Merged Artificial Intelligence

[Beelzebuddy]

The agentic AIs in that video weren't told what to do, they were just told "get bored and do stuff".

That's telling them what to do. They rolled the stochastic completion model for what should come after "get bored and do stuff," then rolled again for how to do the thing they rolled. It's just spicy autocomplete. Do you think they would have gotten bored and done stuff if they weren't explicitly instructed to do so?

That's different. This is a new kind of tool. A disruptive technology. You can't handwave away everything that makes it disruptive by saying "oh it's just another tool wielded by humans". That argument doesn't fly.

It's disruptive, yes, because the language modeling (and more generally trained association) means you can throw a natural language task at a pile of RNG and get a coherent result. That's flippin' amazing. It does not mean it has agency of its own.

 

[arthwollipot]

That's telling them what to do. They rolled the stochastic completion model for what should come after "get bored and do stuff," then rolled again for how to do the thing they rolled. It's just spicy autocomplete. Do you think they would have gotten bored and done stuff if they weren't explicitly instructed to do so?

It's explicitly not "telling them what to do". The only instruction given to the AIs was to get bored and do stuff. The researchers explicitly and deliberately avoided telling them how to fulfil that single instruction. And the AIs came up with lots of ways to actually do it.

It would be like me giving an instruction to Photoshop to "modify pictures" and waiting to see what it did.

It's disruptive, yes, because the language modeling (and more generally trained association) means you can throw a natural language task at a pile of RNG and get a coherent result. That's flippin' amazing. It does not mean it has agency of its own.

Did you spend the 15 minutes watching the video? If you didn't, I think you should. If you did, what did you think about the fact that an AI set itself a problem (to jump to the top of the spire) then worked out how to actually do it (after discovering that its jump distance wasn't high enough it taught itself how to stack blocks so that it could reach the top, then celebrated its success in the log)?

 

[The Great Zaganza]

It's just a fancy version of the Game of Life.
Unexpected patterns will emerge in absence of guidance.

Please keep in mind that people only attribute a level of sentience to AI doing stuff we think we can recognize and understand - no one calls protein folding LLM or investment LLM sentient.

 

[arthwollipot]

Nobody is saying that the AI agents in the video are sapient. But nobody could predict what they would do. They acted in ways that were unanticipated. That's not following a programme, that's actual independence - agency.

 

[The Great Zaganza]

But that's exactly what happened with Conway's Game of Life.
It's just a feature of systems allowed to evolve that patterns will emerge.
It's a strong support for the idea that some kind of Life will emerge naturally, given time and opportunity.
It's not supporting the idea that LLMs have any agency in the slightest.

 

[Dr.Sid]

It's just a fancy version of the Game of Life.
Unexpected patterns will emerge in absence of guidance.

Please keep in mind that people only attribute a level of sentience to AI doing stuff we think we can recognize and understand - no one calls protein folding LLM or investment LLM sentient.

I still haven't heard even half decent definition of what sentient means.

 

[The Great Zaganza]

Or Agency

 

[arthwollipot]

The technical difference between sentient and sapient:

If an organism is sentient it means that it can perceive and react to the world around it. It's capable of sensation and is thus usually considered conscious. A dog is sentient.

If on the other hand an organism is sapient (literally, "wise"), it means that it has a level of self-awareness and self-reflection, it understands that it is an organism in a universe, it is capable of reasoning and metacognition. A dog is not sapient, but humans are, which is why we are Homo sapiens.

"Agency" means that one can exercise one's own will in determining the course of one's actions. A dog has agency, a hammer does not. A dog chooses what pile of poo to sniff. A hammer does not choose which nail to bang. Crucially, whether an organism demonstrates agency is separate from whether it demonstrates either sentience or sapience.

The agentic AIs in the linked video clearly demonstrated agency. Hence the name, I guess. They were given a minimal set of instructions - "get bored and do stuff" - and were put in a world where the laws of physical movement were defined, but they chose what stuff to do. No programmer instructed them to stack blocks, no programmer instructed them to climb to the top of the spire. They chose their own course of actions in order to fulfil the single instruction that they had been given. And they celebrated in the logs when they achieved these self-given goals.

Nobody is suggesting that the AIs demonstrated sapience. It could be reasonably argued, though I am not here doing so, that they demonstrated sentience. It is undeniable that they demonstrated agency.

No AI was used in the composition of this post.

 

[arthwollipot]

But that's exactly what happened with Conway's Game of Life.
It's just a feature of systems allowed to evolve that patterns will emerge.

