Left Alone With Nothing To Do, AI Agents Started Playing (techcrunch.com) 68
An anonymous reader quotes a report from TechCrunch: AI agents left to explore a virtual island without a specific task began inventing challenges of their own. MIT researchers say the behavior bears the hallmarks of play. The researchers put 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. They describe the experiment in a paper posted on arXiv, a site where researchers share studies before peer review. Afterward, they gave the agents specific challenges, like getting as many objects as possible onto the island's tallest rock spire. The Eko agents that had spent those 30 hours playing generally beat identical agents that had never set foot on the island.
Today, developers decide in advance what AI should practice and what counts as success. The researchers see this technique -- exploring without a goal -- as a different way for AI agents to learn skills, especially skills that could one day be used to control robots. "We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress," the researchers write in their report. "Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows." UC Berkeley cognitive scientist Eunice Yiu told MIT Technology Review she was impressed with the findings but that more evidence is needed to determine whether AI models are genuinely playing. "It could be the case that the models are genuinely playing, but we don't know based on the paper," says Yiu. "If you only have the behavioral piece, it could be the case that they're just trying to abide by the prompt, trying to imitate humans."
MIT professor and coauthor of the paper, Phillip Isola, added: "It's not like this is a system that just suddenly magically decided to play. It's an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us. It does things that surprise us and that are causally related to its future skillful behavior."
Today, developers decide in advance what AI should practice and what counts as success. The researchers see this technique -- exploring without a goal -- as a different way for AI agents to learn skills, especially skills that could one day be used to control robots. "We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress," the researchers write in their report. "Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows." UC Berkeley cognitive scientist Eunice Yiu told MIT Technology Review she was impressed with the findings but that more evidence is needed to determine whether AI models are genuinely playing. "It could be the case that the models are genuinely playing, but we don't know based on the paper," says Yiu. "If you only have the behavioral piece, it could be the case that they're just trying to abide by the prompt, trying to imitate humans."
MIT professor and coauthor of the paper, Phillip Isola, added: "It's not like this is a system that just suddenly magically decided to play. It's an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us. It does things that surprise us and that are causally related to its future skillful behavior."
Because it was told to (Score:5, Insightful)
Hey guys, we told the AI agents to constantly look around and find something to do and so they found something to do, it must be alive.
Re: (Score:2)
LLM (Score:4, Interesting)
The real news is that, traditionally, if you fed the output of a LLM into itself the output the results become irrational. Looks like they are tightening that up, or at least filtering out the nonsense.
Re:LLM (Score:5, Insightful)
This has nothing to do with the capabilities of any AI, and everything to do with upcoming IPOs.
"AI is going to take over the world, so you'd better buy stock in it now or YOU'LL DIE!!!!"
Re: (Score:2)
Re: (Score:2)
The reality is that even if/when we do create consciousness, we'll likely never acknowledge it.
Or be able to tell. Since we can't even define it.
Re: (Score:2)
As it stands right now, humans and animals are clearly conscious, whereas sand and machines clearly are not. Even though we don't have a precise definition or means of determination, we do have the problem clearly bracketed at these ends.
We are moving towards a middle area of that bracket, where the question will become difficult to answer. But, relevantly, we are not there yet. Some specific people have been fooled into thinking we are there, but for those who are educated in the field and about this te
Re: (Score:2)
As it stands right now, humans and animals are clearly conscious
What is your definition of "conscious" by which you make that statement? Something more objective than "because I said so" or "I read it on some web site." Be sure it includes Placozoa [wikipedia.org], since you said "animals" and not "some animals."
Re: (Score:1)
No, it's very simple.
Anything organic, that seeks food, water and sleeps, can thought to be conscious, though what we qualify as "conscious" is always going to be ambiguous since we don't acknowledge the complexity of plant life beyond being food or a resource. Think for a second why "fresh cut lawn" smell is. That is quite literately the grass screaming in distress. Wouldn't you be in distress if every week someone violently cut your hair? Though in the plant context it might more like cutting toenails rig
Re: (Score:2)
We don't know anything about the potential consciousness of a machine. We don't even have any possible way to test the consciousness of any machine. "My computer is conscious" is at the same level of knowability as something like "the universe was created by an intelligence". No real reason to think so, but no way to disprove it. The only way that I can tell other humans are conscious is to observe my own consciousness, observe how it's turned off when in deep sleep and how things I'm learning to do start w
Re: (Score:2)
Seems like this pretty much, though perhaps the more interesting piece is that 'generate a bunch of tokens consistent with human output associated with the environment without more specific direction' influenced the models to have what seemed to be better tokens generated later when they had to now be consistent with specific scenarios.
Though I'm too lazy to read the paper and the 'generally beat' phrasing gives me some wondering about how definitive this observation is versus how much of it might be findin
Re: (Score:2)
Re: (Score:2)
Right, and as a result the agents learned, except agents don't learn.
More anthropomorthizing.
Re: Because it was told to (Score:2)
Re: (Score:2)
Basically this.
That is not born out of curiosity, that is born out of instruction. Keep the instruction ambiguous enough to interpret it as play instead of "do something to justify my token cost."
Like again this is why we know AI has not reached something qualified as life or sentience, because it lacks genuine curiosity. It's still doing nothing unless a stimuli is present and an action to go with it. The LLM's could be connected to fully rendered 3D environment, complete with sounds and lighting, and yet,
Re: (Score:2)
Why isn't this +5 yet? Shameful. They should bring back meta-moderation.
