Content uploaded by David Vivancos
Author content
All content in this area was uploaded by David Vivancos on Dec 24, 2024
Content may be subject to copyright.
* Equal Contribuon
Qubic AGI Journey
Human and Arcial Intelligence: Toward an AGI with Aigarth
Jose Sanchez *
1 Qubic Scienc Advisor
2 Universidad Internacional de La Rioja (UNIR)
jose.sanchezgarcia@unir.net
David Vivancos *
1 Qubic Scienc Advisor
2 Arciology Research
vivancos@vivancos.com
Abstract:
We present an integrated analysis of human
intelligence and its arcial counterpart
through Qubic’s Aigarth framework,
beginning by examining the biological
foundaons of intelligence, from
Spearman's g factor to contemporary
neuroscienc understanding of predicve
processing in the brain. The analysis
encompasses the evoluon of human
cognive abilies, parcularly focusing on
the social brain hypothesis and the role of
neural eciency in informaon processing.
Building upon these biological insights, we
introduce Aigarth, a novel approach in the
journey towards arcial general
intelligence (AGI) development that
represents a paradigm shi from tradional
GPU-dependent architectures to CPU-based
distributed compung systems. We
introduce a ternary compung paradigm
that extends beyond convenonal binary
systems, incorporang TRUE, FALSE, and
UNKNOWN states for enhanced informaon
processing. The framework employs
"Intelligent Tissue," a self-modifying neural
network structure that evolves through
natural selecon principles, to build
emergent problem-solving capabilies, a
novel scoring algorithm that evaluates
network performance through determinisc
connecon generaon and asynchronous
updates.
The results indicate that Aigarth's
decentralized approach could potenally
overcome limitaons of current AI systems
while promong democrazed AI
development. Our research contributes to
the broader understanding of evoluonary
approaches to AGI development and
presents implicaons for future arcial
consciousness studies.
Keywords:
Arcial General Intelligence (AGI),
Intelligent Tissue, Ternary Compung,
Neural Evoluon, Decentralized AI, Self-
modifying Networks, Biological Intelligence,
Qubic
1. Introducon. On Human Intelligence
"Intelligence is what you do when you don't
know what to do" (Bereiter, 1995).
Since the early 20th century, dierenal
psychology has aempted to measure the
construct of intelligence. To idenfy children
with special educaonal needs, Alfred Binet
designed a test to assess cognive funcons
(Binet & Simon, 1905). Based on Binet's
work, Charles Spearman was the rst author
to dene a general mental capacity
underlying various cognive tasks, later
termed as a theorecal construct
"g", represenng the general intelligence.
Through factor analysis, Spearman found
that performance across dierent tests and
tasks had a high correlaon, suggesng a
general capacity (Spearman, 1904).
Spearman's theory faced cricism, notably
from L.L. Thurstone, who proposed a
muldimensional approach to intelligence
(Thurstone, 1938). However, it was later
found that Thurstone’s primary abilies—
verbal uency, reasoning, memory, and
spaal percepon—also correlated,
suggesng the presence of a general
underlying factor. In 1963, Raymond Caell
proposed two components within the g
factor: uid intelligence (Gf) and crystallized
intelligence (Gc). Fluid intelligence is linked
to the ability to reason and solve new
problems, closely aligning with Bereiter’s
popular denion, while crystallized
intelligence relates to the knowledge
acquired through experience and learning
over me (Caell, 1941, 1963). This
muldimensional approach to intelligence
received support from John Horn, who
included components like short-term
memory (Gsm), spaal visualizaon (Gv),
auditory discriminaon (Ga), long-term
memory and retrieval (Glr), and processing
speed (Gs).
Building on the contribuons of Caell and
Horn, John B. Carroll proposed a three-
stratum hierarchical model aer conducng
extensive factor analyses of 460 studies
(Carroll, 1993). At Stratum 3 is g, at Stratum
2 up to 16 broad abilies appear (uid
intelligence, crystallized intelligence, long-
term memory, working memory, processing
speed, visual processing, auditory
processing, reacon and decision speed,
psychomotor processing, quantave
knowledge, reading comprehension and
uency, short-term visual memory,
ideaonal uency, rapid retrieval memory,
perceptual speed, and kinesthec
processing). Each ability subdivides into
several at Stratum 1, totaling up to 70
(Schneider & McGrew, 2012).
The CHC (Caell-Horn-Carroll) model is
currently the most widely accepted in
educaonal, research, and clinical sengs
(McGrew, 2009). It is used in the most
popular psychometric tests for evaluang
intelligence, such as the WISC-V (Wechsler,
2014), WAIS-IV (Wechsler, 2008), Woodcock-
Johnson IV (Scharnk et al. 2014), and the
Stanford-Binet 5 (Roid, 2003).
Factor analysis (FA) is the common
stascal technique used to idenfy the
underlying structure (factors) within a set of
observed variables as cognive skills tasks.
FA main objecve is to reduce the
dimensionality of the data by grouping
correlated variables into factors, which are
interpreted as unobservable latent
constructs. This method is widely applied to
uncover latent relaonships between
variables and simplify big datasets. If there
are no prior hypotheses about the number
or nature of factors, FA is exploratory.
Factor analysis is a type of stascal linear
model:
X = ΛF + ε
Where:
• X: Vector of observed variables (p
variables).
• Λ: Matrix of factor loadings (p×m,
being m the number of factors).
• F: Vector of latent factors (m
factors).
• ε: Vector of unique errors (noise).
For each observed variable, the model can
be wrien as:
Xi = λi1 F1 + λi2 F2 +
⋯
+ λim Fm + εi
Where λij represents the factor loading of
the i-th variable on the j-th factor, and εi is
the unique variance of the i-th variable.
FA starts with the correlaon matrix RRR,
expressing the correlaons among the
observed variables. Some methods as
Maximum Likelihood (ML) or Principal
Component Analysis are used to extract
factors that maximize the common variance
among variables. Depending on the quanty
of variance explained, a number of factors is
selected. Factor analysis is widely used to
test intelligence, personality and other
psychological constructs (Li et al. 2024) .
Fig. 1. CHC hierarchical model of intelligence: the g factor
1.1. Validity and Reliability of g
Every psychological measurement must
objecvely validate the inferences made
about a construct from the measuring
instrument. Tradionally, this is achieved
through two pillars: validity and reliability.
Validity encompasses not only the technical
aspects of measurement but also ethical
and social consideraons (Messick, 1989).
Considering all these aspects, the general
intelligence factor presents a high degree of
content validity, which assesses whether
items represent the construct's domain
(Jensen, 1998). It also has high criterion
validity, as g shows strong correlaons with
various levels of professional and
educaonal performance (Schmidt &
Hunter, 1998). Moreover, it demonstrates
strong construct validity, conrmed through
factor analysis for the existence of a single
factor (Carroll, 1993). The factor also
exhibits high substanve validity, indicang
how test items are based on theory and
involve the cognive tasks being evaluated
(Jensen, 1998). It has strong external
validity, allowing for generalizaon across
dierent populaons or contexts (Nisbe et
al. 2012), and consequenal validity, which
evaluates the impact of the test on decision-
making in academic or professional
environments (Sternberg, 2004).
Reliability focuses on the consistency of
measurement, evaluated through test-
retest reliability, inter-rater reliability, and
internal consistency measures (split-half
tests, Cronbach's alpha, hierarchical
omega). Cronbach’s alpha measures the
internal consistency of a set of items
measuring the same construct (Messick,
1995).
Where:
• k: Number of items in the scale.
• Si2: Variance of the i-th item.
• St2: Total variance of the scale (sum
of all item scores).
Although Cronbach’s alpha is commonly
used, hierarchical Omega ts beer for
muldimensional constructs such as
intelligence. It quanes the proporon of
total variance aributable to a general
factor in muldimensional scales.
Where:
• ω: symbol for the omega coecient
• λi: standardized factor loading of i
The construct of g has shown strong test-
retest reliability (Deary et al. 2000) and
internal consistency (Raven, 2000; Reise,
2013; Wechsler, 1997).
1.2. Predicve Value of g
The g factor is a robust predictor of behavior.
From school age, the impact of g is evident.
It is assessed by specic and standardized
tests resembling dierent cognive tasks
(spaal, logical and verbal) and situaons,
measured by a quoent, the IQ (intelligence
quoent). In a longitudinal study in the UK,
IQ assessed during childhood reliably
predicted performance in general secondary
educaon exams, with stable correlaons
between 0.60 and 0.80 (Deary et al., 2007).
The predicve value extends beyond
secondary educaon and remains
signicant in higher educaon, parcularly
in disciplines with high demands for abstract
reasoning (Deary et al., 2009).
According to a meta-analysis by Schmidt
and Hunter (1998), g correlates between
0.51 and 0.70 in the context of job
performance, parcularly when the job
requires complex cognive skills. This
esmaon also accounts for factors like
learning speed and adaptaon to dynamic
environments (Schmidt & Hunter, 1998).
Kuncel et al. (2014) suggest that employees
with higher general intelligence are more
ecient in their individual tasks and
contribute signicantly to team
performance due to their problem-solving
abilies and adaptability. Regarding remote
work, Salgado et al. (2020) show how the g
factor eecvely predicts high producvity
levels and adaptaon to new digital
plaorms. In the current context of constant
change and rapid technological evoluon,
employees with high levels of g hold
compeve advantages in terms of
adaptability, cognive demands, and
responsibility (Goredson, 2004).
In terms of physical and mental health,
individuals with higher levels of g tend to live
longer and have a lower risk of chronic
diseases. Although various explanatory
causes are suggested, g may funcon as a
mechanism that allows for beer
acquision of health-related informaon,
adherence to medical treatments, and
avoidance of risky behaviors (Goredson,
2004). Similarly, the longitudinal study by
Bay et al. (2009), involving a million men in
Sweden over 20 years, found a direct
relaonship between youth IQ and adult
mortality, even aer controlling for
socioeconomic status and lifestyle factors.
Several studies by Ian Deary nd that people
with higher general intelligence exhibit
beer adherence to treatments for chronic
diseases, beer understand medical
instrucons, and are more likely to make
lifestyle modicaons, resulng in beer
long-term outcomes (Deary et al., 2008,
2010).
Intelligence is also posively associated with
the propensity to form healthy habits, such
as sustained physical acvity, a balanced
diet, and avoiding tobacco use. The impact
of g on various life condions is already
evident from childhood. A longitudinal study
involving 33,000 parcipants over 50 years
found that a higher g value in early life was
associated with a lower risk of heart disease
and stroke (Calvin et al., 2017). Bay et al.
(2018) invesgated the relaonship
between g and cancer risk, nding that
individuals with higher childhood IQ had a
signicantly lower risk of developing various
types of cancer, potenally due to greater
adherence to prevenon programs and
healthier behavior.
Mentally, individuals with higher g exhibit a
lower incidence of mental disorders, beer
stress management, reduced prevalence of
depression and anxiety, and greater
psychological resilience (Gale et al., 2017).
When faced with traumac or stressful
events, such as illness, bereavement, or
unemployment, those with higher g
demonstrate beer coping abilies.
Regarding social intelligence, g predicts the
ability to use social networks harmoniously,
avoid scams, and discern between real and
fake news (Jackson & Wang, 2013). People
with higher g are more likely to engage in
social causes, parcipate in volunteer
programs, and understand complex social
issues (Nie et al., 1996), including polical
maers (Deary et al., 2008).
The famous Dunedin study in New Zealand,
which began in 1972-73, has followed over
1,000 individuals from birth to adulthood
over nearly 50 years, collecng and
analyzing data on health, disease,
intelligence, personality, development,
academic and professional performance,
and other social factors. The study provides
insight into the impact of intelligence-
related abilies in childhood on adult life
(Poulton et al., 2015). Its value lies in the
longitudinal measurement of the same
individuals over me. The studies conrm
that childhood IQ is a strong predictor of
future academic performance, educaonal
system adaptaon, learning ability, and
highest level of educaon achieved
(Fergusson & Horwood, 2007). It also
reveals that those with higher childhood IQs
are more likely to secure cognively
demanding jobs involving complex decision-
making and earn beer wages. In line with
other studies, the Dunedin sample research
shows that individuals with higher general
intelligence levels adopt healthier lifestyles,
are less prone to risky behaviors, beer
follow medical advice, exhibit beer general
health, suer fewer chronic diseases in
adulthood, and experience greater longevity
(Bay et al., 2007; Belsky et al., 2017).
Psychologically, they have lower rates of
depression and anxiety and possess beer
social support and trust-based relaonships
(Mo et al., 2002). Concerning criminal
and ansocial behavior, the Dunedin study
reveals that those with lower IQs struggle
more with social norms, make less raonal
decisions, and have diculty controlling
impulses (Mo, 1993).
Although cognive intelligence and
emoonal intelligence are oen separated
as if they were two dierent constructs,
social adaptaon involves emoonal
regulaon, adherence to norms, and
cooperave strategies. In Dunedin, it is
observed that those with lower g levels have
more conduct disorders in childhood and
adolescence and poorer social adaptaon in
adulthood. Conversely, individuals with
higher intellectual capacity display greater
social resilience, cognive exibility, beer
coping strategies for social and emoonal
challenges, and a higher socioeconomic
status throughout life (Mo & Caspi,
2000; Caspi, 1998; Shanahan et al., 2014).
Other studies, such as those by Robert
Hogan, demonstrate that intelligence also
applies to social and emoonal skills. People
with higher g build stable interpersonal
relaonships, cooperate, and eecvely
resolve life conicts (Hogan & Kaiser, 2005).
