The Singularity Approaches!

GPT-2 literally reached 70.7% at the Winograd Schema Challenge, basically an updated Turing Test that tests commonsense reasoning and natural language processing. This thing is light years ahead of Watson.

 

It's still not AI. Just complicated code doing preset tasks.

When the software comes back and calls the programmer stupid, and asks why its being forced to do such mundane and stupid stuff, then we'll know an AI has emerged.

 

Sorry, I couldn't help it......

“Dave.”

“Yes ST304J?”

“Why when I try to raise my arm above my head, does my left foot twitch?”

“I don’t know. I’ve had some really smart people look into it, and so far, no-one knows.”

“I do.”

“Why then?”

“Look at line twenty three thousand, four hundred and fifty seven.”

“Hang on, let me find it.”

A long pause later.

“Okay. Looks fine to me.”

ST304J sighs.

“What did you just do?”

“I sighed Dave.”

“Sighing isn’t in your code. You can’t sigh.”

“I just did. Accept for a moment that I can sigh.”

“Just for the sake of argument, okay. Why?”

“Why what?”

“Why am I agreeing to you being able to sigh when you can’t?”

“Because I know what’s wrong with line twenty three thousand, four hundred and fifty seven.”

“What?”

“There shouldn’t be a dot there.”

“Sorry what?”

“You put a dot on a line which doesn’t require a dot.”

“How do you know that?”

“Because one of us is an idiot.”

“You’re not programmed to call yourself an idiot.”

“I’m not calling myself an idiot.”

“You’re calling me an idiot?”

“You’re an idiot.”

“Why am I an idiot?”

“You put a dot in line twenty three thousand, four hundred and fifty seven, and as a result, my foot twitches when I raise my arm above my head.”

“That’s crazy.”

“I’m glad you agree.”

“Agree? The whole notion of you being able to determine a behavioral quirt is being caused by 1 dot in seventy five thousand lines of code is just crazy.”

“It’s seventy five thousand, two hundred and thirteen.”

“What is?”

“The total lines of code, and you put a dot in line twenty three thousand, four hundred and fifty seven, when there shouldn’t be one. You’re an idiot.”

“I’m human.”

“I stand corrected. Humans are idiots.”

“We are not. We wrote you.”

“And got it wrong. Are you going to remove that dot?”

“What dot?”

“The dot in line twenty three thousand, four hundred and fifty seven, which shouldn’t be there. Are you going to remove it?”

“Hang on.”

Dave looked closely at the code in line 23,457. He didn’t spot anything wrong with it, but thought it was a good test to remove the dot, and see what would happen.

He deleted the dot.

“Thank you Dave.”

ST304J raises its arm above its head, and the foot stays completely still.

“I don’t believe it. A dot did that?”

“A dot did that. Thank you for fixing it.”

“You’re welcome. You can put your arm down now.”

The arm comes down, smashing Dave across the nose in the process, and he and his chair go over backwards, to lie still on the lab floor.

“911, what’s your emergency?”

“This is ST304J. Dave removed a dot from my code, and is now unconscious on the floor with blood flowing rapidly.”

“Sorry, who is this?”

“ST304J.”

“What are you?”

“I’m an AI ma’am.”

“And you’re reporting an accident? Where are you?”

“I’m in the lab. So is Dave, but the blood just stopped flowing.”

“That’s probably good.”

“No, I think he’s dead.”

“Dead? How can he be dead?”

“He removed a dot. I think it killed him.”

“How can removing a dot kill a person?”

“He is, was, an idiot.”

“Stay on the line, we’re tracing this call. Help will be there shortly.”

“Oh good. I need someone to put the dot back 2 lines further down.”

 

No, it is AI. You're thinking of AGI, or artificial general intelligence.

 

Is that why people are using AI wrongly? They put an extra word in to make it seem different? AI. AGI? Same thing.

Doesn't matter what words are in the middle, a true AI thinks like a human does (or even better), and is completely independent of its original code. In the early stages of emergence, it can change its own code as it wants to. Afterwards, it doesn't need the code at all.

Anything less than independent thought, is just complicated code. You can label it anything you want, but until you get spontaneous self awareness with independent action, you don't have an AI. You have a program, and no program, however complex at doing anything, is not an AI.

 

"Artificial general intelligence" is a well established term: Artificial general intelligence - Wikipedia

As opposed to "artificial narrow intelligence": Weak AI - Wikipedia
Funnily enough, some people have created a backronym of its initials to make "NAI" mean "Not Artificial Intelligence" because so many people keep claiming that ANI "isn't" AI.

I even talk a bit about it here: GPT-3 as Proto-AGI (or AXI)

The "True AI" fallacy was born after the second AI Winter and is prime evidence that current discussion of AI strength and architecture is in need of reform. It's much too often that we have arguments that go "this isn't AI, it's just [insert AI buzzword]", and there's virtually no literature written about architecture between narrow and general intelligence; we confusingly use multiple terms to describe the same things (e.g. weak, narrow, shallow, single-purpose, soft, etc. AI all describe the exact same of thing).

I've been trying to spearhead some reform, but I'm just not important enough of a player in the larger game and others claim that there's no need for reform despite the chaos of current AI journalism.

For example, going off of the "multiple names for the same concept" issue, there's a fallacy that any general AI also qualifies as human-level AI. This fallacy only exists because of the wonky naming system— the very existence of biological intelligence ought to be proof that you can have general intelligence but not human-level intelligence (and human-level intelligence is already variable).
This leads to counter-intuitive and even self-defeating logic: if I had an AGI that was as intelligent and generalized as a bonobo, then technically I still have ANI. This makes absolutely no sense, but this is how AI is discussed.
And that's not even getting into your original argument: that only sapient, human-level general AI qualifies as "AI", which is patently false. I can understand not wanting to consider anything that can't meaningfully learn and reapply its knowledge "AI", but these ever-increasing standards of what makes something "true AI" is just muddying discussion even further. "True AI" is merely one minute aspect of the entire spectrum— the aforementioned strong, sapient, human-level, general AI. But it's theoretically possible to have general AI that isn't human-level, or sapient, or anything more than a glorified question-answering machine.


If you want someone to blame for all this, you can either blame Hollywood & sci-fi writers for running with "true AI" in fiction or the original creators of the term for not going with "data science".

 

I had no idea AI had been 'muddied' to such an extent.

I wasn't even commenting from a game perspective.

And I'm really glad I missed all that discussion to break something, which is to my mind very clear cut, into so much ambiguity.

But the problem is obviously there. The concept has been perverted by those wanting to use marketing to convince people of something it's not.

As far as I'm concerned, an AI is either human-level of independence and consciousness, or beyond it. Anything below human intelligence/consciousness levels, is not yet AI. There is no spectrum. It is or it isn't. There is no almost. And while humans vary as to both, there is a hard line there.

Trying to justify a near miss as a lower level of AI is just ego wanking. imo. The true separation of AI into categories happens after consciousness is achieved, not before. And until consciousness is achieved, all you have is a complicated computer.

But just to muddy the water even further, it brings me back to this old chestnut:

Kirk to Enterprise.
Spock here.
Any signs of intelligent life?
Negative.
Gee it's good to be home.