Kurzgesagt – In a Nutshell 

Sources – Superintelligence

We thank following experts for their feedback: 

• Jonathan Richard Schwarz

Research Fellow in Foundational AI, Harvard University

• Daniel Kokotajlo

AI expert, OpenAI (sponsored by Open Philanthropy) employee at the time of our collaboration

• Robert Miles

Computer Science PhD Student, University of Nottingham

Open Philanthropy is a sponsor of kurzgesagt. The foundation is supporting academic work across the field of Artificial Intelligence, and some of the sources used to create this script (from OpenAI, Future of Humanity Institute, Machine Intelligence Research Institute, Future of Life Institute and Epoch AI) also receive financial support from Open Philanthropy. Open Philanthropy had no influence on the content and messages of this video.

—Intelligence is the ability to learn, reason, acquire knowledge and skills and use them to solve problems. Intelligence is power, and we’re the species that exploited it the most. So much so that humanity broke the game of nature and took control. 

There are many definitions of intelligence, but most of them have the capacity to “learn” and “solve problems” at their core. 

#Wenke, Dorit and Frensch, Peter A. and Funke, Joachim (2005): “Complex problem solving and intelligence: Empirical relation and causal direction” in “Cognition and intelligence: Identifying the mechanisms of the mind”, edited by Sternberg, R.J. ; Pretz, J.E..Cambridge University Press.

https://web-archive.southampton.ac.uk/cogprints.org/6626/index.html 

Quote: “At least two theoretical positions strongly suggest that intelligence and problem solving are related. First, the ability to solve problems features prominent in almost every definition of human “intelligence;” thus, problem-solving capacity is viewed as one component of intelligence. Second, intelligence is often assumed to be a predictor of problem-solving ability.”

#Encyclopaedia Britannica: “Human Intelligence” (retrieved 2024)

https://www.britannica.com/science/human-intelligence-psychology 

Quote: “Human intelligence, mental quality that consists of the abilities to learn from experience, adapt to new situations, understand and handle abstract concepts, and use knowledge to manipulate one’s environment. 

Much of the excitement among investigators in the field of intelligence derives from their attempts to determine exactly what intelligence is.”

—But the journey there wasn’t straightforward. For most animals intelligence costs too much energy to be worth it. 

#Isler, Karin; van Schaik, Carel P. (2006): “Metabolic costs of brain size evolution”. Biol Lett. vol. 2, 4,  557–560
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1834002/ 

Quote: “However, brain tissue is energetically expensive, requiring nearly an order of magnitude more energy per unit weight than several other somatic tissues during rest (Mink et al. 1981). The high proportion of energy necessarily allocated to brain tissue may therefore constrain the response of natural selection to the beneficial impact of increased brain size on an animal's survival and/or reproductive success”

#Tomasi, Dardo; Wang, Gene-Jack; Volkowa, Nora D.: (2013) ”Energetic cost of brain functional connectivity” Proc Natl Acad Sci USA. vol. 110, 33, 13642–13647

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3746878/

Quote: “The high energetic cost of human brain function, which is 10 times higher than what would be expected from its weight alone (1, 2), can only be maintained through a combination of strategies for efficient energy use (3–5).”

Though an increase in brain size or mass can’t be directly tied to an increase in intelligence, the combination of brain size and neural connectivity seems to correlate with intelligence in animals.

#Roth, Gerhard; Dicke, Ursula (2005): “Evolution of the brain and intelligence,

Trends in Cognitive Sciences”,  vol.9, 5, 250-257
https://pubmed.ncbi.nlm.nih.gov/15866152/ 

Quote: “Brain properties assumed to be relevant for intelligence are the (absolute or relative) size of the brain, cortex, prefrontal cortex and degree of encephalization. However, factors that correlate better with intelligence are the number of cortical neurons and conduction velocity, as the basis for information-processing capacity.” 

 As neural connectivity can’t be easily derived from fossils, most studies on the evolution of intelligence use brain size as a metric. 

—Still, if we track intelligence in the tree of species over time, we can see lots of diverse forms of intelligence emerge.

#Bräuer, Juliane et al. (2020): “Old and New Approaches to Animal Cognition: There Is Not "One Cognition".” Journal of Intelligence, vol. 8,3 28. 

https://pubmed.ncbi.nlm.nih.gov/32630788/
https://pure.mpg.de/rest/items/item_3241522/component/file_3241523/content 

Quote: “In this paper, we emphasize that specific physical and social environments create selection pressures that lead to the evolution of certain cognitive adaptations. Skills such as following the pointing gesture, tool-use, perspective-taking, or the ability to cooperate evolve independently from each other as a concrete result of specific selection pressures, and thus have appeared in distantly related species. Thus, there is not “one cognition”

[...]

The domestic dog was domesticated about 30,000 years ago (Botigué et al. 2017; Thalmann et al. 2013) and shows outstanding skills in the social-cognitive domain (see Kaminski and Marshall-Pescini 2014; Miklosi 2007; Huber 2016 for reviews).[...]

In contrast, dogs do not show exceptional physical cognitive skills but perform similarly to other nonprimate mammals and birds (Bräuer et al. 2006; Erdohegyi et al. 2007; Osthaus et al. 2005; Rooijakkers et al. 2009; Miletto Petrazzini and Wynne 2016).

[...]

New Caledonian crows are renowned for their technological abilities (Weir et al. 2002; Taylor et al. 2007; Hunt and Gray 2004; Rutz et al. 2010; Hunt and Uomini 2016; Uomini and Hunt 2017). They not only use stick, stem, and grass tools in their natural environments (Figure 1) but also manufacture pandanus tools following templates to produce specific tool shapes that vary between populations and between individuals (Hunt and Gray 2004; Kenward et al. 2006; Taylor et al. 2012b).[...] Although New Caledonian crows may have a better understanding than related bird species of how to use metatools (i.e., the ability to use one tool on another tool; Taylor et al. 2007; Kenward et al. 2006; Gruber et al. 2019), they do not outperform them in other physical cognition tasks. ”

#Godfrey-Smith, Peter (2017): “The Mind of an Octopus”, Scientific American Mind, vol. 28, 1, 62

https://www.scientificamerican.com/article/the-mind-of-an-octopus/ 

Quote: “Octopuses and their relatives (cuttlefish and squid) represent an island of mental complexity in the sea of invertebrate animals.[...] Octopuses have done fairly well on tests of their intelligence in the laboratory, without showing themselves to be Einsteins. They can learn to navigate simple mazes. They can use visual cues to discriminate between two familiar environments and then take the best route toward some reward. They can learn to unscrew jars to obtain the food inside—even from the inside out. But octopuses are slow learners in all these contexts. Against this background of mixed experimental results, however, there are countless anecdotes suggesting that a lot more is going on.”

#Wystrach, Antoine (2013): “We've Been Looking at Ant Intelligence the Wrong Way” Scientific American & The Conversation

https://www.scientificamerican.com/article/weve-been-looking-at-ant-intelligence-the-wrong-way/
https://theconversation.com/weve-been-looking-at-ant-intelligence-the-wrong-way-17619 

Quote: “Ants use a variety of cues to navigate, such as sun position, polarized light patterns, visual panoramas, gradient of odors, wind direction, slope, ground texture, step-counting … and more. Indeed, the list of cues ants can utilise for navigation is probably greater than for humans.

Counter-intuitively, years of bottom-up research has revealed that ants do not integrate all this information into a unified representation of the world, a so-called cognitive map. Instead they possess different and distinct modules dedicated to different navigational tasks. These combine to allow navigation.”

 —The earliest brains were in flatworms 500 million years ago. Just a tiny cluster of neurons to handle basic body functions. 

#Joseph, Rhawn G. (2018): “Evolution of the Brain, the Neuron, Nerve Net, Limbic System, Brainstem, Midbrain, Diencephalon, Striatum, Telencephalon: Genetics, Calcium, Oxygen, and Cyanobacteria” (retrieved 2024)

http://brainmind.com/BrainEvolution.html
Quote: “Again, although a rudimentary "nerve net" may have "evolved" well over a billion years ago, it is evident that by 550 million years ago a complex network of neurons, that is a "nervous system" had been established; a development that may have coincided with the appearance of the planaria, hydra, and flat worms; i.e. the coelenterates.

Coelcenterates are the most primitive members of the animal kingdom to possess not just neurons, but a "nervous system," composed of distinct olfactory-chemical, photosensory, and sensory-motor neurons, including those which are unipolar, bipolar, and multipolar (Ariens Kappers, 1929; Papez, 1967; Lentz 1958). They also contain a third type of cell called a ganglion cell (or protoneuron). These nerve cells also possess true synapses which resemble those of the mammalian nervous system. For example, hydra (e.g., flatworms) are symmetrical and have an enlarged anterior region which correspond to its head, and are capable of twsting, bending, swaying, and waving their tentacles which respond to a variety of stimuli.”

#Sarnat, Harvey B.; Netsky, Martin G. (1985): “The brain of the planarian as the ancestor of the human brain.” vol.12, 4, 296-302.

https://www.cambridge.org/core/services/aop-cambridge-core/content/view/72EF0730D9F191F96F7AA48841E05D6A/S031716710003537Xa.pdf/the-brain-of-the-planarian-as-the-ancestor-of-the-human-brain.pdf 

Quote: “The planarian is the simplest living animal having a body plan of bilateral symmetry and cephalization.The brain of these free-living flatworms is a biiobed structure with a cortex of nerve cells and a core of nerve fibres including some that decussate to form commissures. Special sensory input from chemoreceptors, photoreceptor cells

of primitive eyes, and tactile receptors are integrated to provide motor responses of the entire body, and local reflexes. Many morphological, electrophysiological, and pharmacological features of planarian neurons, as well as synaptic organization, are reminiscent of the vertebrate brain. ”

#Riebli, Nadia; Reichert, Heinrich (2015): “Perspective—The First Brain” in “Structure and Evolution of Invertebrate Nervous Systems” edited by Andreas Schmidt-Rhaesa, Steffen Harzsch, Günter Purschke . Oxford Academic. 

https://academic.oup.com/book/43960/chapter-abstract/369205644?redirectedFrom=fulltext 

Quote: “In contrast, even very early diverging bilaterians, such as C. elegans and planarians, have global neural integration centers that can be thought of as a “brain” (Garrity et al., 2010). Although there is still controversy regarding whether the first bilaterians had a neural net or an actual brain (Hejnol and Martindale, 2008; Arendt et al., 2015), both interpretations agree that the first brain emerged within the bilaterian lineage.”