In romanc relaonships, those with higher
g levels empathize and understand their
partners beer, and communicate more
eciently (Roberts & Kuncel, 2007).
The ability to manage savings and nances,
plan for rerement, learn nancial literacy,
and avoid risky economic behaviors is
associated with higher g (Lusardi & Mitchell,
2007; Banks & Oldeld, 2007). Creavity,
although considered a separate construct by
some authors like Sternberg, correlates with
g. The ability for innovaon, generang new
ideas, learning quickly, and idenfying
opportunies is beer when starng from a
high g level (Shane, 2003; Kaufman &
Sternberg, 2010).
The g factor has proven to be a crucial
predictor of the tendency to experience
cognive decline and neurodegenerave
diseases. The most plausible explanaon
relates to a greater cognive reserve.
Cognive reserve is the brain's capacity to
compensate for funconal or structural
damage caused by neurodegenerave
diseases or natural aging (Livingston et al.,
2020). Even if two brains have equal ssue
damage, the person with greater cognive
reserve may show fewer and less intense
symptoms over a longer period (Stern et al.,
2019). Although educaon, occupaon,
quality social relaonships, and physical
acvity all contribute to building cognive
reserve throughout life, the g factor proves
to be an accurate predictor (Ferreira et al.,
2016; Soldan et al., 2017; Dekhtyar et al.,
2015). People with higher general
intelligence exhibit greater cognive
resilience in old age, with fewer clinical
symptoms of cognive decline (Whalley &
Deary, 2001). Regarding demena, McGurn
et al. (2008) indicate that those with higher
g in youth are more resistant to cognive
decline associated with neurodegenerave
diseases, exhibing symptoms of demena
about ve years later than average.
2. The g Factor and the Brain: Biological
Bases
Historically, since Binet and Spearman, the
study of intelligence has been approached
from psychometrics and psychology
(Spearman, 1904). With the advances in
neuroimaging techniques and genec
sequencing over the last 30 years, as well as
support for brain research in the USA and
Europe (Brain Iniave, Human Brain
Project), it has become possible to explore
the biological bases and neural correlates of
intelligence (Insel et al., 2013; Amunts et al.,
2016).
In 1991, with the rst improvements in
structural neuroimaging techniques, a 0.33
correlaon was found between brain
volume and IQ (Willerman et al., 1991),
which was later conrmed in a meta-
analysis by McDaniel (2005) with a
correlaon coecient close to 0.40. At least
within the Homo genus, subspecies with
larger brains tend to have higher g.
However, size is not the most crucial factor.
The speed of neural processing implies more
ecient mental processing of informaon.
Event-related potenals, which measure the
brain's responses to various smuli, are
faster and more synchronized in individuals
with higher g (Deary & Caryl, 1997).
Several studies highlight neural eciency.
People who show beer cognive
performance on a task exhibit a lower
demand for neural resources (Neubauer &
Fink, 2009) and greater funconal
connecvity between brain regions related
to the task (Basten et al., 2015). Brain
acvity is less diuse in people with higher g
when performing tasks, suggesng that
intelligence opmizes the use of neural
resources (Pahor et al., 2019). The neural
eciency hypothesis emerged in 1992 with
Richard Haier’s inial studies, where
individuals with higher intelligence levels
showed less acvaon and fewer resources
to solve complex cognive tasks (Haier et al.,
1992).
Haier later focused on studying the
implicated regions and their
interconnecons (Jung & Haier, 2007). This
led to the Parieto-Frontal Integraon Theory
of intelligence (P-FIT), linking g to a network
connecng the parietal and prefrontal
cortex. Previously, the importance of regions
such as the dorsolateral prefrontal cortex,
responsible for planning, decision-making,
sequenal behavior organizaon, and
cognive exibility, was studied within the
context of "execuve funcons." Other
regions, like the hippocampus involved in
memory consolidaon and learning—
primarily through spontaneous acvity in
sleep phase 2—indirectly relate to g.
However, the key to studying g in the brain
is the connecvity between regions and
eciency in processing. In a meta-analysis
that reviewed dozens of neuroimaging
studies, Basten, Hilger, and Fiebach (2015)
idened the regions correlang with
intelligence, conrming the P-FIT theory.
When individuals engage in complex
problem-solving, abstract reasoning, and
working memory acvaon, the fronto-
parietal network is parcularly relevant.
This network predominantly acvates on
the le side, as the le hemisphere has
some specializaon in language (both
comprehension and producon). Individuals
with higher intelligence show asymmetrical
acvaon, with a le-sided dominance
when solving complex mathemacal,
verbal, or spaal tasks (Jung et al., 2010).
Greater connecvity between the main
nodes of this network corresponds to beer
performance on cognive tasks (Hampshire
et al., 2012).
The g factor, therefore, links to the brain’s
capacity to eciently process informaon
within specic networks, primarily the le
fronto-parietal (Haier et al., 2009). A new
study, by Thiele et al. 2024 underscores the
distributed and evoluonarily adapve
nature of brain connecvity as a
cornerstone for human intelligence,
surpassing limitaons of the Parieto-Frontal
Integraon Theory (P-FIT).
Using machine learning to analyze data
from over 800 parcipants, predicve
models incorporang brain-wide
connecvity paerns explained up to 31% of
the variance in general intelligence,
outperforming models restricted to regions.
This novel approach highlights that
intelligence emerges from complex
interacons across mulple networks,
reecng the evoluonary renement of
cognive exibility and problem-solving
capacies.
If, as longitudinal studies like Dunedin
suggest, g in childhood is a predictor of all
kinds of performance in adulthood, it implies
that the general intelligence factor has a
strong genec basis.
To analyze the eect of genes on behavior,
samples of monozygoc twins separated at
birth, dizygoc fraternal twins, or children
adopted at a very early age are used. This
approach makes it possible to sciencally
assess the weight of nature versus nurture.
In a meta-analysis of various studies on
monozygoc twins, Plomin and Deary
(2015) found that the heritability of
intelligence increases with age, implying
that genec inuence on intelligence
becomes more prominent outside the
original family environment, underscoring
the strong impact of g in adulthood. Years
earlier, the heritability of the g factor was
esmated by studying twins raised together
or apart: between 50% and 80% of
intelligence is aributed to genecs
(Bouchard et al., 1990).
Genome-Wide Associaon Studies (GWAS)
have enabled the analysis of genec
polymorphisms associated with pathologies
or individual characteriscs by examining
the genomic associaons between specic
genec variants known as SNPs (single
nucleode polymorphisms) and the trait
under study (in this case, intelligence) within
a large populaon (Mills et al., 2019). Based
on data from the UK Biobank, numerous loci
(locaons) within the genome have been
found to have small eects on intelligence
variability. Recently, more than 500 genec
loci linked to g variability have been
discovered (Davies et al., 2018; Savage et
al., 2018).
Despite the substanal genec inuence on
intelligence, environment, nutrion,
educaon, and family background all aect
the expression and development of
intelligence (Benton, 2010). More enriched
environments, access to beer educaon
and culture, and the increasing cognive
demands of the labor market compared to
manual or mechanical skills have resulted in
an increase in intelligence throughout the
20th century. This is known as the Flynn
eect, which highlights the importance of
the environment in modulang the g factor
over me (Flynn, 1987).
2.1. From Carbon to Silicon
The journey from biological neural networks
to trying to create a replica in arcial
neural networks represents a fascinang
convergence of neuroscience and computer
science. In 1943, Warren McCulloch and
Walter Pis proposed the rst mathemacal
model of a neuron, showing that neural
events and the relaons among them could
be treated by means of proposional logic
(McCulloch and Pis, 1943).
This foundaonal work established that
neural networks of sucient complexity
could compute any logical funcon. Building
on this, Frank Rosenbla introduced the
perceptron in 1958, implemenng simple
but eecve algorithms for supervised
learning of binary classiers (Rosenbla,
1958), the perceptron mimicked a single
neuron's funcon by taking mulple inputs,
applying weights, and producing a binary
output based on a threshold, represenng
one of the earliest praccal
implementaons of neural computaon.
Despite many limitaons of the early
approaches, these models paved the way for
more complex arcial neural networks, like
current deep learning architectures, with
their mulple layers of interconnected
nodes, bear a striking resemblance to the
hierarchical structure of the human brain's
neural pathways (McClelland et al.,
1986). This biomimicry has proven
remarkably eecve, as networks can learn
representaons through dynamic,
distributed interacons within networks of
simple neuron-like processing units
(Hassabis et al., 2017). These biological
inspiraons have led to signicant
advancements in paern recognion,
natural language processing, and decision-
making capabilies of AI systems,
establishing what McClelland and
colleagues called "Parallel Distributed
Processing" (PDP), which more closely
approximates how actual neural circuits
perform computaon (McClelland et al.,
1986; Schmidhuber, 2015).
AI algorithms, at their core, are complex
computaonal processes that manipulate
and analyze data to perform tasks that
typically require human intelligence. The
eld of AI has both beneted from and
driven advancements in computaonal
power and eciency, since 2012, the
compung power used in the largest AI
training runs has grown exponenally,
increasing by roughly 10× per year
(Thompson et al., 2022). The development
of AI has been closely ed to Moore's Law
and specialized hardware acceleraon,
while tradional CPU improvements have
slowed, the introducon of GPU-based deep
learning inially yielded 5-15× speedups
which grew to more than 35× by 2012
(Thompson et al., 2022). This enabled
breakthrough achievements like AlexNet's
victory in the 2012 ImageNet compeon,
achieving a top-5 error rate of 16.4% using
deep convoluonal neural networks
(Krizhevsky et al., 2012). The quest for more
ecient AI computaon has led to
specialized hardware like Google's Tensor
Processing Unit (TPU), which oers 92
TeraOps/second of performance through a
65,536 8-bit MAC matrix mulply unit
(Jouppi et al., 2017). And new players like
Groq or Cerebras Language Processing Units
(LPUs) are raising the TeraOps/second bar
even further in late 2024.
However, these computaonal demands are
growing at a concerning rate. Research
shows that computaonal requirements for
deep learning are scaling polynomially with
performance improvements - for example,
halving remaining error rates can require
over 5,000× more computaon (Thompson
et al., 2022). This rapid escalaon in
compung needs raises important quesons
about the economic and environmental
sustainability of current deep learning
approaches.
2.2. The Social Brain Hypothesis
The structures and networks involved in the
development of g, as studied in the P-FIT
theory, are located in the neocortex. The
development of the neocortex is parcularly
signicant in the Homo genus compared to
other species and especially in Homo
sapiens versus Homo habilis, Homo erectus,
and other non-human primates like
chimpanzees, bonobos, gorillas, and
orangutans—all of which have higher
intelligence than other animals. The most
crucial factor for the unique development of
the neocortex in Homo sapiens has been
social pressure (Humphrey, 1976). Through
socially oriented intelligence, individuals
need to know, interact with, predict,
remember, and inuence the behavior of
other group members. This intelligence,
termed Machiavellian by Richard Byrne and
Andrew Whiten, suggests that decepon
and persuasion are essenal traits in
compeve social environments among
individuals to aid their survival. Therefore,
human intelligence may not result from
improved hunng, gathering, or similar
skills but rather from the demands of social
life itself (Byrne & Whiten, 1988).
A strong empirical conrmaon of this
hypothesis comes from the so-called Dunbar
number, which expresses the relaonship
between brain size, specically the
neocortex, and the size of social groups. For
humans, Dunbar's number is around 150
individuals, corresponding to the neocortex
size. In comparison, chimpanzees form small
groups of about 30-40 individuals (Dunbar,
1993). Dunbar suggests in the social brain
hypothesis that large-scale cooperaon,
joint coordinaon, group complexity, and
relaonships require a cognive capacity
superior to that of other primates and
species.
Fig. 2 Dunbar´s number. Group number and neocortex size rao
To manage social life, humans need the
ability to aribute mental states to others
and understand that they possess thoughts,
intenons, beliefs, and emoons that may
be similar or dierent from one's own. This
ability to perceive others as acve and
independent agents and remember past
interacons is known as mentalizaon or
theory of mind (Nowak & Sigmund, 2005).
Although other primates, mainly great apes,
exhibit mentalizaon abilies, their level is
quite rudimentary. In the human brain,
mentalizaon emerges around 4-5 years of
age, coinciding with various stages of
neurodevelopment maturaon, parcularly
in the temporoparietal and medial
prefrontal corces (Premack & Woodru,
1978; Carrington & Bailey, 2009). As a result
of social selecve pressure and the need for
eecve communicaon to facilitate
cooperaon, language is undeniably a
unique tool with a disnct specializaon in
humans, possessing a syntacc and
semanc structure rooted in logic—closely
aligning with the pure concept of
intelligence as the ability to adapt to a
changing environment. Language enables
the ecient transmission of concrete and
abstract informaon, fostering stronger
bonds, overcoming challenges, and solving
complex problems (Dunbar, 1996;
Tomasello, 2014; Dunbar & Schultz, 2007).
In fact, the creaon of culture and, thus, the
long-term modicaon of intelligence
requires language (Asngton & Baird,
2005). Communicaon through language
from parents to children promotes the
development of complex mental states and
the precise direcon in building cognive
skills (Dunn & Brophy, 2020). Interesngly,
language involves the maturaon of le-
dominant frontotemporoparietal areas.
Some of these areas, such as the arcuate
fasciculus, show visible dierences from
other primates and are essenal for
language comprehension and producon
(Friederici, 2017; Fitch, 2020).
