The Apple researchers found out more or less three aspects:<\/p>\n\n\n\n
- \n
- For simple tasks, LLMs are surprisingly better than LRMs;<\/li>\n\n\n\n
- LRMs are good for medium-complex tasks;<\/li>\n\n\n\n
- LRMs collapse with very complex<\/em> ones<\/li>\n<\/ol>\n\n\n\n
The conclusion of the Apple researchers was that neither LLMs nor LRMs would “think” as we expect humans to.<\/p>\n\n\n\n
The reaction from AI critics came quickly. They immediately pointed to this paper, saying “I told you so! AI can’t think! Stop portraying AI as intelligent or thinking!” But this reaction is completely exaggerated and wrong, as I will show after the next part. <\/p>\n\n\n\n
But first a response to the paper, which deals with the experimental setup and the conclusions.<\/p>\n\n\n\n
The Response<\/h3>\n\n\n\n
Under the title “The Illusion of the Illusion of Thinking<\/a>“, the researcher and C. Opus – Anthropics Claude Opus<\/a> LRM – pointed out that the Apple researchers had made several mistakes.<\/p>\n\n\n\n
In the complex Tower of Hanoi<\/a> problem, the Apple researchers confused limitations with reasoning errors. In the experimental setup, the LRM had a limited number of available tokens – i.e. memory storage – which aborted the execution. It cannot be concluded from this that the LRM cannot reason. It is as if a vehicle ran out of gas, but the conclusion would be that this vehicle cannot really drive. <\/p>\n\n\n\n
The criticism of the responding researchers also refers to the fact that the Apple researchers generally do not distinguish between practical restrictions and actual justification restrictions.<\/p>\n\n\n\n
And finally, some of the problems posed, such as the river crossing puzzle<\/a>, are given as a knockout criterion for incorrect reasoning, although there is no solution for this problem from a criterion number N\u22656. In the Apple researchers’ paper, however, the LRM was given exactly such examples with N\u22656 to solve. <\/p>\n\n\n\n
Dans leur r\u00e9ponse, les chercheurs soulignent que 1) les mod\u00e8les ne peuvent pas utiliser plus de jetons que leur contexte ne le permet, que 2) l’\u00e9valuation programmatique peut n\u00e9gliger \u00e0 la fois les capacit\u00e9s du mod\u00e8le et les impossibilit\u00e9s de l’\u00e9nigme et que 3) la longueur de la solution pr\u00e9dit mal la difficult\u00e9 du probl\u00e8me.<\/p>\n\n\n\n
In other words, these are all valuable technical insights, but they do not support the claim that thinking is fundamentally limited.<\/p>\n\n\n\n
Birds and Airplanes<\/h3>\n\n\n\n
However, this discussion overlooks an essential aspect. AI models do not have to think like humans, nor do they think like humans. Debating whether AIs can think is therefore misguided. <\/p>\n\n\n\n
Just because airplanes don’t flap their wings like birds doesn’t mean that airplanes can’t fly. Just because a boat doesn’t move like a fish doesn’t mean that a boat can’t swim. <\/p>\n\n\n\n
They were developed for different purposes and within their design domain they are useful. Just because an AI thinks differently than a human, puts forward a different intelligence than a human, does not make it useless or stupid, especially when we look at all the stupid human intelligences we are surrounded by. <\/p>\n\n\n\n
And why LLMs are better at simple tasks than LRMs is also easily explained with a comparison. A bicycle is enough to get to the bakery around the corner. I don’t have to use a jet plane to pick up my bread from there. <\/p>\n\n\n\n
In my AI book When Monkeys Teach Monkeys<\/a>, published in 2020, I warned against equating or confusing the requirements of artificial intelligence with those of human intelligence.<\/p>\n\n\n\n
\n
Just as Copernicus opened up completely new perspectives for us with the change from the geocentric world view, in which the earth was considered the center of things, to the heliocentric world view with the sun at the center, we should move away from the focus on the anthropocentric concept of intelligence and arrive at a general form and definition. Intelligence is a multidimensional spectrum that can occur in different forms adapted to each species and machine. A species or a machine only moves in a small part of the entire multidimensional intelligence space. <\/p>\n\n\n\n
[…] Our challenge is that we seem to have different expectations of machines than we do of animals. We don’t expect our dog to take out the garbage. We won’t ask our cat to take the parcel from the courier and sign for it. All we ask of these animals is that they chase away burglars or mice and pose for cute dog and cat videos. While intelligence tests for humans also test some abstract thinking, social intelligence is almost never included. Chimpanzees or gorillas, for example, would be vastly superior to us in this respect. <\/p>\n<\/blockquote>\n\n\n\n
In this multidimensional intelligence space, we want complementary intelligence, i.e. intelligence that extends us, and not intelligence that simply replaces us.<\/p>\n\n\n\n
\n\n![\"\"](\"https:\/\/i0.wp.com\/futureangst.com\/wp-content\/uploads\/2020\/05\/WAvAl.jpg?w=720&ssl=1\")
<\/a><\/figure>\n<\/div>\n\n\n\n\nWhat is intelligence in the artificial and human sense? Can machines develop consciousness and how would we know? Are machines capable of showing and feeling empathy?<\/p>\n\n\n\n
The book highlights the many opportunities and positive impacts of AI on all aspects of social and economic life.<\/p>\n\n\n\n
Exciting conversations with AI thought leaders and AI practitioners from Silicon Valley provide readers with valuable new insights and mindsets. An indispensable AI guide for the present and the future!<\/p>\n\n\n\n
\u20ac24,99 | 304 pages | 27.2.2020
Amazon<\/a> | Plassen-Buchverlag<\/a><\/p>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"