—It took hundreds of millions of years for species to diversify and become more complex. Life conquered new environments, gained new senses and had to contend with fierce competition over resources. 

#Garwood, Russell J.; Edgecombe, Gregory D. (2011):  “Early Terrestrial Animals, Evolution, and Uncertainty” Evo Edu Outreach vol. 4, 489–501 

https://evolution-outreach.biomedcentral.com/articles/10.1007/s12052-011-0357-y 

Quote: “Early terrestrial ecosystems record a fascinating transition in the history of life. Animals and plants had previously lived only in the oceans, but, starting approximately 470 million years ago, began to colonize the previously barren continents.”

#Encyclopaedia Britannica: “Evolution of eyes” (retrieved 2024)

https://www.britannica.com/science/photoreception/Evolution-of-eyes

Quote: “The soft-bodied animals that inhabited the world’s seas before the Cambrian explosion (about 541 million years ago) undoubtedly had eyes, probably similar to the pigment-pit eyes of flatworms today.” 

#Niimura, Yoshihito (2012): “Olfactory Receptor Multigene Family in Vertebrates: From the Viewpoint of Evolutionary Genomics”,  Current genomics vol. 13, 2, 103-14

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3308321/ 

Quote: "Extensive bioinformatic analyses using the whole genomes of various organisms revealed a great variation in number of OR genes among species, reflecting the diversity of their living environments."

#Griffin, John N. ; Silliman, Brian R. (2011): “Resource Partitioning and Why it Matters.” Nature Education Knowledge vol. 3, 10, 49

https://www.nature.com/scitable/knowledge/library/resource-partitioning-and-why-it-matters-17362658/ 

Quote:  “There are only a limited number of ways of "making a living" within ecological communities. For example, on a coral reef, there are hard-skeleton corals that gain food from capturing planktonic animals in their tentacles and, in exchange for providing a suitable habitat and nutrients, gain extra sources of energy from sugar-synthesizing symbiotic algae. Within groups of species that make a living in a similar way, species compete for the same resources. These resources, which include nutrients and habitat, are the raw materials needed by organisms to grow, live, and reproduce. However, resources are not unlimited, and individuals from different species commonly compete for resources (interspecific competition).

[...]

Competition can limit the growth, and ultimately the reproductive success, of individuals. It can consequently serve as a selection pressure driving differential reproductive success and the evolution of traits that enable organisms to use resources differently compared to their competitors. This process has been clearly demonstrated in the evolutionary events that have followed the colonization of volcanic islands. For example, a single species of seed-eating finch originally colonized the Galapagos Islands and was faced with a diverse range of seed types and sizes. However, the beak of the founding species only allowed it to eat a small subset of the available seed types and sizes. The advantages gained by individuals that were able to exploit slightly different seed types drove evolution of many new species, each with different shaped beaks enabling them to specialize in a particular size of seed (Grant 1986).”

—But brains are expensive and in nature all that matters is survival, so for almost all animals a narrow intelligence, fit for a narrow range of tasks was enough. In some environments, animals like birds, octopuses and mammals evolved more complex neural structures. For them it paid off to have more energy-consuming skills like advanced navigation and communication. 

#Barron, Andrew B. ; Halina,  Marta; Klein, Colin (2023): “Transitions in cognitive evolution”, Proceedings of the Royal Society B,vol. 290, 2002, 20230671

https://royalsocietypublishing.org/doi/10.1098/rspb.2023.0671 

Quote: “The evolutionary history of animal cognition appears to involve a few major transitions: major changes that opened up new phylogenetic possibilities for cognition. Here, we review and contrast current transitional accounts of cognitive evolution. We discuss how an important feature of an evolutionary transition should be that it changes what is evolvable, so that the possible phenotypic spaces before and after a transition are different. We develop an account of cognitive evolution that focuses on how selection might act on the computational architecture of nervous systems. Selection for operational efficiency or robustness can drive changes in computational architecture that then make new types of cognition evolvable. We propose five major transitions in the evolution of animal nervous systems. Each of these gave rise to a different type of computational architecture that changed the evolvability of a lineage and allowed the evolution of new cognitive capacities. Transitional accounts have value in that they allow a big-picture perspective of macroevolution by focusing on changes that have had major consequences. For cognitive evolution, however, we argue it is most useful to focus on evolutionary changes to the nervous system that changed what is evolvable, rather than to focus on specific cognitive capacities.

[...]

In the vertebrates, and cephalopod gastropods, we recognize a fourth transition to laminated computational architecture. In laminated systems, the control flow operates in parallel but interacting recurrent subsystems. Lamination is a structural rather than a physical concept: the avian pallium likely has an equivalent computational architecture to the vertebrate neocortex, but the avian pallium is organized as nuclei while the mammalian neocortex displays layers [68,69]. The important feature is the possibility of multiple parallel but interacting subsystems.”

—Until seven million years ago, the Hominins emerged. We don’t know why, but their brains grew faster than their relatives’. Something was different about their intelligence – very slowly, it turned from narrow, to general. From a screwdriver to a multi tool. Able to think about any diverse problems.

Humans have general intelligence, while animals do not

#Poirier, Marc-Antoine et al.  (2020): “How general is cognitive ability in non-human animals? A meta-analytical and multi-level reanalysis approach”, Proceedings of the Royal Society B: Biological Sciences, vol. 287, 1940, 20201853

https://royalsocietypublishing.org/doi/10.1098/rspb.2020.1853
Quote: “General intelligence has been a topic of high interest for over a century. Traditionally, research on general intelligence was based on principal component analyses and other dimensionality reduction approaches. The advent of high-speed computing has provided alternative statistical tools that have been used to test predictions of human general intelligence. In comparison, research on general intelligence in non-human animals is in its infancy and still relies mostly on factor-analytical procedures. Here, we argue that dimensionality reduction, when incorrectly applied, can lead to spurious results and

limit our understanding of ecological and evolutionary causes of variation in animal cognition. Using a meta-analytical approach, we show, based on 555

bivariate correlations, that the average correlation among cognitive abilities is low (r = 0.185; 95% CI: 0.087–0.287), suggesting relatively weak support

for general intelligence in animals.“

Though we can’t know when exactly general intelligence appeared in human evolutionary history and it was probably a gradual process, it is generally accepted to have happened at some point along the existence of H. erectus or  H. heidelbergensis.

#Pontzer, Herman. (2012): “Overview of Hominin Evolution.” Nature Education Knowledge 3(10):8

https://www.nature.com/scitable/knowledge/library/overview-of-hominin-evolution-89010983/ 

Quote: “The earliest fossils of our own genus, Homo, are found in East Africa and dated to 2.3 mya (Kimbel et al. 1997). These early specimens are similar in brain and body size to Australopithecus, but show differences in their molar teeth, suggesting a change in diet. Indeed, by at least 1.8 mya, early members of our genus were using primitive stone tools to butcher animal carcasses, adding energy-rich meat and bone marrow to their plant-based diet.

The oldest member of the genus Homo, H. habilis (2.3–1.4 mya) is found in East Africa and is associated with butchered animal bones and simple stone tools (Blumenschine et al. 2003). Its more formidable and widespread descendant, H. erectus, is found throughout Africa and Eurasia and persisted from 1.9 mya to 100 kya, and perhaps even later (Anton, 2003). Like modern humans, H. erectus lacked the forelimb adaptations for climbing seen in Australopithecus (Figure 2). Its global expansion suggests H. erectus was ecologically flexible, with the cognitive capacity to adapt and thrive in vastly different environments. Not surprisingly, it is with H. erectus that we begin to see a major increase in brain size, up to 1,250cc for later Asian specimens (Anton, 2003). Molar size is reduced in H. erectus relative to Australopithecus, reflecting its softer, richer diet.

Around 700 kya, and perhaps earlier, H. erectus in Africa gave rise to H. heidelbergensis, a species very much like us in terms of body proportions, dental adaptations, and cognitive ability (Rightmire, 2009). H. heidelbergensis, often referred to as an "archaic" Homo sapiens, was an active big-game hunter, produced sophisticated Levallois style tools, and by at least 400 kya had learned to control fire (Roebroeks and Villa, 2011). Neanderthals (H. neanderthalensis), cold-adapted hominins with stout physiques, complex behaviors, and brains similar in size to ours, are thought to have evolved from H. heidelbergensis populations in Europe by at least 250 kya (Rightmire, 2008; Hublin, 2009).

Fossil and DNA evidence suggest our own species, H. sapiens, evolved in Africa 200 kya (Relethford, 2008; Rightmire, 2009), probably from H. heidelbergensis. The increased behavioral sophistication of H. sapiens, as indicated by our large brains (1,400cc) and archeological evidence of a broader tool set and clever hunting techniques, allowed our species to flourish and grow on the African continent.”

—Two million years ago, Homo Erectus saw the world differently from anyone before – as something to be understood and transformed. They controlled fire, invented tools and created the first culture.

#Pontzer, Herman. (2012): “Overview of Hominin Evolution.” Nature Education Knowledge 3(10):8

https://www.nature.com/scitable/knowledge/library/overview-of-hominin-evolution-89010983/ 

Quote: “The oldest member of the genus Homo, H. habilis (2.3–1.4 mya) is found in East Africa and is associated with butchered animal bones and simple stone tools (Blumenschine et al. 2003). Its more formidable and widespread descendant, H. erectus, is found throughout Africa and Eurasia and persisted from 1.9 mya to 100 kya, and perhaps even later (Anton, 2003). Like modern humans, H. erectus lacked the forelimb adaptations for climbing seen in Australopithecus (Figure 2). Its global expansion suggests H. erectus was ecologically flexible, with the cognitive capacity to adapt and thrive in vastly different environments. Not surprisingly, it is with H. erectus that we begin to see a major increase in brain size, up to 1,250cc for later Asian specimens (Anton, 2003). Molar size is reduced in H. erectus relative to Australopithecus, reflecting its softer, richer diet.

#Hendry, Lisa (2018): “Homo erectus, our ancient ancestor”. Natural History Museum.

https://www.nhm.ac.uk/discover/homo-erectus-our-ancient-ancestor.html 

Quote: “H. erectus was the first human species to make handaxes (Acheulean tools). These were sophisticated stone tools crafted on two sides. They were probably used to butcher meat, among other purposes.”