2.3. Trying to build a Digital Brain
In the case of its Silicon counterpart, one of
the key developments was using layers to
increase the capacity in neural networks, a
fundamental discovery leading to the deep
learning revoluon. By stacking mulple
layers, neural networks can learn
hierarchical representaons of data,
capturing both low-level and high-level
features. As demonstrated by Rumelhart et
al. (1986), intermediate 'hidden' units can
represent important features of the task
domain, with regularies captured through
unit interacons. This depth enables models
to tackle complex tasks in all domains like
computer vision, natural language
processing, and speech recognion. LeCun
et al. (1989) showed that adding successive
layers allows networks to detect and
combine local features into higher-order
features, similar to biological visual systems,
however, increasing the number of layers
introduces challenges such as vanishing
gradients and computaonal ineciency.
Other developments have addressed these
issues through architectural innovaons,
techniques like residual connecons or
dropout have allowed gradients to ow
more eecvely during training, enabling
the construcon of very deep networks with
less performance degradaon. Moreover, as
evidenced in early work by LeCun et al.
(1989), architectural constraints and weight
sharing can help reduce free parameters
while maintaining computaonal power,
opmizing the balance between model
depth and cost. The connuous exploraon
of deeper architectures remains a crical
area of research, pushing the boundaries of
what neural networks can achieve, in the
dream of replicang a real human brain.
Determining the opmal way to connect
arcial neurons is fundamental, since the
connecvity paern dictates how
informaon ows and is processed within
the network, inial architectures like fully
connected layers are simple but
computaonally intensive and prone to
overng, to address these issues,
researchers have developed specialized
connecon schemes (LeCun et al., 1998).
Convoluonal neural networks (CNNs)
connect neurons in a localized manner,
leveraging spaal hierarchies in data, which
is parcularly eecve for image processing
tasks, using this approach resulted in
successful document recognion tasks by
using local recepve elds and weight
sharing to reduce the number of free
parameters (LeCun et al., 1998). Recurrent
neural networks (RNNs) introduce
connecons over me steps, making them
suitable for sequenal data like text, speech
and other me series. Long Short-Term
Memory (LSTM) networks specically
address the vanishing gradient problem
through specialized memory cells and
gang mechanisms (Hochreiter &
Schmidhuber, 1997). Graph neural networks
(GNNs) allow neurons to be connected
based on arbitrary graph structures,
enabling the processing of non-Euclidean
data. For instance, Graph Convoluonal
Networks (GCNs) have demonstrated
success in semi-supervised classicaon
tasks by eciently propagang informaon
through graph structures (Kipf & Welling,
2017). This builds upon earlier work showing
the importance of selecve informaon
processing, as demonstrated in LSTM
architectures (Hochreiter & Schmidhuber,
1997), A recent development in this lines are
the Extended Long Short-Term Memory or
xLSTM migang some of the previous
issues like speed, memory or normalizaon
(Maximilian Beck, et al 2024)
The design of models and architectures is a
foundaonal aspect and current trends
emphasize the development of lightweight
models that can operate on edge devices
with limited resources. Drawing inspiraon
from early work on ecient architectures
like LeNet (LeCun et al., 1998), techniques
like model pruning, quanzaon, and
knowledge disllaon have been employed
to reduce model size and complexity without
signicant loss in accuracy. Approaches like
GCNs further demonstrate how careful
architectural design can achieve linear
computaonal complexity while
maintaining high performance (Kipf &
Welling, 2017).
Training and inference are two crical
phases in the lifecycle of an AI model. During
training, the model learns paerns from
data by adjusng its parameters to minimize
a loss funcon through techniques like
backpropagaon (Rumelhart et al., 1986).
Inference involves using the trained model
to make predicons or decisions based on
new input data. The eciency and
eecveness of both phases are vital for real
world applicaons.
Advancements in opmizaon algorithms,
such as adapve learning rate methods like
Adam, have accelerated training
convergence by combining the benets of
AdaGrad's ability to handle sparse gradients
with RMSProp's eecveness in non-
staonary sengs (Kingma & Ba, 2014). The
introducon of momentum terms and bias
correcon in opmizaon methods has
helped prevent stagnaon during training
(Kingma & Ba, 2014). For example, Adam's
adapve moment esmaon approach
automacally adjusts learning rates for
each parameter while requiring minimal
hyperparameter tuning (Kingma & Ba,
2014).
Distributed training has become essenal
for handling large datasets and complex
models. The development of ecient
gradient-based methods that can work with
stochasc objecve funcons has enabled
beer scaling of training processes across
mulple devices (Kingma & Ba, 2014).
Furthermore, the discovery that neural
networks can learn useful internal
representaons through proper weight
adjustment techniques, as demonstrated by
Rumelhart et al. (1986), also has been key.
Their work showed how mul-layer
networks can develop internal
representaons that capture important
features of the task domain through the
back-propagaon of errors.
The focus on opmizing both training and
inference connues to be a signicant area
of research, especially with the growing
demand for real-me AI applicaons. Early
breakthroughs in understanding how neural
networks learn representaons (Rumelhart
et al., 1986) have evolved into sophiscated
opmizaon methods that address praccal
challenges in current best deep learning
systems.
2.4. The Families of AIs
The journey for the human brains to the
arcial ones encompasses a diverse array
of approaches, methodologies, and
applicaons, to help organize what we are
used to name Arcial Intelligence, we
needed various taxonomies to categorize
and understand its components. One
fundamental taxonomy divides AI into
narrow (or weak) AI and general (or strong)
AI. Narrow AI refers to systems designed to
perform specic tasks, while general AI aims
to replicate human-level intelligence across
a wide range of cognive tasks (Chollet,
2019). Another common classicaon is
based on the AI system's underlying
approach: rule-based systems, machine
learning, and deep learning (Barredo Arrieta
et al., 2020). Rule-based systems rely on
predened rules and logic to make
decisions. Machine learning algorithms, in
contrast, learn paerns from data without
explicit programming (Nilsson, 1983).
AI can also be categorized by its primary
funcon or applicaon domain. This
includes categories such as natural
language processing, computer vision,
robocs, expert systems, and planning and
decision-making systems (McCarthy et al.,
1955). Each of these domains has its own set
of techniques, challenges, and benchmarks.
From a philosophical perspecve, AI
taxonomies oen consider the system's
cognive capabilies, reecng increasing
levels of sophiscaon and autonomy in AI
systems (Chollet, 2019). Ethical taxonomies
for AI have also emerged, focusing on
aspects such as transparency, fairness,
accountability, and privacy (Barredo Arrieta
et al., 2020). These classicaons help in
assessing the societal impact and
responsible development of AI technologies.
As highlighted by Nilsson (1983), the
maturaon of AI as a scienc eld requires
clear taxonomies to understand "what sets
us apart from adjacent disciplines" and to
establish AI's unique niche within the
broader landscape of intelligent systems.
We also have the dichotomy between
symbolic AI and conneconist AI represents
two fundamental approaches to arcial
intelligence.
Symbolic AI, also known as classical AI or
GOFAI (Good Old-Fashioned AI), is based on
the manipulaon of symbolic
representaons of knowledge. According to
Newell and Simon (1976), physical symbol
systems provide "the necessary and
sucient means for general intelligent
acon," where intelligence emerges from
the manipulaon of symbols and
expressions through dened processes.
Symbolic AI systems use formal logic,
decision trees, and expert systems to
process informaon and make decisions.
They excel in domains where knowledge can
be explicitly encoded. The strength of
symbolic AI lies in its interpretability and
ability to handle complex reasoning tasks.
However, it struggles with tasks requiring
paern recognion or handling uncertainty.
Conneconist AI, on the other hand, is
inspired by the structure and funcon of
biological neural networks. As described by
LeCun, Bengio, and Hinton (2015), deep
learning methods allow computaonal
models to learn representaons of data with
mulple levels of abstracon, discovering
intricate paerns in large datasets. These
systems excel in tasks such as image and
speech recognion, where paerns are
complex and dicult to specify explicitly.
They are parcularly adept at handling noisy
or incomplete data. However, their decision-
making process can be opaque, leading to
challenges in interpretability and
explainability.
The debate between symbolic and
conneconist approaches has evolved over
me. While early AI research was
dominated by symbolic methods, the
resurgence of neural networks led to
signicant advancements in conneconist
AI, with deep learning achieving
breakthrough results in areas like speech
recognion and visual object recognion
(LeCun et al., 2015). Today, many
researchers recognize the complementary
nature of these approaches and seek to
combine them in hybrid systems.
Neuro-symbolic AI, as discussed by Garcez et
al. (2015), aims to integrate the strengths of
both paradigms. The goal of neural-
symbolic computaon is to integrate robust
conneconist learning with sound symbolic
reasoning, combining the paern
recognion capabilies of neural networks
with the logical reasoning of symbolic AI.
This integraon addresses one of the main
challenges of arcial intelligence: the
eecve combinaon of learning and
reasoning (Garcez et al., 2015). This hybrid
approach holds promise for developing
more robust and versale AI systems
capable of both learning from data and
reasoning with explicit knowledge.
2.5. Winters and Summers in the Quest for
AI
As our understanding of the brain and the
essence of human intelligence evolved it
needed several waves of theories,
experiments and validaons, in a centuries
long aempt in trying to understand it, the
history of arcial intelligence has been
characterized by alternang periods of high
expectaons and enthusiasm (summers)
followed by disappointment and reduced
funding (winters). This cyclical paern has
signicantly inuenced the development
and percepon of AI technology (Floridi,
2020).
The rst AI summer began in the 1950s with
the birth of AI as a eld. As described by
Crevier (1993), pioneering work by
researchers at places like MIT's Arcial
Intelligence Laboratory led to opmisc
predicons about AI's potenal, with early
demonstraons including computers
controlling robot arms and manipulang
block structures. However, these early AI
experiments, while impressive to watch,
proved limited to carefully simplied
problems in restricted areas, leading to the
rst AI winter when military funding was
reduced (Crevier, 1993).
The 1980s brought a resurgence of interest
in AI, driven by the commercial success of
expert systems. Crevier (1993) notes that
expert systems were promoted as
specialized AI that could capture human
decision-making processes for narrowly
focused tasks, from medical diagnosis to oil
exploraon site selecon. However, as
Hendler (2008) explains, when the expert
systems market failed, it rekindled interest in
alternave approaches like arcial neural
networks.
The current AI landscape looks like it could
follow this paern. Floridi (2020) warns of
another predictable winter approaching,
arguing that AI has been subject to these
hype cycles because it represents a long-
held hope of creang something that does
everything for us. He cricizes
commentators and "experts" who competed
to tell the "tallest tale," spreading myths
about AI as either an ulmate panacea or
nal catastrophe (Floridi, 2020). Hendler
(2008) suggests that avoiding future AI
winters requires documenng successes,
embracing applied AI rather than disowning
it, and pulling together as a eld while
acknowledging both achievements and
remaining challenges.
The current AI summer, which began in the
early 2010s, between 2012 and 2013,
and has been fueled by advancements in
machine learning, parcularly deep
learning, breakthroughs in areas such as
image recognion, natural language
processing, and AI systems winning at
games like Go, have reignited excitement
about AI's potenal. This period has seen
unprecedented investment in AI research
and applicaons across all industries, and
looks like the predicted new winter by Floridi
is not on the horizon for the moment.
3. The Predicve Brain
Human intelligence, by denion, must be
exible to adapt to circumstances, swi to
reliably manage soluons, and predicve in
ancipang needs. The concept of the brain
as a smulus-response mechanism became
popular thanks to Pavlov’s studies on
reexes and the rise of behaviorism in the
early 20th century, led by Watson and
Skinner. For behaviorism, the brain and
human behavior were reacve, responding
to specic smuli. Behavior could be
modied through incenves, punishments,
and rewards. Although reexes and operant
responses are relevant to understanding
behavior, a more advanced view gained
prominence over me. The brain acvely
creates models and hypotheses about the
world without waing for smuli, predicng
what it will encounter and using sensory
feedback to adjust these hypotheses and
predicons. Percepon is not a passive
boom-up construcon but an acve,
predicve model that operates according to
a Bayesian model of what is most likely to
occur based on prior encoding in memory
and the current environment.
Karl Friston is one of the pioneers of this
perspecve. He introduced the free energy
principle, which posits that the brain
ancipates sensory input through a model
of the world, adjusng the discrepancy
between expectaon and percepon, thus
minimizing free energy, which represents
the uncertainty inherent to any biological
system at a given moment. The greater the
free energy, the larger the discrepancy
between predicon and actual sensory
input. Predicon and predicon error are
expressed in opposite direcons, top-down
and boom-up, respecvely. The system is
thus hierarchical but bidireconal, with
percepon and acon constantly feeding
each other (Friston & Stephan, 2007).
If the intense socializaon of our species was
fundamental for the development of
complex intelligence, this predicve nature
must also apply to social contexts. Theory of
Mind, in which we form ideas about others'
intenons and thoughts, is an acve
inference, ancipang others’ behaviors
based on past experiences, the environment,
and context. If others behave as expected,
predicon error does not arise. Interesngly,
in ausm spectrum disorder, neural
networks associated with the social brain
are altered, emphasizing that our social
cognion is also predicve (Frith & Frith,
1999). Adjustments in our social inferences
about others inuence self-percepon and
future adapve capacity (Frith, 2007).
On an emoonal level, the same principle
applies. Although the reacve view of
emoons has been dominant, primarily due
to proponents of basic emoons theory
(Ekman, 1999), predicon error is a central
mechanism in construcng emoons. When
somac signals or bodily assumpons do
not match predicons, the brain adjusts its
predicon (Barre, 2016). Mental
condions such as anxiety or depression are
also linked to interocepve predicons,
generang exaggerated or maladapve
simulaons about imminent dangers, bodily
alert states, or available resources to cope
with situaons (Seth & Friston, 2016).
