The Jolly Contrarian

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

Template Discussion

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

Revision as of 20:27, 22 November 2023

113 bytes added ,  22 November 2023
→‎If AI is a cheapest-to-deliver strategy you are doing it wrong
Line 176: Line 176:
Of the range of possible ways of discharging your [[contract|contractual obligation]] to the letter, the one that will cost you the least and irritate your customer the most should you choose it.}}
Of the range of possible ways of discharging your [[contract|contractual obligation]] to the letter, the one that will cost you the least and irritate your customer the most should you choose it.}}


Imagine we each had private [[large language model]]s at our personal disposal — free, therefore, of data privacy concerns — that could pattern-match by reference to our individual reading and listening histories, our engineered prompts, our instructions and the recommendations of like-minded readers. Our LLM would search through the entire human ''oeuvre'' — the billions of books, plays, films, recordings and artworks, known and not, that already exist, but instead of using that information to generate random mashups, it would return existing works from the canon of as yet undiscovered delight? 
Imagine we each had private [[large language model]]s at our personal disposal — free, therefore, of data privacy concerns — that could pattern-match against our individual reading and listening histories, our engineered prompts, our instructions and the recommendations of like-minded readers.  


This is not just the Spotify recommendation algorithm, as occasionally delightful as that is. Any commercial algorithm has its own primary goal: revenue maximisation. A certain amount of “customer delight” may be a necessary by-product, but only as far as it intersects with that primary commercial goal. As long as customers are just delighted ''enough'' to keep listening, the algorithm doesn’t care ''how'' delighted they are.<ref>As with the JC’s school exam grades: anything more than 51% is wasted effort.Try as he might, the JC was never able to persuade his dear old ''Mutti'' about this.</ref>
Our LLM would search through the entire human ''oeuvre'' — the billions of existing books, plays, films, recordings and artworks, known and unknown — but instead of using that information to make its own mashups, it would bring back existing human-authored works from the canon that its patterns told it would unusually suit you?  


Commercial algorithms need only follow a ''[[cheapest to deliver]]'' strategy: they “[[satisfice]]”. Being targeted primarily at revenue optimisation, they will tend to converge upon what is likely to be popular, because that is easier to find. Rather than scanning the entire depth of human content, skim the top and keep the punters happy enough.  
This is not just the Spotify recommendation algorithm, as occasionally delightful as that is. Any commercial algorithm has its own primary goal to maximise revenue. A certain amount of “customer delight” might be a by-product, but only as far as it intersects with that primary commercial goal. As long as customers are ''just delighted'' ''enough'' to keep listening, the algorithm doesn’t care ''how'' delighted they are.<ref>As with the JC’s school exam grades: anything more than 51% is wasted effort.Try as he might, the JC  was never able to persuade his dear old ''Mutti'' about this.</ref>


This, by the way, has been the tale of the collaborative internet: despite [[Chris Anderson]]’s wishful forecast in 2006 that universal interconnectedness would change economics forever<ref>[[The Long Tail: How Endless Choice is Creating Unlimited Demand]] ''(2006)''</ref> — that suddenly it would be costless to service the long tail of global demand, prompting some kind of explosion in cultural diversity — what happened in practice has been the exact opposite. The overriding imperative of [[scale]] has obliterated the subtle appeal of diversity, while sudden, unprecedented global interconnectedness has had the [[system effect]] of ''homogenising demand''.
Commercial algorithms need only follow a ''[[cheapest to deliver]]'' strategy: they “[[satisfice]]”. Being targeted at optimising revenue, they converge upon what is likely to be popular, because that is easier to find. Why scan the ocean deeps of human content when you can skim the top and keep the punters happy enough?


Not only has it remained easier to target the fat head than the thin tail, but ''the tail itself got thinner''.<ref>{{author|Anita Elberse}}’s [[Blockbusters: Why Big Hits and Big Risks are the Future of the Entertainment Business|''Blockbusters'']] is excellent on this point. </ref>     
This, by the way, has been the tale of the collaborative internet: despite [[Chris Anderson]]’s forecast in 2006 that universal interconnectedness would change economics forever<ref>[[The Long Tail: How Endless Choice is Creating Unlimited Demand]] ''(2006)''</ref> — that, suddenly, it would be costless to service the long tail of global demand, prompting some kind of explosion in cultural diversity — in practice, the exact opposite has happened.<ref>It is called “{{Plainlink|https://www.theclassroom.com/cultural-convergence-examples-16778.html|cultural convergence}}”.</ref> The overriding imperatives of [[scale]] have obliterated the subtle appeals of diversity, while sudden, unprecedented global interconnectedness has had the [[system effect]] of ''homogenising demand''.
 