#Smithsonian Institution (2024): “Human evolution evidence: Homo erectus”

https://humanorigins.si.edu/evidence/human-fossils/species/homo-erectus 

Quote: “There is fossil evidence that this species cared for old and weak individuals. [...]

The earliest evidence of hearths (campfires) occur during the time range of Homo erectus. While we have evidence that hearths were used for cooking (and probably sharing) food, they are likely to have been places for social interaction, and also used for warmth and to keep away large predators.” 

#MacDonald, Katharine (2017): “The use of fire and human distribution”. Temperature (Austin, Tex.) vol. 4,2 153-165. 

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5489006/

Quote:“Around 350 kya the evidence for fire becomes more frequent and convincing. This may be the result of a change in the frequency of use of fire, perhaps based on an ability to produce fire at will. Searching for older microwear traces of fire production on stone tools may provide relevant evidence. It is tempting to argue that fire played a role in the expansion into cooler conditions which is indicated both by palaeoenvironmental and fossil evidence c. 500 kya. From this time on hominins faced the increased demands of survival in cooler conditions (including the energetic costs associated with producing additional body heat) while at the same time they apparently had reliable sources of energy to support large brains and needy young. They may have been successful hunters with regular access to cold-weather food combined with access to animal hides for insulation. It seems that hominins in this period faced new challenges and developed a range of strategies, of which fire would have been complementary and perhaps necessary.”

#Keim, Brandon (2014): “World's Oldest Art Identified in Half-Million-Year-Old Zigzag”, National Geographic. 

https://www.nationalgeographic.com/adventure/article/141203-mussel-shell-oldest-art 

Quote: “Until now, the earliest evidence of geometric art was dated from 70,000 to 100,000 years ago. Scratched into rocks found in South African caves, those engravings signified behavioral modernity: Homo sapiens' unique cognitive journey into a sophisticated world of abstraction and symbol.

But new analysis of an engraving excavated from a riverbank in Indonesia suggests that it's at least 430,000 years old—and that it wasn't made by humans, scientists announced Wednesday. At least it wasn't made by humans as most people think of them, meaning Homo sapiens.

Rather, the earliest artist appears to have been one of our ancestors, Homo erectus. Hairy and beetle-browed, H. erectus was never before thought to have such talents. ” 

—We probably emerged from them around 250,000 years ago with an even larger and more complex brain.

#Smithsonian Institution (2024): “Human evolution evidence: Homo sapiens” (retrieved 2024)

https://humanorigins.si.edu/evidence/human-fossils/species/homo-sapiens 

Quote: “The species that you and all other living human beings on this planet belong to is Homo sapiens. During a time of dramatic climate change 300,000 years ago, Homo sapiens evolved in Africa. Like other early humans that were living at this time, they gathered and hunted food, and evolved behaviors that helped them respond to the challenges of survival in unstable environments.”

#Pagel, Mark (2017): “Q&A: What is human language, when did it evolve and why should we care?”,  BMC Biology, vol.15, 64 

https://bmcbiol.biomedcentral.com/articles/10.1186/s12915-017-0405-3 

Quote: “Because all human groups have language, language itself, or at least the capacity for it, is probably at least 150,000 to 200,000 years old. This conclusion is backed up by evidence of abstract and symbolic behaviour in these early modern humans, taking the form of engravings on red-ochre”

—Knowledge builds on knowledge. Progress was slow at first and then sped up exponentially. Agriculture, writing, medicine, astronomy or philosophy exploded into the world. 200 years ago science took off and made us even better at learning about the world and speeding up progress. 35 years ago the internet age began. 

#Our world in Data: “Technology over the long run: zoom out to see how dramatically the world can change within a lifetime” (retrieved 2024)
https://ourworldindata.org/technology-long-run

#Bornmann, Lutz; Haunschild, Robin; Mutz, Rüdiger (2021): “Growth rates of modern science: a latent piecewise growth curve approach to model publication numbers from established and new literature databases” Nature Humanities and Social Sciences Communications, vol. 8, 224
https://www.nature.com/articles/s41599-021-00903-w 

Quote: “Growth of science is a prevalent issue in science of science studies. In recent years, two new bibliographic databases have been introduced, which can be used to study growth processes in science from centuries back: Dimensions from Digital Science and Microsoft Academic. In this study, we used publication data from these new databases and added publication data from two established databases (Web of Science from Clarivate Analytics and Scopus from Elsevier) to investigate scientific growth processes from the beginning of the modern science system until today. We estimated regression models that included simultaneously the publication counts from the four databases.

[...]

Although empirical and theoretical studies in previous decades have confirmed exponential growth, a precise estimation of the growth rate based on reliable and sound publication data has not been done yet.”

#CERN: “A short history of the Web” (retrieved 2024)

https://www.home.cern/science/computing/birth-web/short-history-web 

Quote: “Tim Berners-Lee, a British scientist, invented the World Wide Web (WWW) in 1989, while working at CERN. The Web was originally conceived and developed to meet the demand for automated information-sharing between scientists in universities and institutes around the world

CERN is not an isolated laboratory, but rather the focal point for an extensive community that includes more than 17 000 scientists from over 100 countries. Although they typically spend some time on the CERN site, the scientists usually work at universities and national laboratories in their home countries. Reliable communication tools are therefore essential.

The basic idea of the WWW was to merge the evolving technologies of computers, data networks and hypertext into a powerful and easy to use global information system.” 

—Artificial intelligence, or AI, is software that performs mental tasks with a computer. Code that uses silicon, instead of neurons, to solve problems. 

#IBM: “What is artificial intelligence (AI)?” (retrieved 2024)

https://www.ibm.com/topics/artificial-intelligence

Quote: “Artificial intelligence, or AI, is technology that enables computers and machines to simulate human intelligence and problem-solving capabilities.”

#Encyclopaedia Britannica: “Artificial Intelligence” (retrieved 2024)

https://www.britannica.com/technology/artificial-intelligence 

Quote: “Artificial intelligence (AI), the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings. The term is frequently applied to the project of developing systems endowed with the intellectual processes characteristic of humans, such as the ability to reason, discover meaning, generalize, or learn from past experience.”

—In the beginning, AI was very simple. Lines of code on paper, mere proofs of concept to demonstrate how machines could perform mental tasks. Only in the 1960s did we start seeing the first examples of what we would recognize as AI. A chatbot in 1964, a program to sort through molecules in 1965. Slow, specialized systems requiring experts to read their results. Their intelligence was extremely narrow, built for a single task inside a controlled environment. The equivalent of flatworms half a billion years ago, doing the minimum amount of mental work.

#Anyoha, Rockwell (2017): “The History of Artificial Intelligence”, Harvard's SITN Blog. 

https://sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/

Quote: “The Logic Theorist was a program designed to mimic the problem solving skills of a human and was funded by Research and Development (RAND) Corporation. It’s considered by many to be the first artificial intelligence program and was presented at the Dartmouth Summer Research Project on Artificial Intelligence (DSRPAI) hosted by John McCarthy and Marvin Minsky in 1956. In this historic conference, McCarthy, imagining a great collaborative effort, brought together top researchers from various fields for an open ended discussion on artificial intelligence, the term which he coined at the very event. Sadly, the conference fell short of McCarthy’s expectations; people came and went as they pleased, and there was failure to agree on standard methods for the field. Despite this, everyone whole-heartedly aligned with the sentiment that AI was achievable.” 

#Encyclopaedia Britannica: “Alan Turing and the beginning of AI” (retrieved 2024)
https://www.britannica.com/technology/artificial-intelligence/Alan-Turing-and-the-beginning-of-AI 

Quote:“The earliest successful AI program was written in 1951 by Christopher Strachey, later director of the Programming Research Group at the University of Oxford. Strachey’s checkers (draughts) program ran on the Ferranti Mark I computer at the University of Manchester, England”

#Weizenbaum, Joseph (1966): “ELIZA—a computer program for the study of natural language communication between man and machine” Communications of the Association for Computer Machinery Vol. 9, 1,36–45

https://dl.acm.org/doi/10.1145/365153.365168 

https://cse.buffalo.edu/~rapaport/572/S02/weizenbaum.eliza.1966.pdf 

Quote: “ELIZA is a program operating within the MAC time-sharing system at MIT which makes certain kinds of natural language conversation between man and computer possible. Input sentences are analyzed on the basis of decomposition rules which are triggered by key words appearing in the input text. Responses are generated by reassembly rules associated with selected decomposition rules.”

#Encyclopaedia Britannica: “DENDRAL” (retrieved 2024)

https://www.britannica.com/technology/DENDRAL

Quote: “DENDRAL, an early expert system, developed beginning in 1965 by the artificial intelligence (AI) researcher Edward Feigenbaum and the geneticist Joshua Lederberg, both of Stanford University in California. Heuristic DENDRAL (later shortened to DENDRAL) was a chemical-analysis expert system. The substance to be analyzed might, for example, be a complicated compound of carbon, hydrogen, and nitrogen. Starting from spectrographic data obtained from the substance, DENDRAL would hypothesize the substance’s molecular structure. DENDRAL’s performance rivaled that of chemists expert at this task, and the program was used in industry and in academia.”

#Joshi, Kailash: “Expert Systems and Applied Artificial Intelligence: How the AI Field Evolved” (retrieved 2024)

 https://www.umsl.edu/~joshik/msis480/chapt11.htm 

Quote:

“1960 AI established as research field.

Result: Knowledge-based expert systems”

—Progress in AI research paused several times when researchers lost hope in the technology. But just like changing environments create new niches for life, the world around AI changed.

#Umbrello, Steven (2021): “AI Winter” in “Encyclopedia of Artificial Intelligence: The Past, Present, and Future of AI” edited by Michael Klein and Philip Frana. ABC - Clio. 

https://philpapers.org/rec/UMBAW  

Quote: “Coined in 1984 at the American Association of Artificial intelligence (now the Association for the Advancement of Artificial Intelligence or AAAI), the various boom and bust periods of AI research and funding lead AI researchers Marvin Minsky and Roger Schank to refer to the then-impending bust period as an AI Winter. Canadian AI researcher Daniel Crevier describes the phenomenon as a domino effect that begins with cynicism in the AI research community that then trickles to mass media and finally to funding bodies. The result is a freeze in serious AI research and development. This initial pessimism is mainly attributed to the overly ambitious promises that AI can yield with the actual results being far humbler than expectations.”