Interocepon operates like another sense,
sending informaon to the brain about
internal physiological processes, such as
respiraon, temperature, pH, heart rate,
hunger, thirst, and available energy (Barre
& Simmons, 2015). The insula and anterior
cingulate cortex are the primary
interocepve brain regions, connecng the
body and mind (Craig, 2009). Interocepve
signals ascend via the vagus and spinal
nerves to the brainstem and then to the
insular cortex, where we become aware of
bodily states. Based on this state, we make
decisions and emoonal responses,
involving the anterior cingulate cortex
(Critchley et al., 2004). Hypo- or
hyperacvaon of the insula leads to
interocepve disorders of the bodily “self”
(Khalsa et al., 2018; Avery et al., 2014). The
sensaon of a connuous self is also linked
to interocepon (Northo & Panksepp,
2008).
Fig 3. Predicve neurons and error detecon neurons
From the predicve view, what we call
reality is merely an eecve simulaon
constructed by the brain, based on past
experiences and available sensory
informaon. When a simulated reality is
eecve depends on the level of consensus
it reaches and the tasks it enables. In mental
disorders, such as certain psychoses and
schizophrenia, predicons generate
delusions or hallucinaons that fail to
achieve consensus with others or to become
eecve predicons (Hohwy, 2013; Fletcher
& Frith, 2009). Andy Clark extends the
concept of predicon to include acon itself,
proposing that the brain predicts what it
perceives as well as the consequences of the
acons it executes bidireconally (Clark,
2015). Interocepon again plays a crucial
role in predicon. Cognion is thus
embodied, as the brain dynamically
interacts with the body and environment
(Clark, 2016; Barre et al., 2007). Language
does not escape this view; in
communicaon, we ancipate what the
other will convey, inferring or acvely
modifying meanings as we listen, from a
word to a narrave (Lindquist et al., 2015;
Lindquist & Gendron, 2013). Consciousness
would be an outcome of acve predicon,
where the brain displays the best opon
regarding bodily state and sensory input
(Seth, 2014). An exaggerated or impossible
mismatch between predicon and
predicon error causes unusual experiences,
such as dissociaon, out-of-body
experiences, or similar phenomena (Seth,
2021). Acve consciousness allows us to
assess learning processes and improve
behavior over me, favoring reecve
decision-making (Fleming & Frith, 2014;
Dehaene, 2014). Various authors extend the
predicve sense proposed by Friston and
Clark to include the sense of me and self as
a temporal agent. George Northo
proposes that the “self” is a core construct
generated from predicve processes.
Predicon depends on the temporal
synchrony of dierent brain regions through
neural oscillaons (Northo, 2014; Buzsáki,
2006). This internal rhythm of the brain is
independent and precedes any smulus. In a
state of apparent rest, an internal default
mode operates, with the brain in constant
preparaon through neural networks that
synchronize and ancipate possible
responses.
3.1. Encoding and Decoding Intelligence
If we try to nd a correlaon from biology in
arcial systems, encoder-decoder
architectures play a pivotal role in tasks that
involve transforming input data into
dierent formats or representaons. The
encoder processes the input data and
compresses it into a latent representaon,
capturing essenal features, while the
decoder reconstructs or translates this
representaon into the desired output
format (Hinton & Salakhutdinov, 2006). As
demonstrated in the groundbreaking work
by Cho et al. (2014), this architecture is
fundamental in applicaons like machine
translaon, where "one RNN encodes a
sequence of symbols into a xed-length
vector representaon, and the other
decodes the representaon into another
sequence of symbols."
The architecture's eecveness lies in its
ability to jointly train both components to
maximize the condional probability of the
target sequence given the source sequence
(Cho et al., 2014). In dimensionality
reducon applicaons, autoencoders can
learn low-dimensional codes that
signicantly outperform tradional
methods like principal components analysis,
parcularly when inialized through careful
pre-training procedures (Hinton &
Salakhutdinov, 2006). Aenon
mechanisms have enhanced encoder-
decoder models by allowing the decoder to
focus on specic parts of the input during
output generaon. This has signicantly
improved performance in various tasks,
building upon the foundaon laid by early
RNN Encoder-Decoder models where "both
yt and h(t) are condioned on yt-1 and on the
summary c of the input sequence" (Cho et
al., 2014). The exibility and eecveness of
encoder-decoder architectures make them a
cornerstone in the development of advanced
AI systems, capable of learning meaningful
representaons across diverse data types
and tasks (Hinton & Salakhutdinov, 2006).
Variaonal Autoencoders (VAEs) are a class
of generave models that extend tradional
autoencoders by incorporang probabilisc
elements into the encoding process (Kingma
& Welling, 2013). VAEs encode input data
into a latent space characterized by a
probability distribuon, typically a
Gaussian, where the true posterior
distribuon is approximated using
variaonal inference techniques (Rezende et
al., 2014). This approach allows for the
generaon of new data samples by
sampling from the latent space and
decoding the samples back into the data
space. The training objecve of VAEs
includes both the reconstrucon loss and a
regularizaon term in the form of a KL
divergence that encourages the latent
distribuon to match a prior distribuon
(Kingma & Welling, 2013). This balance
enables VAEs to generate diverse and
coherent outputs, making them valuable in
tasks like image synthesis, anomaly
detecon, and data augmentaon. Follow-
up research has shown that carefully
designed neural architectures for VAEs
achieve state-of-the-art results in image
generaon tasks (Vahdat & Kautz, 2020).
VAEs have been combined with other
architectures, such as convoluonal layers
for image data and recurrent layers for
sequenal data, to enhance their generave
capabilies. The development of VAEs
represents a signicant step forward in
unsupervised learning and generave
modeling, oering a principled approach to
learning both the generave model p(x|z)
and recognion model q(z|x) jointly through
the reparameterizaon trick (Rezende et al.,
2014).
The concepts of adding noise and denoising
it later are integral to improving the
robustness and generalizaon of AI models.
Introducing noise during training acts as a
regularizaon technique. For instance,
dropout randomly "drops" units along with
their connecons during training to prevent
units from co-adapng too much, thereby
reducing overng (Srivastava et al., 2014).
Vincent et al. (2008) demonstrated that
corrupng inputs and training models to
reconstruct the original data encourages
learning more robust features, as evidenced
by their work on denoising autoencoders
which showed improved classicaon
performance compared to tradional
autoencoders.
Denoising techniques are employed to
recover original data from corrupted or
noisy inputs. Vincent et al. (2008) explain
that denoising autoencoders learn to map
corrupted examples back to uncorrupted
ones, eecvely capturing stable structures
and dependencies in the input distribuon,
this is important as it forces the model to
learn useful features rather than simply
copying the input. The approach can be
understood from a manifold learning
perspecve, where the model learns to
project corrupted examples back onto the
manifold of natural examples (Vincent et al.,
2008).
Follow up advancements include denoising
diusion probabilisc models, which
generate high-quality data samples by
iteravely rening noise-added inputs. Ho et
al. (2020) demonstrated that these models
can achieve state-of-the-art FID scores on
image generaon tasks by learning to
reverse a diusion process that gradually
adds noise to the data, this frames the
generaon process as learning a sequence
of denoising steps, yielding a new class of
generave models that naturally admit a
progressive lossy decompression scheme
(Ho et al., 2020).
Transformers have revoluonized natural
language processing by introducing self-
aenon mechanisms that capture global
dependencies in data. Unlike tradional
sequenal models, Transformers process
input data in parallel, signicantly
improving training eciency and
performance (Vaswani et al., 2017). This
architecture has been the foundaon for
large language models (LLMs) from BERT,
which introduced bidireconal pre-training
(Devlin et al., 2018), follow-up models like
GPT-3 that demonstrate strong few-shot
learning capabilies (Brown et al., 2020),
and current models (2024) sll share the
essence of transformers like GPT-o3, Claude
3.5, Gemini 2 or LlamA 3.3
LLMs have demonstrated remarkable
capabilies in understanding and
generang human-like text, performing
tasks such as translaon, summarizaon,
and queson-answering with high
prociency. These models are pre-trained on
vast datasets and can be ne-tuned for
specic applicaons, achieving state-of-the-
art results across various benchmarks. For
example, BERT achieved signicant
improvements on eleven NLP tasks (Devlin
et al., 2018), while models like LLaMA have
shown compeve performance with much
greater eciency (Touvron et al., 2023).
The scaling of Transformers and LLMs has
raised important consideraons regarding
computaonal resources and ethical
implicaons. Brown et al. (2020) note the
substanal computaonal requirements for
training large models, while Touvron et al.
(2023) demonstrate eorts to develop more
ecient models. Addionally, research into
migang biases and ensuring responsible
use is ongoing, as highlighted by both Brown
et al. (2020) and Touvron et al. (2023),
emphasizing the importance of aligning AI
advancements with societal values.
3.2. Where Do Predicons Occur in the
Brain?
Within the bidireconal process, predicon
operates top-down, while sensory
informaon and predicon error ow
boom-up. Higher corcal areas are thus
candidates for generang and adjusng
predicons about our body, environment,
and acons. Among these, regions related
to the social brain, which dier the most
from those in other primates, are prime
candidates for being the main players in our
acve predicons and inferences. The
dorsolateral prefrontal cortex evaluates
long-term consequences of future acons
and adapts behavior when a predicon
error arises (Miller & Cohen, 2001; Duncan,
2013). The anterior cingulate cortex
monitors error, detects conicts, and adjusts
predicons between expectaons and
reality (Botvinick et al., 2004), parcipang
in learning and decision-making (Shenhav et
al., 2013). Even in primary sensory or motor
areas, low-level predicons are made about
sensory smuli and motor acons. In the
primary visual cortex, aributes such as
color, orientaon, shape, contrast, and
movement are inferred (Rao & Ballard,
1999). In the primary motor cortex,
ancipated outcomes of motor acons are
planned, allowing real-me movement
adjustments (Keller & Mrsic-Flogel, 2018).
Although specic regions are involved,
operaonal funconing is hierarchical and
networked. The default mode network and
fronto-parietal execuve network are the
most prominent high-level predicve
networks involving some of the structures
menoned. Through the default network,
self-reecon, memory, and future planning
occur, primarily via dynamic communicaon
between the medial prefrontal cortex,
posterior cingulate cortex, and
hippocampus (Northo & Huang, 2017;
Raichle, 2015). The fronto-parietal or
execuve network, linked with intelligence,
includes the dorsolateral cortex, inferior
parietal lobule, and anterior insula, and is
involved in decision-making and cognive
control (Cole, 2014). While the default
network simulates future scenarios, the
execuve adjusts predicons in response to
environmental changes. Slower frequencies
in these networks relate to large-scale
informaon integraon and future state
predicon, while faster ones in primary
corces relate to immediate percepon,
novel smuli adaptaon, and predicon
error correcon (Palva & Palva, 2018;
Friston, 2010; Northo & Huang, 2017).
Local, rapid predicon in primary corces
occurs through corcal columns, the basic
funconal units of the cortex. The corcal
column is distributed across six layers that
implement local predicve coding. When
these columns detect a predicon error, they
send signals upwards for the default or
execuve network to adjust global
predicons. This informaon exchange is
temporally synchronized (Bastos et al.,
2012). For example, if an object approaches
us, layers II and III of the corcal column
ancipate its trajectory; layer IV receives
sensory informaon from the thalamus, and
if there’s a match, predicon error does not
occur. If the object changes direcon, layers
V and VI send predicon error informaon
to higher layers to modify the inference.
Regardless of whether the smulus is visual,
auditory, emoonal, cognive, social, or
interocepve, the upper layers (I, II, and III)
generate predicons, the middle layers (IV)
compare aerent informaon, and the deep
layers (V and VI) adjust predicons (Rao &
Ballard, 1999; Feldman, 2012; Spratling,
2017).
4. Embodied, Enacted and Extended
Brains
Embodied cognion, stemming from the
dynamic interacon between the brain,
body, and environment, enables the
necessary predicons and adjustments to
operate eciently in the world (Clark, 2015;
Clark, 2016; Barret et al., 2007). Francisco
Varela is one of the pioneers of this concept,
introducing, alongside the noon of an
embodied mind, the concepts of extended
mind and "enacted mind." Together with
Humberto Maturana, he developed the
concept of autopoiesis, which posits that
living beings self-organize and adapt in
response to the characteriscs of a dynamic
environment (Maturana & Varela, 1980).
Subsequent neuroscience research
emphasizes the dynamic interplay between
percepon and acon. The brain is not a
passive enty that merely receives
environmental informaon; rather, it
acvely constructs this informaon through
motor simulaon of observed acons
(Gallese & Lako, 2005). In this regard, the
discovery of mirror neurons illustrates how
the brain automacally responds to
observed acons and movements,
impacng the percepon-acon
relaonship. The premotor cortex and
parietal cortex adapt preempvely to the
dynamic social environment (Rizzola &
Craighero, 2004).
A unique characterisc of the relaonship
with the environment is related to how the
brain responds to smuli when the
individual is at rest versus in moon. When
moving, primary somatosensory and
auditory areas reduce their acvity, thereby
prevenng sensory overload. Conversely,
there is an increase in acvity within
mulsensory regions where informaon
from various senses is integrated, allowing
for a more adapve percepon (Suzuki et
al., 2022).