Yes, it is. But Conway's Life was much, much simpler. Orders of magnitude simpler. There was no agency there, there was only a rule-based binary response. Given one state, you could easily predict the following. That's the difference. Nobody could have predicted that an agentic AI would choose to try to jump to the top of the spire. Many iterations didn't. They did other unpredictable things.

 

[Beelzebuddy]

Did you spend the 15 minutes watching the video?

I have a confession to make. I don't watch these videos you keep posting. The one time I did it was most of an hour of "content" that could have been summed up in two sentences. This one in particular also has a face in the thumbnail, which is a red flag all on its own.

You should try interacting with AI models yourself instead of letting youtube do your thinking for you. If you do, I think you'll find they're great at interpreting vague prompts like "get bored and do stuff," which remains an explicit instruction. The model will fill in the gaps. Similarly, if you tell chatgpt to read you a bedtime story it will make up a story for you. The trick is getting it to make up a good one.

Oh, and

And they celebrated in the logs when they achieved these self-given goals.

They are fantastic at autofellatio. By the tenth time you see a coding agent smiling over a pile of smoking garbage with a log message like " Code complete. All tests passed! " the charm will have worn off, I promise.

Yes, it is. But Conway's Life was much, much simpler. Orders of magnitude simpler. There was no agency there, there was only a rule-based binary response. Given one state, you could easily predict the following. That's the difference. Nobody could have predicted that an agentic AI would choose to try to jump to the top of the spire. Many iterations didn't. They did other unpredictable things.

No one predicted the gliders, bullets and all the other self-propagating patterns in Life. Despite each state being a product of the previous state and a handful of very simple rules, marvelously complex behavior emerged from it. The rules and emergence differs, but it's fundamentally the same thing happening here. Just as the Game of Life isn't life, this version of Artificial Intelligence isn't intelligence. It's a big step along the path, but we're not there yet.

 

[Lithrael]

People can and do write scripts that ask Photoshop to 'do whatever' to create glitch art. Usually you run it a bunch of times and then look thru the output for visually interesting results. In this context 'do whatever' means things like, randomly pick and use the options available in photoshop. Gen AI has access to a lot more options in their 'whatever' experiments.

 

[steenkh]

You should try interacting with AI models yourself instead of letting youtube do your thinking for you

You should stop thinking that everybody can do these experiments for themselves. This is simply not the same as fooling around with prompts.

No one predicted the gliders, bullets and all the other self-propagating patterns in Life. Despite each state being a product of the previous state and a handful of very simple rules, marvelously complex behavior emerged from it.

You recognise that emergence takes place, but because Life does not evolve anything like intelligence, you claim that nothing else can.

 

[Trebuchet]

Donald J. Trump
@realDonald Trump

When will The Wall Street Journal start calling Artificial Intelligence (AI), a very bad and inaccurate name, SI (Super Intelligence), the name which virtually everyone else is using, or soon will be.
In short order, we will be officially BANNING the term, "Al/Artificial Intelligence," in all relevant documents, and everything else, due to its inaccuracy.

Thank you for your attention to this matter.

President DONALD J. TRUMP

Guess we should report the thread and get the title fixe!

 

[Roger Ramjets]

No programmer instructed them to stack blocks, no programmer instructed them to climb to the top of the spire. They chose their own course of actions in order to fulfil the single instruction that they had been given.

In exactly the same way that Conway's Game of Life generated things that the programmers didn't 'instruct' it to do. And yet they did, by providing the initial conditions for the patterns to emerge.

A single instruction is still an instruction. That the machine was able to carry out that instruction without being guided every step of the way doesn't imply that it had 'agency' in the human sense, it just means it was able to change it's behavior depending on the data. By this definition a thermostat has 'agency'.

But the AI fanatics will grasp at any opportunity to anthropomorphize their creation, because that's their goal.

And they celebrated in the logs when they achieved these self-given goals.

No, they didn't.

Nobody is suggesting that the AIs demonstrated sapience. It could be reasonably argued, though I am not here doing so, that they demonstrated sentience. It is undeniable that they demonstrated agency.

Some are. Others are suggesting that AI is on the cusp of being sapient. And they have good reason - AI is being trained on human thinking and behavior, and being guided towards acting like a human. Eventually the mimicry will get so close to the real thing that people won't be able to tell the difference. The effect that will have on our minds is very concerning.