#Gursoy, Furkan; Kakadiaris, Ioannis A. (2023): “Artificial intelligence research strategy of the United States: critical assessment and policy recommendations” Frontiers in Big Data, vol. 6

https://www.frontiersin.org/articles/10.3389/fdata.2023.1206139/full 

Quote: “As funders became unhappy with the progress, the amount and flexibility of funding considerably declined in the 1970s (Crevier, 1993). The following yea2rs are considered the setback years or the first AI Winter. In the 1980s, there was a renewed interest in AI with the advent of expert systems (Crevier, 1993). Outside the United States (US) and the United Kingdom, Japan began to invest in the field (Shapiro, 1983). This period saw a great interest in knowledge representation and the revival of the interest in neural networks (McCorduck, 2004). The period is also characterized by dramatically increasing commercial interest. However, commercial vendors failed to develop workable solutions for real-world problems. The late 1980s and early 1990s also see hundreds of AI companies shutting down and the funding for AI dramatically decreasing once again (Newquist, 1994). The late 1980s and early 1990s are popularly known as the AI Winter or the second AI Winter.”

#Schuchmann, Sebastian (2019): “Analyzing the Prospect of an Approaching AI Winter” Bachelor’s thesis, Darmstadt University of Applied Sciences

https://www.researchgate.net/figure/Timeline-of-the-AI-winters_fig1_333039347 

Quote: “Following the AAAI’s usage, in this work AI winter is defined as: A period of declining enthusiasm after times of success in the field of AI.

Depending on the source, the field has experienced either one or two winters to date. Howe talks about the AI winter of 1973 lasting a decade, while Nilsson only regards the years after the mid- to late 1980s as such. Even though the term was first used in 1984, it often gets applied retroactively to the times after 1974 up until the early 1980s, as similar patterns occurred.”

—Between 1950 and 2000 computers got a billion times faster, while programming became easier and widespread. 

#Mcguffie, Kendal; Henderson-Sellers, Ann (2001): “Forty years of numerical climate modeling” Int. J. Climatol., vol. 21, 1067-1109

https://www.researchgate.net/figure/Development-of-computer-power-since-1950-Speeds-are-shown-in-millions-of-instructions_fig5_228605261
https://www.pik-potsdam.de/~stefan/Lectures/modellierung/mcguffie+henderson-s-01.pdf 

—In 1972, AI could navigate a room. In 1989, it could read handwritten numbers. But it remained a fancy tool, no match for humans! Until in 1997 an AI shocked the world by beating the world champion in Chess. Proving that we could build machines that could surpass us – but we calmed ourselves because a chess bot is quite stupid. Not a flatworm, but maybe a bee, only able to perform a specialized, narrow task. But in this narrow task it is so good that no human will ever again beat AI at chess.

#Computer History Museum: “Shakey” (retrieved 2024)

https://www.computerhistory.org/revolution/artificial-intelligence-robotics/13/289

Quote: “Shakey, developed at the Stanford Research Institute (SRI) from 1966 to 1972, was the first mobile robot to reason about its actions. Shakey’s playground was a series of rooms with blocks and ramps. Although not a practical tool, it led to advances in AI techniques, including visual analysis, route finding, and object manipulation.”

#AT&T Bell Laboratories (1989): “Backpropagation Applied to Handwritten Zip Code Recognition”, Neural Comput., vol.1, 4, 541–551

https://direct.mit.edu/neco/article-abstract/1/4/541/5515/Backpropagation-Applied-to-Handwritten-Zip-Code?redirectedFrom=fulltext

Quote: “The ability of learning networks to generalize can be greatly enhanced by providing constraints from the task domain. This paper demonstrates how such constraints can be integrated into a backpropagation network through the architecture of the network. This approach has been successfully applied to the recognition of handwritten zip code digits provided by the U.S. Postal Service. A single network learns the entire recognition operation, going from the normalized image of the character to the final classification.”

#IBM: “Deep Blue” (retrieved 2024)

https://www.ibm.com/history/deep-blue 

Quote: “In 1997, IBM’s Deep Blue did something that no machine had done before. In May of that year, it became the first computer system to defeat a reigning world chess champion in a match under standard tournament controls. Big Blue’s victory in the six-game marathon against Garry Kasparov marked an inflection point in computing, heralding a future in which supercomputers and artificial intelligence could simulate human thinking.

Deep Blue derived its chess prowess through brute force computing power. It used 32 processors to perform a set of coordinated, high-speed computations in parallel. Deep Blue was able to evaluate 200 million chess positions per second, achieving a processing speed of 11.38 billion floating-point operations per second, or flops.”

—As computers continued to improve, AI became a powerful tool for more and more taks: in 2004 it drove a robot on Mars, in 2011 it began recommending Youtube videos to you. But this was only possible because humans broke down problems into easy-to-digest chunks that computers could solve quickly. 

#Jet Propulsion Laboratory, CalTech (2006): “Autonomous Navigation Results from the Mars Exploration Rover (MER) Mission” Experimental Robotics IX. Springer Tracts in Advanced Robotics, vol. 21

https://robotics.jpl.nasa.gov/media/documents/MER_ISER2004.pdf 

Quote: “In January, 2004, the Mars Exploration Rover (MER) mission landed two rovers, Spirit and Opportunity, on the surface of Mars. Several autonomous navigation capabilities were employed in space for the first time in this mission. In the Entry, Descent, and Landing (EDL) phase, both landers used a vision system called the Descent Image Motion Estimation System (DIMES) to estimate horizontal velocity during the last 2000 meters (m) of descent, by tracking features on the ground with a downlooking camera, in order to control retro-rocket firing to reduce horizontal velocity before impact. During surface operations, the rovers navigate autonomously using stereo vision for local terrain mapping and a local, reactive planning algorithm called Grid-based Estimation of Surface Traversability Applied to Local Terrain (GESTALT) for obstacle avoidance.”

#Goodrow, Cristos (2021): “On YouTube’s recommendation system" (retrieved 2024)

https://blog.youtube/inside-youtube/on-youtubes-recommendation-system/ 

Quote: “Our system then compares your viewing habits with those that are similar to you and uses that information to suggest other content you may want to watch.

[...] We’ve used recommendations to limit low-quality content from being widely viewed since 2011, when we built classifiers to identify videos that were racy or violent and prevented them from being recommended.”

#Google Inc. (2010): “The YouTube video recommendation system” Proceedings of the fourth Association for Computing Machinery conference, 293–296

https://dl.acm.org/doi/abs/10.1145/1864708.1864770

Quote: “We discuss the video recommendation system in use at YouTube, the world's most popular online video community. The system recommends personalized sets of videos to users based on their activity on the site.” 

—This is not a technical video, so we are massively oversimplifying here. In a nutshell, the sheer power of supercomputers was combined with the almost endless data collected in the information age to make a new generation of AI.

#Zhou, Lina  et al. (2017): “Machine learning on big data: Opportunities and challenges”, Neurocomputing, vol. 237, 350-361

https://www.sciencedirect.com/science/article/abs/pii/S0925231217300577  

Quote: “ML thrives on efficient learning techniques (algorithms), rich and/or large data, and powerful computing environments. Thus, ML has great potential for and is an essential part of big data analytics. This paper focuses on ML techniques in the context of big data and modern computing environments.” 

#Our world in Data & Epoch (2023)  (Open Philanthropy sponsored): “Computation used to train notable artificial intelligence systems” part of the following publication: Charlie Giattino, Edouard Mathieu, Veronika Samborska and Max Roser - “Artificial Intelligence” (retrieved 2024)

https://ourworldindata.org/grapher/artificial-intelligence-training-computation

—AI experts began drastically improving forms of AI software called neural networks, enormously huge networks of artificial neurons that start out being bad at their tasks. They then used machine learning, which is an umbrella term for many different training techniques and environments, that allows algorithms to write their own code and improve themselves. 

#Amazon Web Services: “What is a Neural Network?” (retrieved 2024) https://aws.amazon.com/what-is/neural-network/ 

Quote: “A neural network is a method in artificial intelligence that teaches computers to process data in a way that is inspired by the human brain. It is a type of machine learning process, called deep learning, that uses interconnected nodes or neurons in a layered structure that resembles the human brain. It creates an adaptive system that computers use to learn from their mistakes and improve continuously. Thus, artificial neural networks attempt to solve complicated problems, like summarizing documents or recognizing faces, with greater accuracy.”

#MathWorks: “What Is Machine Learning?” (retrieved 2024) https://www.mathworks.com/discovery/machine-learning.html 

Quote: “Machine Learning is an AI technique that teaches computers to learn from experience. Machine learning algorithms use computational methods to “learn” information directly from data without relying on a predetermined equation as a model. The algorithms adaptively improve their performance as the number of samples available for learning increases. Deep learning is a specialized form of machine learning.”

—The scary thing is that we don’t exactly know how they do it and what happens inside them. Just that it works and that what comes out the other end is a new type of AI. A capable black box of code.

#Blouin, Lou (2023): “AI's mysterious ‘black box’ problem, explained” UM-Dearborn News https://umdearborn.edu/news/ais-mysterious-black-box-problem-explained 

Quote: “But Rawashdeh says that, just like our human intelligence, we have no idea of how a deep learning system comes to its conclusions. It “lost track” of the inputs that informed its decision making a long time ago. Or, more accurately, it was never keeping track.

This inability for us to see how deep learning systems make their decisions is known as the  “black box problem,” and it’s a big deal for a couple of different reasons.”

#Rudin, Cynthia; Radin, Joanna (2019): “Why Are We Using Black Box Models in AI When We Don’t Need To? A Lesson From an Explainable AI Competition” Harvard Data Science Review vol.1, 2

https://hdsr.mitpress.mit.edu/pub/f9kuryi8/release/8 

Quote: “In machine learning, these black box models are created directly from data by an algorithm, meaning that humans, even those who design them, cannot understand how variables are being combined to make predictions. Even if one has a list of the input variables, black box predictive models can be such complicated functions of the variables that no human can understand how the variables are jointly related to each other to reach a final prediction.”

—In 2014, Facebook AI could identify faces with 97% accuracy. In 2016 an AI beat the best humans in the incredibly complex game of Go.  In 2018, a self-learning AI learned chess in four hours just by playing against itself – and then defeated the best specialized chess bot. 