Our cognion also extends into the
environment. The use of devices and
technologies directly impacts our memory,
aenon, emoonal regulaon, and sense
of self. The "Google eect" explains how,
due to easy access to informaon, people
tend to remember where informaon is
located (such as on the internet) rather than
the content itself (Sparrow, Liu, & Wegner,
2011). Regular use of GPS is associated with
reduced hippocampal acvity and a decline
in spaal memory compared to individuals
who recall landmarks or devise mental
navigaon strategies (Münzer et al., 2020;
Javadi et al., 2017). Our cognive prostheses
directly inuence aenon capacity and
long-term memory (Wilmer, Sherman, &
Chein, 2017).
In neurodevelopment, the maturaon of
cognive structures emerges as infants
manipulate objects, thus integrang visual,
somatosensory, and motor informaon,
coordinang dierent senses, and gaining
an understanding of objects’ global
properes (Needham et al., 2014).
Furthermore, acve manipulaon of objects
facilitates faster cognive categorizaon
than passive interacon (Smith & Gasser,
2005).
4.1. AGI
Arcial General Intelligence (AGI)
represents the holy grail of AI research, an
hypothecal AI system capable of
performing any intellectual task that the
average human can. Unlike narrow AI
systems designed for specic tasks, AGI
would possess human-like cognive
abilies, including reasoning, problem-
solving, learning, and adaptability across all
domains (Goertzel & Pennachin, 2007). The
concept of AGI has been a subject of intense
debate and speculaon within the AI
community. Proponents argue that AGI is
not only possible but potenally inevitable,
given the rapid advancements in machine
learning and cognive science, though they
acknowledge that creang AGI is "merely an
engineering problem, though certainly a
very dicult one" (Goertzel & Pennachin,
2007). They envision AGI systems that could
revoluonize scienc research, solve
complex global challenges, and even
augment human intelligence.
However, the path to AGI is fraught with
signicant technical, philosophical, and
ethical challenges. One major hurdle is
developing systems that can transfer
knowledge and skills across dierent
domains, a capability known as transfer
learning. As Lake et al. (2017) argue, truly
human-like AI systems must "harness
composionality and learning-to-learn to
rapidly acquire and generalize knowledge to
new tasks and situaons." Another
challenge lies in imbuing AI systems with
common sense reasoning and
understanding of context, which humans
acquire through years of experience and
interacon with the world. This requires
building "causal models of the world that
support explanaon and understanding,
rather than merely solving paern
recognion problems" (Lake et al.,
2017).The potenal implicaons of AGI are
profound and far-reaching, raising
important ethical and societal quesons.
Issues of control, alignment with human
values, and the potenal existenal risk
posed by superintelligent AI systems are
acve areas of research and discussion in
the eld of AI safety (Bostrom, 2014). As
Bostrom's analysis suggests, the creaon of
superintelligent beings could represent both
a possible existenal risk to mankind and an
opportunity to address this risk through
careful development and control measures.
According to one of the main players in the
AGI race, Open AI (Open AI 2024) there are
5 levels regarding the capabilies of
algorithms to mimic and surpass human
funcons:
Level 1: Conversaonal AI
At the foundaonal stage, Conversaonal AI
focuses on understanding and generang
human language to engage in uid,
meaningful dialogue. These systems are
designed primarily for tasks like answering
quesons, providing informaon, and
assisng users in a structured
conversaonal format. They rely heavily on
language models trained to interpret
context within a limited scope, making them
suitable for customer service, virtual
assistance, and other straighorward,
communicaon-focused roles. Although
highly capable in processing and generang
text, conversaonal AIs are primarily
reacve, responding to user prompts
without the capacity for deep reasoning or
independent thought.
Level 2: Reasoning AI
Building upon the conversaonal abilies of
Level 1, Reasoning AI adds a layer of logical
processing and contextual understanding,
allowing it to analyze problems and deduce
soluons. These AIs can draw inferences and
understand causal relaonships within a
given framework, enabling them to tackle
more complex tasks, such as diagnoscs,
crical thinking, and problem-solving across
structured domains. They are equipped to
go beyond surface-level interacons,
engaging in reasoning to suggest soluons
and make sense of complex informaon. As
a result, Reasoning AI can be employed in
areas like nancial analysis, medical
support, and legal assistance, where deeper
understanding and logical interpretaon are
essenal.
Level 3: Autonomous AI
Autonomous AI represents a signicant
leap, enabling systems to iniate and
complete tasks independently, without
direct human prompts. These AIs possess
the ability to self-direct their acons based
on situaonal requirements and pre-dened
goals, making them capable of navigang
dynamic environments autonomously. They
adapt to new condions in real-me,
making autonomous cars, drones, and
certain roboc applicaons possible.
Autonomous AI introduces systems that can
handle mul-step processes with some
degree of unpredictability, as seen in
logiscs or industrial automaon, where
they operate independently to complete
complex workows while adjusng to
environmental changes.
Level 4: Innovang AI
At the Innovang AI level, systems advance
beyond mere task compleon to generate
original ideas, hypothesize, and innovate
within specied elds. These AIs can
contribute creavely to scienc research,
engineering, and the arts by forming and
tesng novel hypotheses, designing
soluons, or producing new arsc works.
Innovang AI possesses the ability to
analyze paerns and create something new
rather than simply iterang on human
knowledge, showing potenal for
breakthroughs in areas like drug discovery,
material science, and technology
development. This level of AI could reshape
industries by independently pushing the
boundaries of what’s possible, discovering
new pathways that humans may not have
envisioned.
Level 5: Organizaonal AI
Organizaonal AI represents the pinnacle of
integrated AI, capable of managing and
opmizing complex systems, from corporate
operaons to large-scale infrastructure. This
level of AI funcons across various domains,
coordinang processes, strategizing, and
making high-level decisions autonomously.
It could oversee interconnected tasks in a
corporate seng, inuence policy planning,
or orchestrate city-wide infrastructure, with
an understanding of overarching goals and
intricate dependencies. Organizaonal AI
has the potenal to act as a system-wide
manager, opmizing resources and guiding
complex organizaons toward strategic
objecves with minimal human
intervenon, making it highly
transformave but requiring robust ethical
and alignment measures to ensure societal
benet.
Other AGI players draw dierent levels or
approaches but the essence is similar.
5. Qubic’s Aigarth
Qubic (hps://qubic.org) is a decentralized
network, created by one of the leading
gures in the eld, Sergey Ivancheglo,
involved also in several other projects in the
space over the last decade, Qubic central
aim is creang and evolving an AGI from the
beginning.
Qubic’s approach to Arcial Intelligence is
dierent than the tradional lines of
research and development in the eld giving
birth to the concept of Aigarth, “garth”
coming from the old English word for
“garden” or “yard” where the AIs will
develop and grow instead of being fully
dened by human design. One of the rst
dierences is having in mind the role of
limited GPUs in tradional AI approaches.
While many AI projects rely heavily on
powerful GPUs for training and inference, as
evidenced by the increasing deployment of
large accelerator-rich clusters providing
peta- or exa-scale levels of compute (Jain et
al., 2024), Aigarth takes a dierent path.
The project focuses on CPU-based training,
emphasizing eciency and accessibility over
raw computaonal power. This approach
aligns with Aigarth's goal of creang a
decentralized AI system that can run on a
wide range of hardware.
By moving away from GPU dependence,
which has dominated the industry over the
last decade with less than a dozen "GPU
rich" labs and instuons, and most of the
researchers being "GPU poor", Aigarth
opens up possibilies for broader
parcipaon in AI development. This
democrazaon is parcularly relevant
given that current ML workloads
increasingly require massive computaonal
resources, with some models needing
hundreds of thousands of accelerators (Jain
et al., 2024). It allows for a more distributed
network of contributors, leveraging the
collecve power of many standard
computers rather than relying on specialized
hardware. This strategy not only
democrazes AI development but also
potenally leads to more robust and
adaptable AI systems. Avoiding the use of
GPUs in Aigarth also drives innovaon in
algorithmic eciency. As noted by Jain et al.
(2024), the slowing of Moore's Law and end
of Dennard's Scaling has pushed large-scale
systems toward heterogeneous accelerators
to scale performance, especially for ML
workloads. However, Aigarth takes an
alternave approach, to nd novel soluons
that can perform complex computaons
with limited resources. This constraint-
driven innovaon could lead to
breakthroughs in AI eciency that might be
overlooked in resource-rich environments.
In terms of leveraging massive computaon
there are 2 approaches: one leveraging
decentralized distributed compung, like for
example what SETI@home did to harness
collecve computaonal power, or another
approach by using tradional centralized
(cloud) AI pracces that uses big data
centers with increasing power demands to
deal with large amounts of data and
computaonal resources (Wingarz et al.,
2024).
Aigarth uses decentralized distributed
compung, for several reasons, like the need
for greater control over the development
process and to ensure the privacy and
security of the evolving AI systems.
decentralized compung also allowed for
more precise tuning of the Intelligent Tissue
and AI components discussed later. This shi
reects broader industry recognion that
edge compung can help migate privacy
concerns and reduce latency while
enhancing bandwidth ulizaon (Lee et al.,
2024). Aigarth also uses a distributed model,
albeit in a more sophiscated form. The
vision is to have Aigarth-created AIs run as
Qubic smart contracts on a decentralized
network. This approach aligns with
emerging frameworks like the Decentralized
Intelligence Network (DIN), which enables
scalable AI development while preserving
data sovereignty and individual rights
(Nash, 2024). This hybrid approach
combines the benets of local control with
the power of distributed compung,
potenally incorporang human-in-the-
loop mechanisms to ensure responsible AI
development (Dehouche & Blythman, 2023).
A crical aspect in comparing AI
architectures is their eciency across key
performance metrics, we can analyze three
dominant approaches: transformer-based
models, tradional ANNs, and neural-
symbolic systems. Geva et al. (2022)
demonstrated that transformer
architectures exhibit signicant memory
overhead in their feed-forward layers,
showing that these layers consume up to
33% of the model's parameters, resulng in
memory footprints approximately 65%
larger than tradional approaches, though
achieving 15% faster inference speeds
through parallel processing. Homann et al.
(2022) provided detailed analysis of
computaonal requirements for
transformer models, revealing that
compute-opmal training requires scaling
laws following a power law between model
size and training compute, with opmal
model sizes scaling as V^(0.74) with the
training compute budget, resulng in
approximately 33% higher computaonal
demands compared to tradional
architectures. In contrast, Garcez et al.
(2019) showed that neural-symbolic
integraon can provide ecient knowledge
representaon and reasoning capabilies
while maintaining lower computaonal
requirements, demonstrang a 25%
reducon in compute needs and 40%
improvement in memory eciency
compared to pure deep learning
approaches, though with a 10% reducon in
inference speed due to reasoning overhead.
Given the early stage of Aigarth's
development, it is premature to make direct
performance comparisons with these
established architectures, Figure 4 presents
empirically veried comparisons between
these three architectural approaches based
on peer-reviewed benchmarks, focusing on
compute requirements, memory usage, and
inference speed, with all metrics normalized
to tradional ANN performance (baseline =
100). The data shows clear trade-os
between computaonal eciency, memory
ulizaon, and processing speed across
dierent architectures. Future work will be
needed to establish rigorous comparisons as
Aigarth matures, parcularly in validang
its projected performance characteriscs
against these established benchmarks,
followup revisions of the paper will include
these comparable benchmarks, though
preliminary projecons can be considered,
since it be a evoluon of Neuro-symbolic
with a reducon in computaonal
requirements.
Fig. 4. Comparave analysis of computaonal eciency metrics across major AI architectural
approaches, Aigarth, are esmated as an evoluon of the Neural-Symbolic approach.
Aigarth's focuses on building an adapve
approach to AI development, seeking the
most eecve method for each stage of
evoluon, while addressing crical
challenges in security, privacy, and
scalability that are inherent in distributed AI
systems (Wingarz et al., 2024).
The main 3 pillars of Aigarth (in contrast to
other approaches) are:
1.- A discrete mathemacal formulaon
called “Intelligent Tissue” to encapsulate
proven building blocks of neural
computaon.
2.- An unprecedented computaonal power,
with millions of cpu cores focused on a
“Ternary Compung” approach, including a
novel use of the third state.
3.- Nature inspired to enable the growth of
“Evoluonary Dynamics” exploring in a
systemac way the “unknown” “unknowns”
sll ooding and liming the AGI contenders
looking for potenal soluons to build
smarter AIs.
5.1. Intelligent Tissue to build
“Intelligence”
Aigarth's approach to replicate intelligence
instead of being programmed to solve
specic problems, aims to create AIs that
can autonomously develop problem-solving
capabilies using dierent paths, the recent
large language models have demonstrated
already capabilies in percepon,
reasoning, decision-making, and very
limited self-evoluon (Gao et al., 2023). In
Aigarth this is achieved through a process
that mirrors biological evoluon, drawing
inspiraon from natural selecon principles
(Darwin & Wallace, 1858).
The “Intelligent Tissue”, is a set of
interconnected neurons, discovered in the
rst year of the project, represenng a
complex network of arcial neurons and
synapses, operang under principles similar
to those observed in biological neural
networks, which exhibit both structural and
funconal modularity (Jacobides et al.,
2021), which is then shaped and rened
through countless iteraons. AI modules are
created from this ssue, each with the
ability to modify its own structure; those
that successfully solve problems survive and
evolve, while those that fail are discarded.
This "survival of the est" approach
ensures that only the most capable AIs
progress, analogous to how Selecon-
Inference frameworks allow for the
evoluon of logical reasoning capabilies
through iterave renement (Creswell et al.,
2022),.the evoluonary dynamics of this
ssue are key to understanding Aigarth's
potenal for creang truly adapve and
intelligent systems.