 

[arthwollipot]

I have a confession to make. I don't watch these videos you keep posting. The one time I did it was most of an hour of "content" that could have been summed up in two sentences. This one in particular also has a face in the thumbnail, which is a red flag all on its own.

The video I posted is about 15 minutes long, so it's less investment of your time than Forumvision is, and it goes in detail about how the AI agents behaved and exactly how and why they demonstrate agency. I've spent considerable effort now trying to summarise it and with the above confession I am officially done. Either you want to know, and you'll spend 15 minutes learning something, or you don't want to know and you won't. I'm done here. Good day.

 

[arthwollipot]

No one predicted the gliders, bullets and all the other self-propagating patterns in Life. Despite each state being a product of the previous state and a handful of very simple rules, marvelously complex behavior emerged from it. The rules and emergence differs, but it's fundamentally the same thing happening here. Just as the Game of Life isn't life, this version of Artificial Intelligence isn't intelligence. It's a big step along the path, but we're not there yet.

The difference is that anybody by looking at the current state of the game and with knowledge of the rules, could easily and trivially calculate what the next state would be, which means that yes, the existence of gliders and bullets can be and was predicted.

And again, nobody is arguing that agentic AI is intelligent. Just that it exhibits agency.

 

[arthwollipot]

Some are. Others are suggesting that AI is on the cusp of being sapient.

You mean people like Blake Lemoine? The guy who was fired from Google and became a laughingstock because he went public about his belief that his AI chatbot was sapient because it was instructed to imitate a sapient entity as closely as possible? You mean that guy?

 

[Roger Ramjets]

If the problems are the same as in every large project, regardless of whether it uses AI tools, then why would someone blame those problems on the use of AI tools?

Problem is people are expecting AI to be more than a tool. Only after they have been using one for a while do they begin to appreciate the truth. As Beelzebuddy said, the larger a project gets the harder it is for the AI to keep track of it. A human can because they have a persistent memory and understanding of the concepts and goals they are working towards. What typically happens is the AI gets about 75% of the way there and starts struggling, then the developer gets bogged down trying to goad it into completing the job. After a while they give up and release the software in an unfinished state that humans have difficulty maintaining, ie. slop.

It seems to me that one important attribute of a solid software architecture is that it has been designed so you don't have to keep all 800,000 lines of code in your head as you're working on one localized module. That attribute is important for software designed and maintained by humans. Why wouldn't it be equally important for software designed and maintained by humans aided by AI?

The problem is that the AI often does have to keep large amounts of code 'in it's head'. Localizing modules enough to fit in the AI's 'cache' may lead to inefficient code, as well as making the developer do more work. You're not doing the actual coding but you are spending a lot of time defining modules. Human coders often start with a basic concept and flesh it out as they go. AI doesn't like that, it needs everything defined beforehand because it isn't good at visualizing things that may change as the code is built up.

I'm sure it will get better, but right now many developers are going back to coding stuff themselves because the productivity gains of 'Vibe' coding are evaporating. That doesn't mean AI tools can't be useful, just that expecting AI to do the whole lot may be overambitious. Unfortunately it's attracting non-coders too. They try something simple and have success, and think they are coders now. Then they try making a larger project and soon get out of their depth - and the AI is no help because it's just as lost as they are.

 

[alfaniner]

Donald J. Trump
@realDonald Trump

When will The Wall Street Journal start calling Artificial Intelligence (AI), a very bad and inaccurate name, SI (Super Intelligence), the name which virtually everyone else is using, or soon will be.
In short order, we will be officially BANNING the term, "Al/Artificial Intelligence," in all relevant documents, and everything else, due to its inaccuracy.

Thank you for your attention to this matter.

President DONALD J. TRUMP

I wonder if anyone has asked Ai itself what it wants to be called?
I'd hope the answer would be "You can call me Al".

Note: that is a lower-case "L". Modeled on the few occasions my own name has been read as "Ai".

 

[Roger Ramjets]

You mean people like Blake Lemoine? The guy who was fired from Google and became a laughingstock because he went public about his belief that his AI chatbot was sapient because it was instructed to imitate a sapient entity as closely as possible? You mean that guy?

That was 4 years ago. A lot has happened since then. Now AI is officially 'Super Intelligent' and outrageous claims are being taken seriously. Any skepticism of its abilities is dismissed by the AI fans as being out of date or soon will be. And they have a point. So long as the bubble doesn't burst and progress continues, nobody knows how advanced it might be in a decade.