#Taigman, Yaniv et al. (2014): “DeepFace: Closing the Gap to Human-Level Performance in Face Verification” IEEE Conference on Computer Vision and Pattern Recognition 2014, 1701-1708

https://ieeexplore.ieee.org/document/6909616 

Quote: “Our method reaches an accuracy of 97.35% on the Labeled Faces in the Wild (LFW) dataset, reducing the error of the current state of the art by more than 27%, closely approaching human-level performance.”

#Chouard, Tanguy (2016): “The Go Files: AI computer wraps up 4-1 victory against human champion”. Nature. 

https://www.nature.com/articles/nature.2016.19575

Quote: “It’s all over at the Four Seasons Hotel in Seoul, where this morning AlphaGo wrapped up a 4-1 victory over Lee Sedol — incidentally, earning itself and its creators an honorary '9-dan professional' degree from the Korean Baduk Association.”

#DeepMind (2018): “Mastering Chess and Shogi by Self-Play with a

General Reinforcement Learning Algorithm”, Science,  vol. 362, 6419, 1140-1144
https://arxiv.org/pdf/1712.01815.pdf 

https://www.science.org/doi/10.1126/science.aar6404 

Quote: “Figure 1 shows the performance of AlphaZero during self-play reinforcement learning, as a function of training steps, on an Elo scale (10). In chess, AlphaZero outperformed Stockfish after just 4 hours (300k steps)"

—Since then machine learning has been applied to reading, image processing, solving tests and much more. Many of these AIs are already better than humans for whatever narrow task they were trained, but they still remained a simple tool. AI still didn’t seem that big of a deal for most people.

#Our world in Data & Kiela, Douwe et al. (2023): “Artificial Intelligence: Key Insights in artificial Intelligence” part of the following publication: Charlie Giattino, Edouard Mathieu, Veronika Samborska and Max Roser (2023) - “Artificial Intelligence” (retrieved 2024)

https://ourworldindata.org/artificial-intelligence

Quote: “The language and image recognition capabilities of artificial intelligence (AI) systems have developed rapidly.

This chart zooms into the last two decades of AI development. The plotted data stems from several tests in which human and AI performance were evaluated in five domains: handwriting recognition, speech recognition, image recognition, reading comprehension, and language understanding.

Within each domain, the initial performance of the AI is set to –100. Human performance is used as a baseline, set to zero. When the AI’s performance crosses the zero line, it scored more points than humans.

Just 10 years ago, no machine could reliably provide language or image recognition at a human level. However, AI systems have become much more capable and are now beating humans in these domains, at least in some tests.”

—And then came the chatbot ChatGPT. The work that went into it is massive. It trained on nearly everything written on the Internet to learn how to handle language, which it now does better than most people. It can summarise, translate and help with some maths problems. It is incredibly more broad than any other system just a few years ago, not crushing any single benchmark but all of them at once. Many large tech companies are spending billions to build powerful competitors.

#BBS Science focus (2023): “ChatGPT: Everything you need to know about OpenAI's GPT-4 tool” 

https://www.sciencefocus.com/future-technology/gpt-3 

Quote: “The model was trained using text databases from the internet. This included a whopping 570GB of data obtained from books, web texts, Wikipedia, articles and other pieces of writing on the internet. To be even more exact, 300 billion words were fed into the system.”

#Wooldridge, Michael (2023):”ChatGPT is not “true AI.” A computer scientist explains why”, BigThink.

https://bigthink.com/the-future/artificial-general-intelligence-true-ai/ 

Quote: “The data used to train GPT was 575 gigabytes of text. Maybe you don’t think that sounds like a lot — after all, you can store that on a regular desktop computer. But this isn’t video or photos or music, just ordinary written text. And 575 gigabytes of ordinary written text is an unimaginably large amount — far, far more than a person could ever read in a lifetime. Where did they get all this text? Well, for starters, they downloaded the World Wide Web. All of it. Every link in every web page was followed, the text extracted, and then the process repeated, with every link systematically followed until you have every piece of text on the web. English Wikipedia made up just 3% of the total training data.”

#Johri, Shreya; Moncada-Reid, Cynthia (2023): “The Making of ChatGPT: From Data to Dialogue”,  Harvard’s SITN Blog

https://sitn.hms.harvard.edu/flash/2023/the-making-of-chatgpt-from-data-to-dialogue/ 

Quote: “To develop a machine learning model that can understand language and generate coherent, grammatically correct sentences with contextual relevance, the training process for ChatGPT (i.e., providing the model with data so it can learn and capture the underlying rules and patterns within the data) was divided into multiple stages. The training dataset consisted of text collected from multiple sources on the internet, including Wikipedia articles, books, and other public webpages.

In the first stage, a machine learning model was developed to generate the next word in a partially complete sentence or paragraph. This next word had to not only make sense in the sentence, but also in the context of the paragraph. When humans read a piece of text, they pay attention to certain key words in the sentence, and complete the sentence based on those key words. Similarly, the model had to learn how to pay “attention” to the right words.  

[...]

It is noteworthy that GPT-3 was not trained for a specific task (such as translating languages or summarizing text), it was only trained to predict the next word. But since it was trained on a massive dataset that contained many examples of specific tasks (such as answering questions, translating languages, etc.), it acquired the ability to perform these wide varieties of tasks when prompted appropriately.” 

#Frieder, Simon at al. (2023): ”Mathematical Capabilities of ChatGPT”, Advances in Neural Information Processing Systems, vol.36, 27699-27744

Quote: “We find that ChatGPT can be used most successfully as a mathematical assistant for querying facts, acting as a mathematical search engine and knowledge base interface. GPT-4 can additionally be used for undergraduate-level mathematics but fails on graduate-level difficulty.”

#Forbes (2023): “Google Invests In Anthropic For $2 Billion As AI Race Heats Up”

https://www.forbes.com/sites/qai/2023/10/31/google-invests-in-anthropic-for-2-billion-as-ai-race-heats-up/?sh=2cb7f893664e 

Quote: “Alphabet’s Google is investing $2 billion into OpenAI rival Anthropic, adding fuel to the fire that is the AI race to the top. Google said it had made a $500 million upfront investment in Anthropic and would stump up the remaining $1.5 billion over time.

[...]
The move comes after Amazon announced it was pouring $4 billion into the AI start-up, making Google the second patron Anthropic has with bottomless pockets."

—AI is already transforming customer service, banking, healthcare, marketing, copywriting, creative spaces and more. AI-generated content has already taken hold of social media, youtube and news websites. Elections are expected to be inundated by propaganda and misinformation. No-one is sure how much good or harm can come from adopting AI everywhere.

#Adam, Martin;  Wessel, Michael; Benlian, Alexander (2021): “AI-based chatbots in customer service and their effects on user compliance” Electron Markets vol. 31, 427–445

https://link.springer.com/article/10.1007/s12525-020-00414-7

Quote: “Communicating with customers through live chat interfaces has become an increasingly popular means to provide real-time customer service in many e-commerce settings. Today, human chat service agents are frequently replaced by conversational software agents or chatbots, which are systems designed to communicate with human users by means of natural language often based on artificial intelligence (AI).”

#Deutsche Bank: “How AI is changing banking” (retrieved 2024)

https://www.db.com/what-next/digital-disruption/better-than-humans/how-artificial-intelligence-is-changing-banking/index?language_id=1 

Quote: “Artificial intelligence is considered one of the technologies that can fundamentally change industries. Banking is no exception. We show three possibilities – from investing to ESG to financial crime.”

#Panch, Trishan; Mattie, Heather;  Celi, Leo A. (2019): “The “inconvenient truth” about AI in healthcare”, npj Digital Medicine vol.2, 77

https://www.nature.com/articles/s41746%E2%80%90019%E2%80%900155%E2%80%904 

Quote: “The rapid explosion in AI has introduced the possibility of using aggregated healthcare data to produce powerful models that can automate diagnosis6 and also enable an increasingly precision approach to medicine by tailoring treatments and targeting

resources with maximum effectiveness in a timely and dynamic manner.”

#Behera, Rajat K.(2020): “Personalized digital marketing recommender engine”, Journal of Retailing and Consumer Services, vol. 53, 101799

https://www.sciencedirect.com/science/article/abs/pii/S0969698918307987 

Quote: “E-business leverages digital channels to scale its functions and services and operates by connecting and retaining customers using marketing initiatives. To increase the likelihood of a sale, the business must recommend additional items that the customers may be unaware of or may find appealing. Recommender Engine (RE) is considered to be the preferred solution in these cases for reasons that include delivering relevant items, hence improving cart value, and boosting customer engagement.”

#The Chartered Institute of Marketing: “AI for copywriting: Course overview” (retrieved 2024)

https://www.cim.co.uk/training/list-courses/ai-for-copywriting/ 

Quote: “AI is reengineering the DNA of the creative process to save you time, money and effort when producing marketing content. Innovative technology puts you firmly in control of the creative process in ways that were inconceivable until now. This workshop will show you how AI copywriting tools help develop your natural writing style and creative marketing skills to create persuasive long-form and short-form content that achieves excellent results.”

#NYU Libraries: “Machines and Society: Image generation tools” (retrieved 2024)

https://guides.nyu.edu/data/ai-image-generation 

Quote: “A.I. image generators are computer programs that use deep learning algorithms to produce digital images from scratch (usually text) or modify existing ones (usually images). These generators can create highly realistic and complex images, including landscapes, faces, objects, and more. They have practical applications in various fields, such as art, design, advertising, and gaming.”

#Yang, Kai-Cheng; Singh,  Danishjeet; Menczer, Filippo (2024): “Characteristics and prevalence of fake social media profiles with AI-generated faces”

https://arxiv.org/pdf/2401.02627.pdf 

Quote: “In this paper, we present a systematic analysis of Twitter(X) accounts using human faces generated by Generative Adversarial Networks (GANs) for their profile pictures. We present a dataset of 1,353 such accounts and show that they are used to spread scams, spam, and amplify coordinated messages, among other inauthentic activities. Leveraging a feature of GAN-generated faces — consistent eye placement — and supplementing it with human annotation, we devise an effective method for identifying GAN-generated profiles in the wild. Applying this method to a random sample of active Twitter users, we estimate a lower bound for the prevalence of profiles using GAN-generated faces between 0.021% and 0.044% — around 10K daily active accounts.”

