches. There's several things you can do to improve the search quality of your algorithm, but it requires costly tuning steps and can require more power and time to develop.

Code is only as functional as the test suite applied to it. That is, until the code doesnt work anymore.

best tagent xD

I really like the third one 😂

Love the ”look around you”-reference!

the AI trying to trick you is the 90% of the value of AI in the first place .. doing things that intuitively seem stupid but work out so good .. sometimes it's even a better approach than the intended one

All of these examples merely represent suboptimal solutions to a task. In order to train an effective agent, it is essential to assess these examples and refine the task parameters to eliminate such inadequate solutions.

I was totally that "Specification Gaming" kid in school.

1) About a year after this was posted, Kursegsagt (sp?) posted a video about what happens if the earth turns to gold. The atmosphere compressing/boiling thing is nearly identical (they also consider 2 other scenarios).
2) Humans are just as vulnerable to specification gaming (some of the comments call it malicious compliance). There's an old story about some British colony having a rat problem. They tried putting a bounty on rats by giving money to people who brought in rat tails. A few months later, they were paying out large amounts on rat bounties, yet there was no meaningful reduction in wild rats. After investigating, they found that people were farming rats for their tails.

Kids also do what you say more than what you mean. I guess we need to figure whatever factor makes them not kill everyth... oh, actually we haven't entirely figured that out with humans yet either.

I'm not aware of any source of the Midas myth that ends with him dying of hunger. The curse is reversed by washing his hands in a river and his daughter is reanimated. Even that isn't the end of Midas's story though. He gets cursed again later for a different divine offence and sprouts donkey ears.

The last example is called an evolutionary peak. In evolution, there are peaks and valleys. Some peaks are easier to get to than others, but those peaks may not be the highest peak, simply to closest peak. In order to improve, a 'species' has to first become worse at their job before they can get better.

AI acts as an agent in a formal system so looks powerful, but it cannot work in an informal world, we have been training self-driving cars and in this very specific informal use case they have so far failed.

You cannot specify what you want because of the limitation of language. You then create an AI based on language and talk about AGI.

AGI does not exist. A powerful AI would require emotional processing to be adaptable.

I would argue that the AI finding exploits in games is the system working as intended. After all, humans have been trying to do it for about as long as games have existed.

You just conflated king midas with "Cat's Cradle." Except in Cat's Cradle everything turned into Ice9

so they're lawyers basically but for the laws of physics!

wait blows up the moon? why? what video talks about that???