While it has become ever easier to target the “[[fat head]]”, ''the [[long tail]] has grown thinner''.<ref>{{author|Anita Elberse}}’s [[Blockbusters: Why Big Hits and Big Risks are the Future of the Entertainment Business|''Blockbusters'']] is excellent on this point. </ref>     


A [[cheapest-to-deliver]] strategy will have the counter-intuitive effect of ''truncating'' the “[[long tail]]” of consumer choice. As the tail contracts, the [[commercial imperative]] to target common denominators gets stronger. ''This is a highly undesirable feedback loop''. It will homogenise ''us''. ''We'' will become less diverse. We will become more [[Antifragile|fragile]].   
A [[cheapest-to-deliver]] strategy will have the counter-intuitive effect of ''truncating'' the “[[long tail]]” of consumer choice. As the tail contracts, the [[commercial imperative]] to target common denominators gets stronger. ''This is a highly undesirable feedback loop''. It will homogenise ''us''. ''We'' will become less diverse. We will become more [[Antifragile|fragile]].   
Line 207: Line 209:
All mammals can do this to a degree; even retrievers.<ref>Actually come to think of it Lucille, bless her beautiful soul, didn’t seem to do that very often. But still. </ref> Humans happen to be particularly good at it. This is our great superpower: it took three and a half billion years to get from amino acid to the wheel, but 6,000 years to get from the wheel to the Nvidia  RTX 4090 GPU.  
All mammals can do this to a degree; even retrievers.<ref>Actually come to think of it Lucille, bless her beautiful soul, didn’t seem to do that very often. But still. </ref> Humans happen to be particularly good at it. This is our great superpower: it took three and a half billion years to get from amino acid to the wheel, but 6,000 years to get from the wheel to the Nvidia  RTX 4090 GPU.  


Now. [[Large language model]]<nowiki/>s are, like evolution, ''undirected''. They are a brute force method, using colossal amounts of energy and processing power. They work by a stochastic algorithm not dissimilar to evolution by natural selection. They can get better by chomping yet more data, faster, in parallel, with batteries of server farms in air-cooled warehouses full of lightning-fast multi-core graphics processors. But this takes colossal amounts of processing power and energy. This is starting to get expensive and hard. We are bumping up against computational limits as Moore’s law conks out, and environmental consequences as the planet does.  
Now. [[Large language model]]<nowiki/>s are, like evolution, ''undirected''. They are a brute force method, using colossal amounts of energy and processing power. They work by a stochastic algorithm not dissimilar to evolution by natural selection. They can get better by chomping yet more data, faster, in parallel, with batteries of server farms in air-cooled warehouses full of lightning-fast multi-core graphics processors. But this takes colossal amounts of processing power and energy. This is starting to get expensive and hard. We are bumping up against computational limits as Moore’s law conks out, and environmental consequences as the planet does.


For the longest time, computing power has been the cheap, efficient model. That is ceasing to be true. More silicon is not a zero-cost option. We will start to see the opportunity cost to devoting all these resources to something that, at the moment, creates diverting sophomore mashups we don’t actually need.<ref>We would do well to remember Arthur C. Clarke’s law here. The parallel processing power an LLM requires is already massive. It may be that the cost of expanding it in the way envisioned would be unfeasibly huge — in which case the original “business case” for [[technological redundancy]] falls away. See also the [[simulation hypothesis]]: it may be that the most efficient way of simulating the universe with sufficient granularity to support the simulation hypothesis is ''to actually build and run a universe'' in which case, the hypothesis fails.</ref>
For the longest time, computing power has been the cheap, efficient model. That is ceasing to be true. More silicon is not a zero-cost option. We will start to see the opportunity cost to devoting all these resources to something that, at the moment, creates diverting sophomore mashups we don’t actually need.<ref>We would do well to remember Arthur C. Clarke’s law here. The parallel processing power an LLM requires is already massive. It may be that the cost of expanding it in the way envisioned would be unfeasibly huge — in which case the original “business case” for [[technological redundancy]] falls away. See also the [[simulation hypothesis]]: it may be that the most efficient way of simulating the universe with sufficient granularity to support the simulation hypothesis is ''to actually build and run a universe'' in which case, the hypothesis fails.</ref>
Retrieved from "https://jollycontrarian.com/index.php?title=Special:MobileDiff/80581"