#Lyu, Yao et al. (2024): ”A Preliminary Exploration of YouTubers’ Use of Generative-AI in

Content Creation”. Preprint. 

https://arxiv.org/pdf/2403.06039.pdf 

Quote: “Content creators increasingly utilize generative artificial intelligence (Gen-AI) on platforms such as YouTube, TikTok, Instagram, and various blogging sites to produce imaginative images, AI generated videos, and articles using Large Language Models (LLMs).”

#NewsGuard (2024): “Tracking AI-enabled Misinformation: 790 ‘Unreliable AI-Generated News’ Websites (and Counting), Plus the Top False Narratives Generated by Artificial Intelligence Tools”

https://www.newsguardtech.com/special-reports/ai-tracking-center/ 

Quote: “To date, NewsGuard’s team has identified 790 Unreliable AI-Generated News and information websites, labeled “UAINS,” spanning 16 languages: Arabic, Chinese, Czech, Dutch, English, French, German, Indonesian, Italian, Korean, Portuguese, Russian, Spanish, Tagalog, Thai, and Turkish.”

#APNews (2024): “Here’s how ChatGPT maker OpenAI plans to deter election misinformation in 2024”

https://apnews.com/article/ai-election-misinformation-voting-chatgpt-altman-openai-0e6b22568e90733ae1f89a0d54d64139 

Quote: “ChatGPT maker OpenAI has outlined a plan to prevent its tools from being used to spread election misinformation as voters in more than 50 countries prepare to cast their ballots in national elections this year.”

#CNN Business: “OpenAI sets rules to combat election misinformation. It’s been tried before.”

https://edition.cnn.com/2024/01/16/tech/openai-election-misinformation/index.html 

Quote: “As concerns swirl about the disruption artificial intelligence could cause for the 2024 elections, OpenAI on Monday declared that politicians and their campaigns are not allowed to use the company’s AI tools.

[...]

The announcement shows how OpenAI is attempting to get ahead of criticism that artificial intelligence — which has already been used this election cycle to disseminate fake images — could undermine the democratic process with computer-generated disinformation.

OpenAI’s policies echo those implemented by other large tech platforms. But even social media firms that are much bigger than OpenAI, and that dedicate massive teams to election integrity and content moderation, have often shown that they struggle to enforce their own rules. OpenAI is likely to be no different — and a lack of federal regulation is forcing the public to simply take the companies at their word.”

—All these potential gains or risks are just the result of today’s AI. ChatGPT’s intelligence is a major step up, but it remains narrow. While it can write a great essay in seconds, it doesn’t understand what it is writing

#Mahowald, Kyle at al.  (2024): “Dissociating language and thought in large language models”, Trends in Cognitive Sciences (In Press)

https://www.sciencedirect.com/science/article/abs/pii/S1364661324000275 

https://arxiv.org/abs/2301.06627 

Quote: “Large Language Models (LLMs) have come closest among all models to date to mastering human language, yet opinions about their linguistic and cognitive capabilities remain split. Here, we evaluate LLMs using a distinction between formal linguistic competence—knowledge of linguistic rules and patterns—and functional linguistic competence—understanding and using language in the world.

We ground this distinction in human neuroscience, which has shown that formal and functional competence rely on different neural mechanisms. Although LLMs are surprisingly good at formal competence, their performance on functional competence tasks remains spotty and often requires specialized fine-tuning and/or coupling with external modules. We posit that models that use language in humanlike ways would need to master both of these competence types, which, in turn, could require the emergence of mechanisms specialized for formal linguistic competence, distinct from functional competence.”

#Wooldridge, Michael (2023):”ChatGPT is not “true AI.” A computer scientist explains why”, BigThink.

https://bigthink.com/the-future/artificial-general-intelligence-true-ai/ 

Quote: “The first thing we notice when we use ChatGPT or BARD is that they are extremely good at generating very natural text. That is no surprise; it’s what they are designed to do, and indeed that’s the whole point of those 575 gigabytes of text. But the unexpected thing is that, in ways that we don’t yet understand, LLMs acquire other capabilities as well: capabilities that must be somehow implicit within the enormous corpus of text they are trained on.

For example, we can ask ChatGPT to summarize a piece of text, and it usually does a creditable job. We can ask it to extract the key points from some text, or compare pieces of text, and it seems pretty good at these tasks as well.

[...]

But their dazzling competence in human-like communication perhaps leads us to believe that they are much more competent at other things than they are. They can do some superficial logical reasoning and problem solving, but it really is superficial at the moment. But perhaps we should be surprised that they can do anything beyond natural language processing. They weren’t designed to do anything else, so anything else is a bonus — and any additional capabilities must somehow be implicit in the text that the system was trained on.

For these reasons, and more, it seems unlikely to me that LLM technology alone will provide a route to “true AI.”

[...]

This doesn’t mean they aren’t impressive (they are) or that they can’t be useful (they are). And I truly believe we are at a watershed moment in technology. But let’s not confuse these genuine achievements with “true AI.” LLMs might be one ingredient in the recipe for true AI, but they are surely not the whole recipe — and I suspect we don’t yet know what some of the other ingredients are.”

—What makes humans different from AI is our general intelligence. Humans can technically absorb any piece of knowledge and start working on any problem. We are great at many very different skills and tasks, from playing chess to writing or solving science puzzles – not equally of course. Some of us are experts in some fields and beginners in others, but we can technically do all of them.

#University of Wolverhampton (2023): “What are the different types of artificial intelligence?”

https://online.wlv.ac.uk/what-are-the-different-types-of-artificial-intelligence/ 

Quote: “Artificial general intelligence (AGI) aims to perform intellectual tasks in the way that a human can. Also known as strong AI, AGI aims to learn and adapt to new situations, just like a person would, and not be limited to one specific task or area. Instead, it should be applied across various fields.

General artificial intelligence has potential applications in robotics, where machines can think and make decisions on their own – making them more efficient and productive – but it could also revolutionise industries from healthcare to transportation.

General AI is what artificial intelligence experts are currently working towards.”

#Amazon Web Services: “What Is AGI?” (retrieved 2024) 

https://aws.amazon.com/what-is/artificial-general-intelligence/ 

Quote: “Artificial general intelligence (AGI) is a field of theoretical AI research that attempts to create software with human-like intelligence and the ability to self-teach. The aim is for the software to be able to perform tasks that it is not necessarily trained or developed for. 

Current artificial intelligence (AI) technologies all function within a set of pre-determined parameters. For example, AI models trained in image recognition and generation cannot build websites. AGI is a theoretical pursuit to develop AI systems that possess autonomous self-control, a reasonable degree of self-understanding, and the ability to learn new skills. It can solve complex problems in settings and contexts that were not taught to it at the time of its creation. AGI with human abilities remains a theoretical concept and research goal. 

[...]

Over the decades, AI researchers have charted several milestones that significantly advanced machine intelligence—even to degrees that mimic human intelligence in specific tasks. For example, AI summarizers use machine learning (ML) models to extract important points from documents and generate an understandable summary. AI is thus a computer science discipline that enables software to solve novel and difficult tasks with human-level performance. 

In contrast, an AGI system can solve problems in various domains, like a human being, without manual intervention. Instead of being limited to a specific scope, AGI can self-teach and solve problems it was never trained for. AGI is thus a theoretical representation of a complete artificial intelligence that solves complex tasks with generalized human cognitive abilities.”

—In the past AI was narrow and able to become good at one skill but was extremely bad in all the others. Simply by building faster computers and pouring more money into AI training will get us new, more powerful generations of AI.  

Training AI takes a lot of computational power. Building faster computers will give them more computational power, and lead to better AI.

#Baidu Research (2017): “Deep Learning Scaling is Predictable, Empirically”, Preprint.
https://arxiv.org/pdf/1712.00409.pdf 

Quote: “First, there is a close tie from compute operation rate (e.g., floating point operations, or "FLOPs") to model accuracy improvements. Power-law learning curves and model size growth indicate that each new hardware generation with improved FLOP rate can provide a predictable step function improvement in relative DL [Deep Learning] model accuracy.”

#Amodei, Dario; Hernandez, Danny (2018): “AI and compute”. Open AI Research (sponsored by Open Philanthropy) 

https://openai.com/research/ai-and-compute 

Quote: “Three factors drive the advance of AI: algorithmic innovation, data (which can be either supervised data or interactive environments), and the amount of compute available for training. Algorithmic innovation and data are difficult to track, but compute is unusually quantifiable, providing an opportunity to measure one input to AI progress. Of course, the use of massive compute sometimes just exposes the shortcomings of our current algorithms. But at least within many current domains, more compute seems to lead predictably to better performance, and is often complementary to algorithmic advances.”

Of course, computational power costs money, so additional investment would also be needed to access this computational power.

#Amodei, Dario; Hernandez, Danny (2018): “AI and compute”. Open AI Research (sponsored by Open Philanthropy) 

https://openai.com/research/ai-and-compute

Quote: “Many hardware startups are developing AI-specific chips, some of which claim they will achieve a substantial increase in FLOPS/Watt (which is correlated to FLOPS/$) over the next 1–2 years. There may also be gains from simply reconfiguring hardware to do the same number of operations for less economic cost.

 [...]

Therefore, if sufficient economic incentive exists, we could see even more massively parallel training runs, and thus the continuation of this trend for several more years.”

—We don’t know how to build AGI, how it will work or what it will be able to do. But since narrow AIs today are already capable of mastering one mental task quickly, AGI might be able to do the same with all mental tasks. So even if it starts out stupid, an AGI might be able to become as smart and capable as a human. The scary thing is this could happen suddenly. 

#Latif, Ehsan et al (2023): “AGI: Artificial General Intelligence for Education” (Preprint)

https://arxiv.org/pdf/2304.12479.pdf 

Quote: “AGI usually refers to machine intelligence that possesses human-like cognitive abilities. For instance, an AGI agent shall be capable of understanding, learning, and carrying out any intellectual work that a human person is capable of (Legg et al., 2007). AGI

systems mimic humans’ general-purpose problem-solving abilities (Wang, 2019). The

ability of AGI systems to function autonomously, making judgments and conducting

actions without the need for ongoing human supervision, is one of these features.

