Template:M intro technology rumours of our demise: Difference between revisions

Template:M intro technology rumours of our demise (view source)

Revision as of 10:46, 24 November 2023

54 bytes removed , 24 November 2023

no edit summary

Amwelladmin

Bureaucrats, Interface administrators, Administrators

82,891

edits

@@ Line 247: / Line 247: @@
 Last point on literary theory is that the “[[Bayesian priors]]” argument which fails for Shakespeare also fails for a [[large language model]].
-Just as most of the intellectual energy needed to render a text into the three-dimensional [[metaphor]]ical universe we know as ''King Lear'' comes from the surrounding cultural milieu, so it does with the output of an LLM. The source, after all, is entirely drawn from the human canon. A model trained only on randomly assembled ASCII characters would return only randomly assembled ASCII characters.
+Just as most of the intellectual energy needed to render a text into the three-dimensional [[Metaphor|metaphorical]] universe we know as ''King Lear'' comes from the surrounding cultural milieu, so it does with the output of an LLM. The source, after all, is entirely drawn from the human canon. A model trained only on randomly assembled ASCII characters would return only randomly assembled ASCII characters.
 But what if the material is not random? What if the model augments its training data with its own output? Might that create an apocalyptic feedback loop, whereby LLMs bootstrap themselves into some kind of hyperintelligent super-language, beyond mortal cognitive capacity, whence the machines might dominate human discourse?
 Are we inadvertently seeding ''Skynet''?
-Just look what happened with [[Alpha Go]]. It didn’t require ''any'' human training data: it learned by playing millions of games against itself. Programmers just fed it the rules, switched it on and, with indecent brevity, it worked everything out and walloped the game’s ruling grandmaster.
+Just look at what happened with [[Alpha Go]]. It didn’t require ''any'' human training data: it learned by playing millions of games against itself. Programmers just fed it the rules, switched it on and, with indecent brevity, it worked everything out and walloped the game’s reigning grandmaster.
-Could LLMs do that? This fear has been with us for a while.
+Could LLMs do that? This fear has been with us for a while.{{Quote|{{rice pudding and income tax}}}}
-{{Quote|{{rice pudding and income tax}}}}
+But brute-forcing outcomes in fully bounded, [[Zero-sum game|zero-sum]] environments with simple, fixed rules — in the jargon of [[Complexity|complexity theory]], a “tame” environment — is what machines are designed to do. We should not be surprised that they are good at this, nor that humans are bad at it. ''This is exactly where we would expect a Turing machine to excel''.
-But brute-forcing outcomes in a fully bounded, [[Zero-sum game|zero-sum]] environments with simple, fixed rules — in the jargon of [[Complexity|complexity theory]], a “tame” environment — is exactly what machines are designed to do. We should not be surprised that they are good at this, nor that humans are bad at it.
+By contrast, LLMs must operate in complex, “[[wicked]]” environments. Here conditions are unbounded, ambiguous, inchoate and impermanent. ''This is where humans excel''. Here, the situation continually changes. The components interact with each other in non-linear ways. The landscape dances. Imagination here is an advantage: brute force mathematical computation won’t do.{{Quote|Think how hard physics would be if particles could think.
-To see this as a fair comparison is to misdirect ''ourselves'': willingly, to suspend disbelief. ''This is exactly where we would expect a Turing machine to excel''.
-By contrast, LLMs must operate in complex, “[[wicked]]” environments. Here conditions are unbounded, ambiguous, inchoate and impermanent. ''This is where humans excel''. The situation continually changes. The components interact with each other to make the landscape dance. Here, narratising is an advantage: brute force mathematical computation won’t do.
-{{Quote|Think how hard physics would be if particles could think.
 :— Murray Gell-Mann}}
-And nor does it: an LLM works by compositing a synthetic output from a massive database of pre-existing text. It must pattern-match against well-formed human language. Degrading its training data will progressively degrade its output. Such “model collapse” is an observed effect.<ref>https://www.techtarget.com/whatis/feature/Model-collapse-explained-How-synthetic-training-data-breaks-AI</ref> LLMs will only work for humans if they’re fed human generated content.
+An LLM works by compositing a synthetic output from a massive database of pre-existing text. It must pattern-match against well-formed human language. Degrading its training data with its own output will progressively degrade its output. Such “model collapse” is an observed effect.<ref>https://www.techtarget.com/whatis/feature/Model-collapse-explained-How-synthetic-training-data-breaks-AI</ref> LLMs will only work for humans if they’re fed human generated content. [[Alpha Go]] is different.
 {{Quote|{{AlphaGo v LLM}}}}