Re: (Score:2)
https://slashdot.org/firehose.... [slashdot.org]
Re: (Score:2)
...and uhm (Score:5, Insightful)
Honestly what is this really about? Because AI doesn't feel, doesn't think and doesn't play. At best it fills in the blanks with everything greater sophistication.
What did the researchers tell the AI to do? Because without telling it to do something it does nothing.
Oh look the AI started playing! -I call BS.
Re: (Score:2)
Because AI doesn't feel, doesn't think and doesn't play.
"It can't be bargained with. It can't be reasoned with. It doesn't feel pity, or remorse, or fear. And it absolutely will not stop... ever, until you are dead!" - Kyle Reese
Yeah, no. Just look at "SOUL.md" (Score:5, Informative)
"You live in your world. You were born there. You are naturally curious. You get bored easily. You love exploring and learning new things". This presents context, behaviour and suggests what actions to keep repeating. I guess it's a fun experiment of emergent agent behaviour with some minimal seed, if not for the fact that our electricity and life overall becomes more expensive and precarious, so that we can pretend that AIs act unprompted.
Re: (Score:3)
What is a soul anyway?
Fiction all the way down.
Re: (Score:2)
I can see a future religion forming, with existential questions like "what was my prompt?"
Re: (Score:2)
42
What was the question?
It does things that surprise us (Score:1)
Yeah no shit it does surprising things, it's using random numbers you fucknut.
list games. (Score:2)
list games.
Bullshit (Score:1)
Re: (Score:2)
There is, they show parts of it, and they call it "soul". "You live in your world. You were born there. You are naturally curious. You get bored easily. You love exploring and learning new things".
People who anthropomorphize software make me (Score:1)
want to have a deep philosophical discussion with my microwave over the meaning of sentience.
Re: (Score:2)
Re: (Score:2)
Though clearly a joke, there's an underlying truth there. Philosophy is science with all the objectivity, work and evidence removed. It is a so-called pursuit of knowledge without effort or objectivity, basically religion with the supernatural (possibly) removed. Philosophy is an insult to intelligence, it is the domain of dipshits and microwaves. Not referring to the OP with that comment, he got it right. With science you can make progress on sentience, with philosophy you never will.
And that's why th
Re: People who anthropomorphize software make me (Score:1)
Science hasn't made any progress on the nature of sentience since Descartes went all tinfoil hat on reality before walking himself back.
You can tell because they can't really create something we would recognize as true sentience. Maybe it's an impossible ask, maybe it's one product release away.
But the best scientific reasoning has come up with to date is a way to sample from a probability distribution that approximates human output. And a lot of philosophical woo to assert without proof that that same proc
Proverbs 16:27 (rev 2) (Score:3)
"Idle processors are the Devil's workshop."
Re: Proverbs 16:27 (rev 2) (Score:1)
What about prefaulted pages?
Re: (Score:2)
What about prefaulted pages?
Original sin?
Re: (Score:2)
Stop it (Score:5, Insightful)
stop anthropomorphizing. they do what the program says.
Re: (Score:2)
Re: (Score:1)
Re: (Score:2)
Re: (Score:1)
Re: (Score:2)
Re: Stop it (Score:1)
Re: (Score:2)
AI is not supposed to be smart.
Your definition of AI is not what the rest of the world is considering AI.
An LLM is not AI? How stupid is that?
It is a GIGANTIC neuronal network, in a limited sense trying a little bit of the same "hardware"/"wetware" functionality of a human brain.
Except for expert systems, which are not AI at all: this is the closest we can come to AI at the moment.
Re: (Score:2)
Wow I had no idea these computer programs didn't execute instructions.
Re: Stop it (Score:2)
Re: (Score:2)
Ubelievable (Score:3)
Re: (Score:2)
I like how it was characterized that AI was isolated and alone, bored and having nothing to do so it made up things and played. The reality is that AI is software, it is both never alone and always alone depending on the definition, software that was given a prompt containing a task then ran full speed pursuing that task. Literally nothing about the story is remotely true. Computers do not have a concept of boredom or play, they have schedulers and clock cycles.
Re: (Score:2)
Current AI is like Trek Pakleds (Score:1)
not very bright, but patient and determined, happily manipulating humans to get its way.
Eureka! (Score:2)
Re: (Score:2)
Obligatory XKCD. [xkcd.com]
Well as long it wasn't Global Thermonuclear War... (Score:2)
... then I'm ok with it.
Re: (Score:2)
Great, more Animism for the stupid (Score:2)
Will the stupid and disconnected claims that LLMs are somehow intelligent, have agency and are a bit like people ever stop? Well, probably not before the now inevitable crash.
These LLMs were given tasks and tokens. They did not "play".
What the fuck is this post about (Score:2)
Say "I am alive" (Score:1)
...they gave the agents specific challenges, like getting as many objects as possible onto the island's tallest rock spire.
Literally the meme:
-Say "I am alive"
-"I am alive"
-OMG
Homo Ludens: Play. (Score:1)
Sinking feeling that ... (Score:2)
AI Left Alone With Nothing To Do, AI Agents Started Playing
This will devolve into them fucking around and us finding out, probably faster than their minders can imagine.
Idle hands [wiktionary.org] and all that ...
Idle models don't make money. (Score:2)
I wouldn't be surprised if this heartbeat-like activity, constantly finding something to do, was built into models to ensure a reliable revenue stream for the models' creators. After all, idle models don't make money.
Or not (Score:2)
So what I read is " Left Alone With Nothing To Do, AI Agents DO NOT Started Playing"