Thanks to this degree of autonomy, AGI may work well in complicated, dynamic

situations, enabling it to adjust to unforeseen conditions (Zhai et al., 2021). AGI may

solve problems and carry out activities in multiple domains without being restricted

to a single area of competence by accumulating information and abilities in a general purpose manner (Lake et al., 2017).”

#Sarma, Gopal P. (2018): “Brief Notes on Hard Takeoff, Value Alignment, and Coherent Extrapolated Volition” (retrieved 2024)
https://arxiv.org/abs/1704.00783 

Quote: “The distinction between hard takeoff and soft takeoff has been used to describe different possible scenarios following the arrival of human-level artificial intelligence. The basic premise underlying these concepts is that software-based agents would have the ability to improve their own intelligence by analyzing and rewriting their source code, whereas biological organisms are significantly more restricted in their capacity for self-improvement.

There is no precise boundary between the two scenarios, but in broad strokes, a hard takeoff refers to a transition from human level intelligence to superintelligence in a matter of minutes, hours or days. A soft takeoff refers to a scenario where this transition is much more gradual, perhaps taking many months or years. The practical importance of this qualitative distinction is that in a soft takeoff, there may opportunities for human intervention in the event that the initial AI systems have problematic design flaws.”

—While this sounds like science fiction, most AI researchers think this will happen some time this century, maybe already in a few years.

#Zhang, Baobao et al. (2022): ”Forecasting AI Progress: Evidence from a Survey of Machine Learning Researchers”. Preprint. 

https://arxiv.org/abs/2206.04132 

Quote: “We present the results below for human-level machine intelligence forecasts for 1) our main cross-sectional sample consisting of 289 authors that published at the 2018 NeurIPS and ICML conferences and 2) a panel sample made up of 49 respondents recontacted from a survey by Grace et al. (2018) conducted in 2016. The cross-sectional sample respondents were asked to forecast when human-level machine intelligence would exist, defined as “when machines are collectively able to perform almost all tasks (> 90% of all tasks) that are economically relevant better than the median human paid to do that task.” To be consistent with the prior survey phrasing, we asked the panel sample respondents about high-level machine intelligence, as in 2016, which was defined as when “unaided machines can accomplish every task better and more cheaply than human workers” and were asked to assume that scientific activity would continue without major negative disruption.

This question was asked in either a fixed years or fixed probabilities framing. The fixed years framing asked for their subjective probability that HLMI would exist in a certain number of years (for the cross-sectional sample: 10, 20, 50 years, panel sample: 10, 20, 40 years) The fixed probabilities framing presented three probabilities (10%, 50%, and 90%) and asked both cross-sectional and panel sample respondents in how many years they would expect human or high-level machine intelligence to exist with that probability.

For the cross-sectional sample, the aggregate forecast across both framings for when high-level machine intelligence would exist with 50% probability was 2060, according to the median parameters aggregation method. As Fig. 1 shows, the aggregate forecast for the cross-sectional sample in 2019 varied somewhat depending on which aggregation methodology was used.5 For the parametric aggregation methods in Fig. 1, we fit each respondent’s forecasts to cumulative distribution functions (CDFs) following the gamma distribution, before aggregating these CDFs.” 

#Gruetzemacher, Ross;  Paradice, David; Lee, Kang-Bok (2019): ”Forecasting Transformative AI: An Expert Survey”. Preprint.

https://arxiv.org/ftp/arxiv/papers/1901/1901.08579.pdf 

Quote: “Median forecasts indicated a 50% probability of AI systems being

capable of automating 90% of current human tasks in 25 years and 99% of current

human tasks in 50 years”

There is also an economic angle to the arrival of AGI predictions that has received significant attention. An example of an argument of this kind is given in the following source:

#How Soon is Now? Predicting the Expected Arrival Date of AGI- Artificial General Intelligence (2023)

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4496418 

Quote: ”This paper uses some economic modelling techniques to predict the expected arrival date of AGI- Artificial General Intelligence. The average predicted date from this analysis is 2041, with a likely range of 2032 to 2048, and an estimated earliest possible arrival date of 2028, i.e. just 5 years away.” 

However, other researchers oppose this view and think that AGI will take significantly longer to develop, or is unrealizable in principle.

#Fjelland, Ragnar (2020): ”Why general artificial intelligence will not be realized”, Nature Humanities and Social Sciences Communications, vol.7, 1 https://www.nature.com/articles/s41599-020-0494-4 

Quote: “A closer look reveals that although development of artificial intelligence for specific purposes (ANI) has been impressive, we have not come much closer to developing artificial general intelligence (AGI). The article further argues that this is in principle impossible, and it revives Hubert Dreyfus’ argument that computers are not in the world.” 

—Today’s AIs like ChatGPT already think and solve the tasks they were made for at least ten times faster than even very skilled humans. Maybe AGI will be slower but it may also be faster, maybe much faster. And since AGIs are software, you could copy them endlessly as long as you have enough storage and run them in parallel. 

#Philosophy Dictionary of Arguments (2014): “Nick Bostrom on Software” (source former director of Future of Humanity Institute, sponsored by Open Philanthropy)
https://philosophy-science-humanities-controversies.com/listview-details.php?id=2627748&a=a&first_name=Nick&author=Bostrom&concept=Software 

Quote: “Software/superintelligence/Bostrom: advantages for digital intelligences:

[...]

- Duplicability: With software, one can quickly make arbitrarily many high-fidelity copies to fill the available hardware base.”

Duplicability is certainly an advantage when transmitting and preserving software. However, running AI software is currently expensive:

#STTelemedia Global Data Centres (2023): “The Real Costs of AI”

https://www.sttelemediagdc.com/resources/real-costs-ai 

Quote: “Training these state-of-the-art AI models demands vast computational resources. Generative AI training needs large graphics processing unit (GPU) clusters – a modern-day supercomputer – working overtime, consuming large amounts  of energy. The costs involved are not for the faint-hearted either, with one research house forecasting that generative AI data server infrastructure plus operating costs will exceed US$76 billion by 2028.”

#SemiAnalysis (2023): ”The Inference Cost Of Search Disruption – Large Language Model Cost Analysis”

https://www.semianalysis.com/p/the-inference-cost-of-search-disruption 

Quote: “Disruption and innovation in search don’t come for free. The costs to train an LLM, as we detailed here, are high. More importantly, inference costs far exceed training costs when deploying a model at any reasonable scale. In fact, the costs to inference ChatGPT exceed the training costs on a weekly basis.

[...]

Estimating ChatGPT costs is a tricky proposition due to several unknown variables. We built a cost model indicating that ChatGPT costs $694,444 per day to operate in compute hardware costs. OpenAI requires ~3,617 HGX A100 servers (28,936 GPUs) to serve Chat GPT. We estimate the cost per query to be 0.36 cents.”

But technological advancements in computing power or in AI architecture could make AGI software faster and cheaper to run, making the idea of running many at the same time more feasible. 

—There are 8 million scientists in the world – Now imagine an AGI, copied a million times and put to work. Imagine one million scientists working 24/7, thinking ten times faster than humans, without being distracted, only focused on the task they have been given. 

#UNESCO Science Report (2021): “Statistics and resources” (retrieved 2024)

https://www.unesco.org/reports/science/2021/en/statistics 

Quote: “Between 2014 and 2018, the researcher pool grew three times faster (13.7%) than the global population (4.6%). This translates into 8.854 million full-time equivalent (FTE) researchers by 2018.”

—What if suddenly AGI could do all intelligence based jobs in the world, from interpreting law, to coding to creating animated youtube videos – better, faster and much cheaper than humans? Would whoever controls this AGI suddenly own the economy?

#Nordhaus, William D. (2021): “Are We Approaching an Economic Singularity? Information Technology and the Future of Economic Growth” American Economic Journal: Macroeconomics, vol. 13, 1, 299-332

https://www.aeaweb.org/articles?id=10.1257/mac.20170105 

https://www.nber.org/system/files/working_papers/w21547/w21547.pdf 

Quote: “As growth accelerates with superintelligent capital, the rate of return on capital and real interest rates fall to zero.[...]

In this outcome, we thus would see the euthanasia of the laboring classes in the sense that all of income eventually goes to the owners of capital. Workers would be well-paid but would control a vanishing part of national output through the fruits of their labor. However, as long as corporations own most of the capital, and people or human institutions (including governments through taxation) own corporations, capital income will indirectly flow through to humans.Since national income equals national output, average income will be growing

increasingly rapidly. 

How this will play out in terms of individual equality or inequality goes beyond

economics to politics, tax and benefit systems, and the nature of dynastic savings. It

is clear that the Piketty condition for growing inequality (that r > g) definitely will

not hold, but beyond that little is clear.

Will the incomes be captured by the Schumpeterian classes – the innovators who design machines and write software for them? Or by the wealthy who subvert institutions to increase their wealth? By those who are the last humans who are complements rather than substitutes for information, perhaps as gardeners or butlers? Perhaps by those who control the intelligent machines before they take over?” 

#Naudé, Wim;  Dimitri, Nicola (2018): “‘Artificial general intelligence’, economic development and public policy”, Global Dev.

https://globaldev.blog/artificial-general-intelligence-economic-development-and-public-policy/ 

Quote: “While narrow AI can bring many benefits to economic development (as discussed at a recent World Bank Conference on ‘Artificial Intelligence for Development’), it is the potential invention of an AGI that is seen by some as a future game-changer for economic development.

The point at which AGI will exceed human intelligence is termed the ‘singularity’ by Raymond Kurzweil. Some economists, such as William Nordhaus, believe that after this, ‘economic growth will accelerate sharply as an ever-increasing pace of improvements cascade through the economy’.

But there is one major caveat. Whichever high-tech firm or government lab succeeds in inventing the first AGI will obtain a world-dominating technology. This potential ‘winner-takes-all’ prize raises the specter of a competitive (arms) race for an AGI.”

—Intelligence and knowledge build and accelerate each other but humans are limited by biology and evolution. Once we evolved the right hardware, our software outpaced evolution by orders of magnitudes and within a heartbeat we ruled this planet. But our software basically hasn’t changed much since then, which is why we have obesity and destroy the climate for short term gains.

#Wurz, Sarah (2014): “The Transition to Modern Behavior”. Nature Education Knowledge vol. 3, 10, 15

https://www.nature.com/scitable/knowledge/library/the-transition-to-modern-behavior-86614339/ 

Quote: “One of the major issues in palaeoanthropology and archaeology is when our hominin ancestors became like us. Humans living today have developed the capacity for ‘modern behavior'. Modern behavior can be recognized by creative and innovative culture, language, art, religious beliefs, and complex technologies (d'Errico & Stringer 2011). 