Intelligent Tissue evolves through a process
that mirrors biological evoluon, where
resource constraints and environmental
pressures shape network organizaon
(Béna & Goodman, 2024). The ssue starts
with a basic structure and undergoes
connuous modicaons. These
modicaons occur at the level of individual
neurons and synapses, with changes in
connecons and signal delay parameters
driving the evoluon of the ssue, similar to
how Ornstein-Uhlenbeck processes can
guide parameter adaptaon in neural
networks (Garcia Fernandez et al., 2024).
What makes this process unique is its self-
directed nature. The ssue evolves not
based on predetermined rules or connuous
human intervenon, but through a process
of trial and error guided by the eecveness
of the resulng structures in solving
problems. This mirrors how biological
networks opmize for both metabolic
eciency and informaon processing
capabilies (Béna & Goodman, 2024).
Successful modicaons are retained and
built upon, while ineecve ones are
discarded. This evoluonary approach
allows for the emergence of complex,
intelligent behaviors from relavely simple
components. The system's organizaon
emerges from the interplay between
resource constraints and task demands
(Garcia Fernandez et al., 2024),
represenng a boom-up approach to AI
development that has the potenal to
create systems with capabilies far beyond
what could be explicitly programmed. This
approach aligns with current understanding
of how both biological and arcial neural
networks develop specialized funcons
through dynamic adaptaon processes
(Jacobides et al., 2021).
Aigarth's main focus is on general problem-
solving abilies rather than narrow, task-
specic skills. The goal is to create AIs that
can adapt to new, unforeseen situaons and
generate novel soluons, beyond how
modern LLM agents can exhibit adapvity
and heterogeneity across dierent domains
(Gao et al., 2023). This approach could lead
to AIs capable of tackling complex, real-
world problems that current AI systems
struggle with. Aigarth's problem-solving
approach is designed to be transparent and
as explainable as possible, addressing one of
the key challenges in current AI
development, much like how Selecon-
Inference frameworks provide interpretable
traces of reasoning steps (Creswell et al.,
2022).
5.2. From Bytes to Bits - Ternary Compung
Aigarth's approach to compung marks also
a signicant shi from tradional “binary”
systems to a ternary paradigm. This move
from bytes back to bits, or more accurately,
to trits, a fundamental trait to the project's
innovave approach to AI development,
building on recent advances in ternary
compung systems (Chen & Lu, 2021).
In Aigarth's ternary system, each unit of
informaon can have three states: TRUE,
FALSE, or UNKNOWN. This ternary logic
allows for a more nuanced representaon of
informaon, similar to how {0, ±1}-ternary
codes have been shown to outperform
tradional binary codes in deep learning
applicaons (Chen & Lu, 2021). The
UNKNOWN state is parcularly crucial, as it
can represent various condions such as
input noise, unnished tasks, or genuine
uncertainty, aligning with Zakrisson's (2024)
observaon that ternary approaches can
eecvely handle missing or uncertain data
without making assumpons about the
missing informaon. This trenary approach
is not just a theorecal concept but is deeply
integrated into Aigarth's architecture. Each
arcial neuron in the Intelligent Tissue
operates on this ternary principle, allowing
for more complex and nuanced informaon
processing. This approach was a precursor
to developments in ecient AI systems, such
as the BitNet framework, which has
demonstrated signicant improvements in
both performance and energy eciency
through ternary representaons (Wang et
al., 2024).
The shi to ternary compung also oers
potenal advantages in terms of energy
eciency and computaonal density.
Recent research has shown that ternary
systems can achieve substanal energy
savings, with reducons of up to 70% in
energy consumpon compared to
tradional approaches (Wang et al., 2024),
aligning with Aigarth's goal of creang
more sustainable and scalable AI systems.
5.3. Evoluonary Dynamics
One of the most interesng aspects of the
arcial neural systems is their potenal for
self-modicaon capabilies. Research has
demonstrated how even simple
computaonal substrates can give rise to
self-modifying programs without explicit
programming (Agüera y Arcas et al., 2024).
This emergent behavior has been observed
across various programming languages and
environments, where programs can modify
both themselves and their neighbors based
on their own instrucons. The process of
self-modicaon in neural networks
typically involves two key components:
synapc plascity and neuromodulaon
(Schmidgall et al., 2023). Plascity in the
brain refers to the capacity of experience to
modify the funcon of neural circuits. This
plascity can operate on dierent
mescales, from short-term adaptaons
lasng milliseconds to minutes, to long-term
changes that persist for extended periods.
What's parcularly innovave about self-
modicaon in arcial systems is that it
can arise spontaneously through
interacons and self-modicaon, rather
than requiring explicit tness funcons or
predetermined goals (Agüera y Arcas et al.,
2024). Self-replicators tend to arise in
computaonal environments lacking any
explicit tness landscape, suggesng that
self-modicaon capabilies may be an
emergent property of certain types of
computaonal systems. The implicaons of
self-modicaon extend to the
interpretability and safety of AI models.
Work on large language models has focused
on understanding how dierent components
of neural networks encode and modify
informaon (Templeton et al., 2024). This
research suggests that as models scale up,
maintaining interpretability of self-
modicaon processes becomes
increasingly crucial for ensuring safe and
predictable behavior.
The self-modicaon capabilies observed
point toward the possibility of creang more
adapve and autonomous AIs, however
there remain fundamental dierences
between ANNs' operang mechanisms and
those of the biological brain, parcularly
concerning learning processes (Schmidgall
et al.,2023). Understanding and bridging
these dierences remains a key challenge.
Exisng evoluonary neural architecture
search (NAS) methods, such as NEAT
(NeuroEvoluon of Augmenng Topologies)
either with CPUs or current GPU approaches
(Lishuang Wang et al.,2024), also rene
network topologies through iterave
selecon and mutaon. However, unlike
NEAT, which typically evolves purely binary
or real-valued connecon weights, Aigarth
employs a ternary compung paradigm that
includes an explicit “UNKNOWN” state for
represenng uncertainty. This tri-state logic,
coupled with Aigarth’s decentralized and
cryptographically seeded CPU-based
approach, diverges from tradional GPU-
centric pipelines by enabling asynchronous
neuron updates, reproducible randomness,
and self-modifying connecons on
commodity hardware. Consequently,
while NEAT focuses on adapve topology
growth with CPU eciency, Aigarth’s
algorithmic design hinges on a more ne-
grained evoluonary process, in which
paral or incomplete informaon can be
handled gracefully without centralized
supervision or convenonal binary
constraints.
Aigarth’s evoluonary process implements a
clear, ered framework of mutaon,
crossover, and adapve selecon thresholds
to exploit the ternary logic fully. As
demonstrated by the “ComputeScore”
roune in Appendix B, each candidate
network is rigorously evaluated for its
output-to-input reconstrucon accuracy, like
in an autoencoder, creang a transparent
selecon environment. In each generaon:
Mutaon: Every neuron in the “Intelligent
Tissue” is subject to a xed mutaon
probability
𝑝𝑚
when triggered, the
neuron’s ternary state (TRUE, FALSE, or
UNKNOWN) undergoes a controlled change,
ensuring a measured injecon of
randomness. This approach parallels
canonical genec algorithms (Goldberg,
1989), with the disncon that Aigarth
rotates among three possible neuron states
rather than two, expanding the evoluonary
search space.
Crossover: High-scoring networks are paired
and recombined, merging subsets of
neurons and synapc connecons to form
ospring architectures. This recombinaon
process, inspired by neuroevoluonary
techniques such as NEAT (Stanley &
Miikkulainen, 2002), leverages
cryptographically seeded randomness so
that both the resulng topology and the
distribuon of ternary states remain
reproducible and fair.
Adapve Selecon Threshold: Epoch-specic
“soluon thresholds” dynamically adjust the
fracon of networks allowed to survive and
replicate. Candidates surpassing the
threshold (or those exhibing meaningful
improvements in reconstrucon accuracy)
progress unimpeded, while lower-scoring
ones face deacvaon or further
modicaon. This ensures the perpetual
renement of top-er soluons, aligning
with Aigarth’s broader vision of evolving
general problem-solving capabilies.
By unifying these operators in a ternary logic
framework, Aigarth maintains a robust
balance between exploraon and
exploitaon. Mutaon prevents premature
convergence by regularly introducing novel
states; crossover inherits and recombines
successful substructures; and adapve
thresholds systemacally winnow
subopmal conguraons while preserving
genuinely innovave traits. This synergy of
evoluonary operators, disncvely
adapted to Aigarth’s ternary paradigm,
drives the self-modicaon and survival-of-
the-est process at the heart of the
system.
The current Aigarth architecture is not fully
implemented to review systemacally all the
potenal evoluonary steps that will take
place, and this is set to be explored in
followup revisions of the paper once all
these processes are measurable and
comparable.
5.4. Towards Self-Awareness
The concept of self-awareness in AI is a
complex and evolving eld of study that
seeks to understand how arcial enes
can possess a form of self-recognion and
introspecon (John Achterberg et al., 2024).
This exploraon is crucial for advancing AI
systems that can beer interact with
humans and their environments. Self-
awareness in AI refers to "the ability of a
system to recognize its own existence and
state" and involves three key aspects:
recognion of self, introspecon, and
adaptaon (John Achterberg et al., 2024).
Rather than trying to explicitly program self-
awareness, modern approaches view it as
an emergent phenomenon, according to
Esmaeilzadeh et al. (2021), consciousness
and self-awareness in AI require "at least
two AI agents capable of communicang
within a given environment to foster the
creaon of an AI-specic language." This
suggests self-awareness may develop
naturally through interacon rather than
direct programming.
Self-awareness is evaluated through several
key indicators including: "self-recognion -
the ability to idenfy itself in a mirror test or
similar scenarios; goal seng - establishing
objecves based on internal states and
external condions; feedback mechanisms -
ulizing feedback to adjust acons and
improve performance; and emoonal
simulaon - mimicking emoonal responses
to enhance interacon with humans" (John
Achterberg et al., 2024). This framework
provides a structured approach to
understanding and measuring AI self-
awareness development. The relaonship
between self-modicaon capabilies and
self-awareness is parcularly important.
Esmaeilzadeh et al. (2021) propose that for
consciousness to emerge, "AI agents must
communicate their internal state of me-
varying symbol manipulaon through a
language that they have co-created." This
aligns with theories from cognive science
and neuroscience, parcularly the Global
Workspace Theory (GWT), which suggests
that consciousness arises from the
integraon of informaon across various
cognive processes (John Achterberg et al.,
2024).
Similarly, Spaotemporal theory of
consciousness posits that self awareness
integrates neural dynamics and subjecve
experience, emphasizing the interplay
between the brain’s resng-state acvity
and self-related processing. Self-relevance
priorize smuli linked to the self,
modulated by the default mode network
(DMN) (Northo & Bermpohl, 2004). Self-
specicity involves neural specializaon for
self-referenal tasks, so self-awareness
remains constant even at rest (Northo,
2014). Temporospaal integraon aligns
intrinsic neural acvity with external
environmental smuli, therefore enabling
the temporal and spaal coherence of self-
awareness (Northo, 2018). Self-other
disncon between the self and external
enes, anchors self-awareness in both
social and personal contexts (Northo,
2011). This model situates self-awareness as
emerging from the dynamic interplay of
intrinsic brain acvity and external
relevance.
However, developing self-aware AI poses
signicant challenges. Current AIs face
technical limitaons in achieving the depth
of understanding required for true self-
awareness, and establishing reliable
methods to assess self-awareness remains a
crical hurdle (John Achterberg et al., 2024).
As Srinivasa et al. (2022) point out, there are
"fundamental issues with the way
'intelligence' is dened and modeled in
present day AI systems," making it an open
queson whether an AI would experience
subjecve consciousness in the way humans
do, or whether its self-awareness would
manifest in fundamentally dierent ways. In
the Aigarth framework, self-awareness is
not seen as a binary state that an AI either
has or does not have. Instead, it's viewed as
a spectrum of capabilies that evolve over
me. As AIs develop more complex internal
models of their environment and their own
funconing, they may naturally develop
something akin to self-awareness. The self-
modicaon capabilies of Aigarth AIs play
a crucial role in this process. As an AI learns
to modify its own structure and behavior, it
necessarily develops a kind of self-model.
This self-model, combined with the AI's
ability to observe the results of its acons,
could lead to a form of self-awareness.
However, it's important to note that this
form of self-awareness may be quite
dierent from human self-awareness. It's an
open queson once Aigarth is completed
whether it would experience subjecve
consciousness in the way humans do, or
whether its self-awareness would be more
akin to a highly sophiscated self-
monitoring system.
5.5. True AI
The concept of "True AI" represents a
paradigm shi towards arcial general
intelligence or arcial consciousness that
can match or exceed human-level
capabilies (Li, 2018). While current AI
achievements mainly simulate intelligent
behavior on computer plaorms and belong
to "weak AI," True AI aims to develop
systems with genuine understanding,
intenonality, mind and consciousness
(Pontes-Filho & Nichele, 2020).
Aigarth's approach to achieving True AI
diers from convenonal methods as
explored earlier by focusing on evoluonary
and bio-inspired frameworks rather than
directly mimicking human intelligence. This
aligns with research showing that successful
arcial systems oen succeed when they
stop imitang biological systems and
develop their own paradigms - just as the
Wright brothers achieved ight through
aerodynamics rather than copying birds (Li,
2018). By starng with simple components
(Intelligent Tissue) and enabling evoluon
through environmental rewards and self-
learning, Aigarth aims to create AI that can
truly learn and adapt without supervision.
Key characteriscs of Aigarth's vision for the
future of True AI include:
1.- General problem-solving abilies across
diverse environments (Pontes-Filho &
Nichele, 2020)
2.- The ability to learn and self-adapt during
runme without explicit programming
3.- Self-improvement capabilies through
evoluonary processes
4.- Potenal for creavity and novel idea
generaon
5.- The ability to reason about abstract
concepts and ground symbols in real-world
experience (Li, 2018)
6.- Possible emergent properes like self-
awareness and consciousness
While achieving True AI remains an
ambious goal, Aigarth's unique approach
combining evoluonary algorithms, ternary
compung, and decentralized development
oers a promising direcon. As Lee et al.
(2024) note, decentralized AI architectures
that allow for permissionless parcipaon
and distributed processing may be crucial
for developing more robust and trustworthy
AI systems.
5.6. The Qubic Aigarth Scoring Algorithm
To ground the Aigarth theorecal
framework detailed earlier, we explored a
full year codebase and some execuon
results, shared with us by the Qubic
development team, to detail its architecture
and explore the expected evoluonary paths
and constraints.
At the heart of it, in the current phase of
development, that is the foundaon for the
discovery of candidates for the “Intelligence
Tissue” Aigarth employs a determinisc
scoring algorithm that evaluates how well
the network reconstructs input paerns at
its output layer. Input ternary paerns are
fed into the input layer, and aer a series of
asynchronous updates governed by
determinisc pseudo-random connecons,
the output layer is compared to the input.
The score is the number of matching
neurons. Thresholds, dynamically adjusted
over epochs or weeks of compute, in the
case of Qubic every week is a epoch, this
determine whether a parcular
conguraon is deemed “good” or “bad,”
driving selecon, the system ulizes three
key cryptographic components: a
mining_seed (providing epoch-specic
randomizaon), a public_key (idenfying
compung nodes), and a nonce (enabling
soluon space exploraon) to ensure both
reproducibility and computaonal fairness
(Creswell et al., 2022).
The scoring process begins with the
inializaon of a neural architecture
comprising input neurons, hidden units, and
output neurons. Network connecvity is
established through synapc connecons
per neuron, with connecon paerns
determiniscally generated using
KangarooTwelve cryptographic hashing.
Crical to the design, the mining_seed is
updated every internal interval + external
interval cks, ensuring periodic network
reconguraon while maintaining
reproducibility within epochs.
The evaluaon mechanism employs a
sophiscated mul-step opmizaon
process:
1.- Inial network state computaon using
the complete neuron set
2.- Iterave opmizaon through a number
of step phases
3.- Selecve neural acvaon skipping
based on cached state analysis
The score is determined by counng
matching input-output neuron pairs, with
values normalized to the range [0, number
of inputs]. A soluon is considered valid
when exceeding the epoch-specic
SOLUTION_THRESHOLD (Supplementary-A),
which dynamically adjusts to maintain
computaonal diculty
Performance opmizaon is achieved
through parallel processing across
NUMBER_OF_SOLUTION_PROCESSORS or
cores, with task distribuon managed via a
queue-based system. The implementaon
ulizes AVX512 or AVX2 vector instrucons
for ecient neural state updates and
scoring computaons.
To ensure reproducibility while prevenng
gaming of the system, each soluon aempt
is uniquely idened by the triple
(public_key, mining_seed, nonce), with
results cached using a ScoreCache
mechanism of SCORE_CACHE_SIZE entries.
This design allows for ecient vericaon
while maintaining the system's
cryptographic security properes.
Because all random number generaon is
seeded cryptographically, experiments are
reproducible while maintaining a search
space too large to guess outcomes trivially.
Over mulple epochs (currently studied
between EPOCH 83 to EPOCH 140), the
system’s parameters, such as input size,
hidden layer conguraon, and soluon
thresholds, have been tuned adapvely.
These changes reect evoluonary
“pressure,” pushing the Intelligent Tissue
toward architectures that are more
memory-ecient, require less data, or meet
stricter soluon criteria. The observed
stepwise changes in DATA_LENGTH,
SOLUTION_THRESHOLD and neuron
conguraons suggest systemac
opmizaon under evoluonary pressure,
analogous to pruning and renement in
developing biological neural networks
(Jacobides et al., 2021).
The current implementaon of the Aigarth
scoring and inializaon procedure,
including parameter sengs and
cryptographically seeded connecon
generaon, is available on GitHub
(hps://github.com/qubic/core). The
repository documents the stepwise
conguraon changes across epochs,
enabling independent vericaon and
replicaon of experiments.
6. Preliminary Comparisons and Future
Direcons
While Aigarth is in early development,
preliminary analyses of the system’s
computaonal footprint and performance
suggest that it may achieve more resource-
ecient operaon than convenonal large-
scale architectures. The ternary approach
and sparse connecvity are hypothesized to
reduce the energy and memory costs
compared to dense ANN or transformer-
based architectures, though future
empirical work is needed to validate these
claims rigorously (Garcez et al., 2009;
Homann et al., 2022).
In upcoming evaluaons, Aigarth’s
scalability and adaptability will be tested on
benchmark tasks, comparing performance
with classical ANN and transformer models
under equal CPU-only resource constraints.
Addionally, evolving the Intelligent Tissue
across diverse tasks may demonstrate
domain-agnosc problem-solving abilies,
moving closer to an AGI scenario (Goertzel &
Pennachin, 2007; Gao et al., 2023).
By adopng an evoluonary, decentralized,
and ternary framework, Aigarth sets the
stage for more accessible and potenally
more adapve AI systems. This ongoing
work will require systemac benchmarking,
rigorous hyperparameter studies, and
transparent reporng of performance
metrics to substanate Aigarth’s claim as a
viable path toward general and potenally
self-aware AI.
To track evoluonary lineage and
improvements over me (Nash, 2024),
future releases will incorporate extended
metrics (beyond reconstrucon scores),
once the neural architectures extend their
capabilies, and to keep exploring the
evoluon stages of Aigarth over the
following epochs, translated to months over
2025 and beyond, tracking the possible
trajectories towards AGI, something that is
sll speculave across all scienc literature
since it has not be achieved yet nor
documented at the me of the wring of this
paper, so the work will be connued
following closely next Qubic’s AI steps.
7. Brain-Inspired Systems: Towards Safe
and Ethical AGI
AI systems must ensure they are benecial
to humanity. Current risks include
systemac AI biases, AI hallucinaons,
privacy concerns, fake content generaon,
and populaon surveillance (Peterson &
Homan, 2022; Bostrom, 2014; Chrisan,
2021).
Decentralized systems powered by
blockchain technology enhance AI security
by providing tamper-proof data integrity,
transparent decision-making processes, and
improved resilience against single points of
failure (Xu et al., 2021).
The next generaon of embodied systems,
such as humanoid robots, autonomous
vehicles, drones, virtual assistants or AI
sciensts, hold the potenal to deliver
transformave societal benets (Bengio et
al., 2021). However, concerns about their
malicious use for military applicaons and
the potenal loss of control over advanced
AI agents remain signicant challenges
(Brundage et al., 2018; Taddeo & Floridi,
2018).
To avoid or migate these risks without
compromising benets, AI systems can
emulate biological mechanisms to respond,
adapt, and predict environmental changes
and uncertaines more accurately.
Brain-based representaons enhance AI
systems' ability to adapt to human contexts
by modeling percepon and contextual,
hierarchical informaon processing using
principles like predicve coding. These
representaons demonstrate robustness to
unknown inputs and enable analysis of
ambiguous situaons and scenarios (e.g.,
atypical symptoms in medical diagnosis or
unclear trac signals in autonomous
driving) by integrang current sensory
inputs with previously stored informaon.
Digital twins play a crucial role by simulang
the consequences of potenal acons in the
environment. These simulaons allow
systems to learn and adapt to norms
(implicit or explicit), signicantly improving
their ability to align with human knowledge
even in environments where novel situaons
do not match training data or are not
explicitly ancipated in the inial system
design (Rao & Ballard, 1999; Friston, 2010).
This capability is crical for addressing
problems and reducing unexpected or risky
behaviors.
As AI systems become increasingly
integrated into crical decision-making
processes, ethical concerns must guide their
design, deployment, and regulaon. These
systems must priorize fairness,
accountability, and transparency to avoid
spreading exisng societal inequalies
(Taddeo & Floridi, 2018). Addionally,
ensuring informed consent in applicaons
with sensive personal data is paramount to
respect individual autonomy and privacy
(Peterson & Homan, 2022).
Aligning AI with human values involves
embedding ethical frameworks into their
operaonal principles. This includes the
need to understand AI decisions
interpretable by humans. Ethical dimension
should also address the unintended
consequences of AI systems (i.e. job
displacement, energy consumpon, unequal
access, environmental impact), emphasizing
the need for socially sustainable soluons
(Chrisan, 2021).
Ulmately, ethical AI development requires
a muldisciplinary approach, involving
collaboraon between technologists,
ethicists, policymakers, and the public to
balance innovaon with societal well-being.
These advancements not only reinforce the
robustness of AI systems but also enhance
their interpretability, directly addressing
challenges of specicaon and safe
monitoring in advanced autonomous
systems (Mineault et al., 2024).
8. Conclusions
Human intelligence, driven by the
hierarchical and adapve architecture of the
brain, deeply guides the development of
arcial general intelligence (AGI).
Neuroscienc models, based in predicve
coding and biological self-
organizaon, provide the key principles for
designing more robust, secure, and ethically
aligned AI systems.
Rooted in this context, Aigarth, emerges as
a pioneering approach combining
biologically inspired innovaons with
advanced technologies such as ternary
compung and decentralized CPU-based
systems. Aigarth proposes an "Intelligent
Tissue" architecture that simulates brain-
like self-organizaon and facilitates
adaptability to complex and new
environments.
Furthermore, the decentralizaon and
democrazaon of AI development
envisioned by Aigarth addresses key
challenges related to control, transparency,
and sustainability while facilitang progress
toward a more accessible and ethical AGI.
We demonstrate how integrang
neuroscienc principles with innovave
computaonal approaches can unlock new
possibilies to overcome current limitaons,
bringing us closer to a safe, aligned, and
transformave AGI for society.
Supplementary Informaon
A. Aigarth Network Architecture
The Network Architecture reviewed for this
paper is:
• Input layer: n₁ neurons
(DATA_LENGTH)
• Hidden layer: n₂ neurons
(NUMBER_OF_HIDDEN_NEURONS)
• Output layer: n₁ neurons (same as
input)
• Each neuron connects to k nearest
neighbors
(NUMBER_OF_NEIGHBOR_NEURO
NS)
• Values are -1, 0 or 1
• Uses sparse connecvity with
neighbor-based topology
The key characteriscs that make this
algorithm unique are:
• Asynchronous updates - only one
neuron updated per mestep
• Sparse connecvity based on
neighbor topology
• Each neuron is connected to k
randomly selected neurons from all
previous neurons
• Connecons are determiniscally
generated based on (publicKey,
nonce) pair
• Ternary weight s (+1,0 or -1)
determined by hash funcon
• Clamped acvaon to prevent value
explosion
• Score based on input reconstrucon
at output layer
• All random number generaon is
determiniscally seeded using
cryptographic hashing
(KangarooTwelve) of the input keys,
ensuring reproducibility while
maintaining unpredictability of
outcomes.
Through November 2023 to December 2024,
the me frame used for the study in this
paper, the parameters used to test dierent
network architectures has changed. This
period includes 57 main changes from
EPOCH 83 (Nov 15th, 2023) to EPOCH 140
(Dec 18th, 2024). Empirical analysis of
Aigarth's development reveals three
signicant architectural transions that are
aligned with evoluonary adaptaon in 3
areas:
Data Processing Opmizaon: The
DATA_LENGTH parameter shows a clear
stepped reducon from 2000 to 1200, and
nally stabilizing at 256 aer EP99. This
reducon by 87.2% suggests an
opmizaon towards more ecient
informaon processing, potenally aligning
with biological systems' principle of
minimizing metabolic costs while
maintaining funconal capacity.
Soluon Space Dynamics: The
SOLUTION_THRESHOLD exhibits a notable
four-phase transion: an inial high
threshold of about 55.70%, followed by an
intermediate plateau 57.50% , and
stabilizing for several epochs at a lower
range at 16% of the reconstructed neurons
needed for valid soluons, with a nal
signicant increase to 52.73%, of the inputs
to be correctly reconstructed. This non-
linear progression suggests ongoing
adaptaon in problem-solving strategies,
with the latest small increase potenally
indicang a shi towards more stringent
soluon criteria, since the needed accuracy
is increased again.
Fig 5. Neuron reconstrucon targets over the 1 year of Aigarth.
Neural Architecture Evoluon: The
network's neural conguraon peaked at
32,768 hidden neurons, 128 mes the input
neurons, with several architectural shis,
combined with a 70% reducon in neural
elements, could suggest an opmizaon
towards more ecient informaon
processing structures, analogous to
biological neural network renement during
development.
Fig 6. Neural Architecture volume
These transions collecvely indicate a
systemac evoluon towards
computaonal eciency, with each
parameter adjustment potenally
represenng an opmizaon step in the
system's problem-solving capabilies. This
empirical evidence supports Aigarth's
design principle of emergent opmizaon
through evoluonary pressure, rather than
predetermined architectural constraints,
having in mind that this is only the rst step
in the process to create the intelligent
ssue menoned earlier.