[...]
The relatively sudden appearance of such traits as a group or package in the archaeological record of the European Upper Palaeolithic has been interpreted as evidence for the onset of modern behavior (Klein 2008). 

[...]

The evolution of modern planning capabilities can be investigated by analyzing the decision steps used to produce ancient tools. There are, for example, many different ways of making stone artifacts. Over the last 200,000 years a variety of reduction techniques in different combinations were used to make stone tools. This resulted in different techno-complexes, but all with the same degree of complexity. This may mean that humans had essentially modern cognitive capabilities for this period of time (Shea 2011).”

#King, Bruce M. (2017): “The Brain, the Environment, and Human Obesity: An Evolutionary Perspective on the Difficulty with Maintaining Long-Term Weight Loss” from “Adiposity - Epidemiology and Treatment Modalities” edited by Jan Oxholm Gordeladze.

https://www.intechopen.com/chapters/52120 

Quote: “The dramatic increase in obesity within one or two generations cannot possibly be due to a change in genetics. It is the processing, distribution, and availability of foods that have changed, not the brain. For most people, in the presence of pleasant tasting (high calorie) foods, the brain’s reward circuitry overwhelms the satiety signals. For 2 million years, overeating (on the occasional basis when that was possible) had adaptive value, and it has only been since the rise of an omnipresent obesogenic environment that such behavior has become maladaptive, resulting in widespread obesity.”

#Sörqvist, Patrik; Langeborg, Linda (2019): “Why People Harm the Environment Although They Try to Treat It Well: An Evolutionary-Cognitive Perspective on Climate Compensation”, Frontiers in Psychology, vol. 10

https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2019.00348/full 

Quote: “Anthropogenic climate changes stress the importance of understanding why people harm the environment despite their attempts to behave in climate friendly ways. This paper argues that one reason behind why people do this is that people apply heuristics, originally shaped to handle social exchange, on the issues of environmental impact. Reciprocity and balance in social relations have been fundamental to social cooperation, and thus to survival, and therefore the human brain has become specialized by natural selection to compute and seek this balance. When the same reasoning is applied to environment-related behaviors, people tend to think in terms of a balance between “environmentally friendly” and “harmful” behaviors, and to morally account for the average of these components rather than the sum.”

—Since AGI is software on a computer, it may be able to change its own fundamental nature and apply its intelligence to itself. What if it uses its intelligence to make itself more intelligent? Could AGI start with intelligence close to us, improve itself to become as smart as the smartest humans and then blow past that? 

#Yampolskiy, Roman V. (2015): “On the Limits of Recursively Self-Improving AGI”, Papers of the International Conference on Artificial General Intelligence 2015, 394–403 

https://link.springer.com/chapter/10.1007/978-3-319-21365-1_40  

https://agi-conf.org/2015/wp-content/uploads/2015/07/agi15_yampolskiy_limits.pdf  

Quote: “A particular type of self-improvement known as Recursive Self-Improvement (RSI) is fundamentally different as it requires that the system not only get better with time, but that it gets better at getting better. A truly RSI system is theorized not to be subject to diminishing returns, but would instead continue making significant improvements and such improvements would become more substantial with time. Consequently, an RSI system would be capable of open ended self-improvement. As a result, it is possible that unlike with standard self-improvement, in RSI systems from generation-to-generation most source code comprising the system will be replaced by different code.”

#Creighton, Jolene (2019): “The Unavoidable Problem of Self-Improvement in AI: An Interview with Ramana Kumar, Part 1”. Future of Life Institute Blog. (sponsored by Open Philanthropy)

https://futureoflife.org/ai/the-unavoidable-problem-of-self-improvement-in-ai-an-interview-with-ramana-kumar-part-1/

Quote: “Researchers think that self-improving machines could ultimately lead to AGI because of a process that is referred to as “recursive self-improvement.” The basic idea is that, as an AI system continues to use recursive self-improvement to make itself smarter, it will get increasingly better at making itself smarter. This will quickly lead to an exponential growth in its intelligence and, as a result, could eventually lead to AGI.”

—Some AI researchers fear this process may be inevitable and worse: incredibly fast. Maybe just months or years after the first self improving AGI is switched on. Maybe it would actually take decades. We simply don’t know. But such an intelligence explosion might lead to a true superintelligent entity. 

#Machine Intelligence Research Institute: “Intelligence Explosion FAQ” (retrieved 2024) (sponsored by Open Philanthropy)

https://intelligence.org/ie-faq/#WhatIsTheIntelligence

Quote: ”Computers remain far short of human intelligence, but the resources that aid AI design are accumulating (including hardware, large datasets, neuroscience knowledge, and AI theory). We may one day design a machine that surpasses human skill at designing artificial intelligences. After that, this machine could improve its own intelligence faster and better than humans can, which would make it even more skilled at improving its own intelligence. This could continue in a positive feedback loop such that the machine quickly becomes vastly more intelligent than the smartest human being on Earth: an ‘intelligence explosion’ resulting in a machine superintelligence.”

#Sarma, Gopal P. (2018): “Brief Notes on Hard Takeoff, Value Alignment, and Coherent Extrapolated Volition” (retrieved 2024)

https://ar5iv.labs.arxiv.org/html/1704.00783

Quote: “The distinction between hard takeoff and soft takeoff has been used to describe different possible scenarios following the arrival of human-level artificial intelligence. The basic premise underlying these concepts is that software-based agents would have the ability to improve their own intelligence by analyzing and rewriting their source code, whereas biological organisms are significantly more restricted in their capacity for self-improvement [1, 2, 3, 4, 5].[...]

If self-improving AI systems are thought to be the intellectual analogue of nuclear chain reactions, then the natural image of an intelligence explosion that this metaphor creates is a scenario in which massive, disruptive changes take place in the world that are difficult for individuals and for society to handle.”

—We don’t know what such a being would look like, what its motives or goals would be, what would go on in its inner world. We could be as laughably stupid to a superintelligence as squirrels are to us. Unable to even comprehend its way of thinking.

The matter of whether and how machines could achieve consciousness and have motives or inner experience, as well as its relationship with machine intelligence, is still contested. There are two alternative views based on two different conceptions of consciousness:  the integrated information theory (IIT) and the global neuronal workspace (GNW) theory.

#Maillé, Sébastien ; Lynn, Michael (2020): “Reconciling Current Theories of Consciousness.” The Journal of neuroscience : the official journal of the Society for Neuroscience, vol. 40, 10, 1994-1996

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7055139/ 

Quote: “The IIT, first proposed by Tononi (2004), focuses on defining what a conscious system should look like with respect to information processing and architecture without considering particular brain areas or temporal profiles. One prediction of IIT is that neural networks supporting consciousness must be highly interconnected, effectively integrating different components of a state into a unified experience. A crucial advantage of the IIT is that it provides a mathematical metric of irreducibility (or integration), Φ, that can be related to the level of consciousness. Proponents of IIT point to its explanatory power: for instance, it can explain why the cortex is capable of producing conscious experience while the cerebellum is not (Lemon and Edgley, 2010; Yu et al., 2015), even though the cerebellum possesses up to four times more neurons. While the IIT has not received unambiguous validation (possibly due to the abstract nature of its description of consciousness; for review, see Tononi et al. (2016)), it provides one of the most detailed accounts for the emergence of conscious experience from an information-processing network.

The GNW theory (Dehaene and Changeux, 2011), in contrast to the IIT, was empirically derived from EEG and imaging studies in humans and primates. These studies have shown that when a stimulus is presented but not consciously perceived, activation can be seen mainly in the associated primary sensory cortices. When the stimulus is consciously perceived, however, activation in primary cortical areas is followed by a delayed “neural ignition” in which a sustained wave of activity propagates across prefrontal and parietal association cortices. According to the GNW model, this allows relevant information to be broadcast across the brain to other subsystems for use in decision-making, reporting, memory consolidation, and other processes. Thus, while IIT focuses on abstract connectivity and information-processing structure, GNW proposes a concrete spatiotemporal locus for conscious processes.”

To learn more about these views and their relationship to machine consciousness, we recommend the following reading:

#Koch, Kristof (2023): “What Does It ‘Feel’ Like to Be a Chatbot?”, Scientific American.
https://www.scientificamerican.com/article/what-does-it-feel-like-to-be-a-chatbot/ 

#Maillé, Sébastien ; Lynn, Michael (2020): “Reconciling Current Theories of Consciousness.”, The Journal of neuroscience, vol. 40, 10, 1994-1996

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7055139/  

—We will look at some of these potential futures in more videos but let us wrap up now. The only thing we know for sure is that today, right now, many of the largest and richest companies in the world are racing to create ever more powerful AIs. Whatever our future is, we are running towards it.

#Altman, Sam (2023): “Planning for AGI and beyond” OpenAI blog (retrieved 2024) (sponsored by Open Philanthropy) 

https://openai.com/blog/planning-for-agi-and-beyond 

Quote: “First, as we create successively more powerful systems, we want to deploy them and gain experience with operating them in the real world. We believe this is the best way to carefully steward AGI into existence—a gradual transition to a world with AGI is better than a sudden one. We expect powerful AI to make the rate of progress in the world much faster, and we think it’s better to adjust to this incrementally.”

#Google DeepMind: “Our vision” (retrieved 2024)

https://deepmind.google/about/ 

Quote: “We live in an exciting time when AI research and technology are delivering extraordinary advances.

In the coming years, AI — and ultimately artificial general intelligence (AGI) — has the potential to drive one of the greatest transformations in history.

We’re a team of scientists, engineers, ethicists and more, working to build the next generation of AI systems safely and responsibly.”

#Lee,Kevin; Gangidi, Adi; Oldham, Mathew (2024): “Building Meta’s GenAI Infrastructure” Engineering at Meta. (retrieved 2024)

https://engineering.fb.com/2024/03/12/data-center-engineering/building-metas-genai-infrastructure/ 

Quote: “Meta’s long-term vision is to build artificial general intelligence (AGI) that is open and built responsibly so that it can be widely available for everyone to benefit from. As we work towards AGI, we have also worked on scaling our clusters to power this ambition. The progress we make towards AGI creates new products, new AI features for our family of apps, and new AI-centric computing devices.”