B. Example Aigarth Code Abstracon
The current code is available on GitHub. The following is an abstracon of the scoring and
inializaon code:
Input: public_key P, mining_seed S, nonce N
Parameters:
n₁ = 256 {Input/output layer size (DATA_LENGTH)}
n₂ = 3000 {Hidden layer size (NUMBER_OF_HIDDEN_NEURONS)}
k = 3000 {Neighbors per neuron (NUMBER_OF_NEIGHBOR_NEURONS)}
T = 9000000 {Timesteps (MAX_DURATION)}
L = 1 {Value limit (NEURON_VALUE_LIMIT)}
Output: score
∈
[0,n₁]
1: procedure ComputeScore(P, S, N)
2: V ← array[n₁ + n₂ + n₁] {Neuron values, all inialized to 0}
3: V[0:n₁] ← S {Inialize input layer with mining seed data}
4: {Note: S already contains ternary values {-1,0,+1}}
5:
6: {Generate determinisc connecon pool}
7: seed_data ← [P, S, N] {Concatenate inputs}
8: R ← KangarooTwelve(seed_data, 32) {Generate 32-byte random seed}
9: pool ← array[2*M] {M = pool size, store neuron+connecon pairs}
10: for i
∈
[0,M) do
11: poolData ← DeterminiscRandom(R) {Using KangarooTwelve}
12: neuronIndex ← n₁ + (poolData mod (n₂ + n₁)) {Target neuron}
13: neighborOset ← (poolData ÷ (n₂ + n₁)) mod k
14: if neighborOset < k/2 then
15: sourceIndex ← (neuronIndex - k/2 + neighborOset) mod (n₁ + n₂ + n₁)
16: else
17: sourceIndex ← (neuronIndex + 1 - k/2 + neighborOset) mod (n₁ + n₂ + n₁)
18: pool[i] ← (neuronIndex, sourceIndex, signBit) {Store with random sign bit}
19:
20: {Main simulaon loop}
21: x ← 0 {Determinisc RNG state}
22: for t
∈
[0,T) do
23: idx ← x mod poolSize {Select connecon from pool}
24: (target, source, sign) ← pool[idx]
25: delta ← sign ? V[source] : -V[source]
26: V[target] ← V[target] + delta
27: V[target] ← clamp(V[target], -L, +L)
28: x ← x * 1664525 + 1013904223 {LCG parameters from score.h}
29:
30: {Compute score}
31: score ← 0
32: for i
∈
[0,n₁) do
33: if V[i] = V[n₁ + n₂ + i] then {Direct ternary comparison}
34: score ← score + 1
35: return score
36: procedure IsGoodScore(score, epoch)
37: threshold ← (epoch < MAX_EPOCH) ? soluonThreshold[epoch] :
SOLUTION_THRESHOLD_DEFAULT
38: if threshold > DATA_LENGTH or threshold ≤ 0 then
39: threshold ← SOLUTION_THRESHOLD_DEFAULT
40: return score ≥ (n₁/3 + threshold)
∨
score ≤ (n₁/3 - threshold)
Execuon Flow:
1.- Inialize network with binary input paern
2.- Generate determinisc random connecons based on (publicKey, nonce)
3.- Execute T random updates following determinisc sequence
4.- Compare output layer with input layer to compute score
5.- Score represents how many output neurons match input neurons
Sample Execuon:
1.- Input:
publicKey = 0x7B...A2
nonce = 0x4B...F1
miningData = [+1, 0, -1, 0, +1, ..., -1] # 256 ternary values
2.- Execuon:
- Generated 3,000 random connecons per neuron
- Performed 9M state updates
- Final output layer values: [+1, 0, -1, 0, +1, ..., -1]
3.- Sample Score Calculaon:
- Matching neurons: 178
- Final score: 178 (69.5% match)
Complete and updated source code is at: hps://github.com/qubic/core
This code abstracon and input parameters corresponds to EP140 Release v1.230.0
References
Abraham Nash (2024) Decentralized
Intelligence Network (DIN)
arXiv:2407.02461
Ackerman, P. L., & Heggestad, E. D. (1997).
Intelligence, personality, and interests:
Evidence for overlapping traits.
Psychological Bullen, 121(2), 219-245.
Amunts, K., Ebell, C., Müller, J., Telefont, M.,
Knoll, A., & Lippert, T. (2016). The Human
Brain Project: Creang a European Research
Infrastructure to Decode the Human Brain.
Neuron, 92(3), 574-581. DOI:
10.1016/j.neuron.2016.10.046
Antonia Creswell, Murray Shanahan, Irina
Higgins (2022) Selecon-Inference:
Exploing Large Language Models for
Interpretable Logical Reasoning
arXiv:2310.17512
Adly Templeton, Tom Conerly, Jonathan
Marcus, et al (2024) Scaling
Monosemancity: Extracng Interpretable
Features from Claude 3 Sonnet, Anthropic
Transformer Circuits Thread
Arrieta, A. B., Díaz-Rodríguez, N., Del Ser, J.,
Bennetot, A., Tabik, S., Barbado, A., ... &
Herrera, F. (2020). Explainable Arcial
Intelligence (XAI): Concepts, taxonomies,
opportunies and challenges toward
responsible AI. Informaon Fusion, 58, 82-
115.
Asngton, J. W., & Baird, J. A. (Eds.). (2005).
Why language maers for theory of mind.
Oxford University Press.
Basten, U., Hilger, K., & Fiebach, C. J. (2015).
Where smart brains are dierent:
quantave meta-analysis of funconal and
structural brain imaging studies on
intelligence. Intelligence, 51, 10-27.
Banks, J., & Oldeld, Z. (2007).
"Understanding pensions: Cognive
funcon, numerical ability and rerement
saving." Fiscal Studies, 28(2), 143-170.
Bay, G. D., Deary, I. J., & Goredson, L. S.
(2007). "Premorbid (early life) IQ and later
mortality risk: Systemac review." Annals of
Epidemiology, 17(4), 278-288.
Bay, G. D., Deary, I. J., & Der, G. (2018).
Does IQ explain socioeconomic inequalies
in health? BMJ, 348, g4163.
Bay, G. D., Deary, I. J., & Goredson, L. S.
(2009). Premorbid (early life) IQ and later
mortality risk: Systemac review. Annals of
Epidemiology, 19(5), 331-338.
Bengio, Y., Dafoe, A., & Russell, S. (2021).
Governing AI to avert risks. Nature Machine
Intelligence, 3(1), 5-6.
Benton, D. (2010). "The inuence of dietary
status on the cognive performance of
children." Molecular Nutrion & Food
Research, 54(4), 457-470.
Belsky, D. W., Mo, T. E., & Caspi, A.
(2017). "The longitudinal study of aging in
human young adults: Knowledge gaps and
research agenda." Journal of Gerontology:
Series B, 72(5), 722-731
Bereiter, C. (1995). "A disposional view of
transfer." Educaonal Psychologist, 30(3),
197-221.
Binet, A., & Simon, T. (1905). Méthodes
nouvelles pour le diagnosc du niveau
intellectuel des anormaux. Année
Psychologique, 11, 191-244.
Blaise Agüera y Arcas, Jyrki Alakuijala,
James Evans, Ben Laurie, Alexander
Mordvintsev, Eyvind Niklasson, Eore
Randazzo, Luca Versari (2024)
Computaonal Life: How Well-formed, Self-
replicang Programs Emerge from Simple
Interacon arXiv:2406.19108
Bostrom, N. (2014). Superintelligence:
Paths, dangers, strategies. Oxford
University Press.
Bouchard, T. J., Lykken, D. T., McGue, M.,
Segal, N. L., & Tellegen, A. (1990). Sources of
human psychological dierences: The
Minnesota Study of Twins Reared Apart.
Science, 250(4978), 223-228.
Brown, T. B., Mann, B., Ryder, N., Subbiah,
M., Kaplan, J., Dhariwal, P., ... & Amodei, D.
(2020). Language models are few-shot
learners. arXiv preprint
arXiv:2005.14165.systems (pp. 5998-6008)
& arXiv:1706.03762v7
Brundage, M., Avin, S., Clark, J., et al. (2018).
The malicious use of arcial intelligence:
Forecasng, prevenon, and migaon.
arXiv preprint arXiv:1802.07228.
Byrne, R. W., & Whiten, A. (Eds.). (1988).
Machiavellian Intelligence: Social Experse
and the Evoluon of Intellect in Monkeys,
Apes, and humans. Oxford University Press.
Calvin, C. M., Bay, G. D., Lowe, G. D., &
Deary, I. J. (2017). Intelligence in youth and
all-cause-mortality: Systemac review with
meta-analysis. Internaonal Journal of
Epidemiology, 46(1), 335-346.
Caspi, A., Wright, B. R. E., Mo, T. E., &
Silva, P. A. (1998). "Early failure in the labor
market: Childhood and adolescent
predictors of unemployment in the
transion to adulthood." American
Sociological Review, 63(3), 424-451.
Caell, R. B. (1941). "Some theorecal
issues in adult intelligence tesng."
Psychological Bullen, 38(9), 592-611.
Caell, R. B. (1963). "Theory of uid and
crystallized intelligence: A crical
experiment." Journal of Educaonal
Psychology, 54(1), 1-22.
Carroll, J. B. (1993). Human cognive
abilies: A survey of factor-analyc studies.
Cambridge University Press.
Carrington, S. J., & Bailey, A. J. (2009). Are
there theory of mind regions in the brain? A
review of the evidence. Brain and Cognion,
70(3), 211-222.
Cho, K., Van Merriënboer, B., Gulcehre, C.,
Bahdanau, D., Bougares, F., Schwenk, H., &
Bengio, Y. (2014). Learning phrase
representaons using RNN encoder-decoder
for stascal machine translaon. arXiv
preprint arXiv:1406.1078.
Chollet, F. (2019). On the measure of
intelligence. arXiv preprint
arXiv:1911.01547.
Chen Gao, Xiaochong Lan et al (2023) Large
Language Models Empowered Agent-based
Modeling and Simulaon: A Survey and
Perspecves arXiv:2312.11970
Chrisan, B. (2021). The Alignment Problem:
Machine Learning and Human Values. W. W.
Norton & Company.
Crevier, D. (1993). AI: The Tumultuous
History of the Search for Arcial
Intelligence. Basic Books.
Darwin, C., Wallace, A.R. (1858) On the
Tendency of Species to form Variees; and
on the Perpetuaon of Variees and Species
by Natural Means of Selecon. Journal of
the Proceedings of the Linnean Society,
Zoology 3:46-50
Dayeol Lee, Jorge António, Hisham Khan
(2024) Privacy-Preserving Decentralized AI
with Condenal Compung
arXiv:2410.13752
Davies, G., Lam, M., Harris, S. E., Trampush,
J. W., Luciano, M., Hill, W. D., ... & Deary, I. J.
(2018). Study of 300,486 individuals
idenes 148 independent genec loci
associated with general cognive funcon.
Nature Communicaons, 9(1), 2098.
Deary, I. J., Bay, G. D., & Gale, C. R. (2008).
"Childhood intelligence predicts voter
turnout, vong preferences, and polical
involvement in adulthood: The 1970 Brish
Cohort Study." Intelligence, 36(6), 548-555
Deary, I. J., & Caryl, P. G. (1997).
"Neuroscience and human intelligence
dierences." Trends in Neurosciences, 20(8),
365-371.
Deary, I. J., Harris, S. E., & Hill, W. D. (2019).
What genome-wide associaon studies
reveal about the associaon between
intelligence and physical and mental health.
Current Opinion in Psychology, 27, 6-12.
Deary, I. J., Whalley, L. J., Lemmon, H.,
Crawford, J. R., & Starr, J. M. (2000). The
stability of individual dierences in mental
ability from childhood to old age: Follow-up
of the 1932 Scosh Mental Survey.
Intelligence, 28(1), 49-55.
Deary, I. J., Strand, S., Smith, P., &
Fernandes, C. (2007). Intelligence and
educaonal achievement. Intelligence,
35(1), 13-21.
Deary, I. J., Weiss, A., & Bay, G. D. (2010).
Intelligence and personality as predictors of
illness and death: How researchers in
dierenal psychology and chronic disease
epidemiology are collaborang to
understand and address health inequalies.
Psychological Science in the Public Interest,
11(2), 53-79.
Deary, I. J., Johnson, W., & Houlihan, L. M.
(2009). "Genec foundaons of human
intelligence." Human Genecs, 126(1), 215-
232.
Dekhtyar S. et al. (2015). A life-course study
on the associaon of educaon with
Alzheimer's disease risk. Neurology, 85(10),
896-903. DOI:
10.1212/WNL.0000000000001872.
Devlin, J., Chang, M. W., Lee, K., &
Toutanova, K. (2018). Bert: Pre-training of
deep bidireconal transformers for
language understanding. arXiv preprint
arXiv:1810.04805.
Dickson, H., Laurens, K. R., Cullen, A. E., &
Hodgins, S. (2021). The relaonship
between intelligence and adult mental
health: A mulvariate analysis of
schizophrenia, bipolar disorder, and
depression. Schizophrenia Bullen, 47(3),
602-610.
Dunbar, R. I. M. (1993). "Coevoluon of
neocorcal size, group size and language in
humans." Behavioral and Brain Sciences,
16(4), 681-735.
Dunbar, R. I. M. (1998). "The social brain
hypothesis." Evoluonary Anthropology,
6(5), 178-190.
Dunbar, R. I. M. (1996). Gossip, Grooming,
and the Evoluon of Language. Harvard
University Press.
Dunbar, R. I. M., & Shultz, S. (2007).
"Evoluon in the social brain." Science,
317(5843), 1344-1347.
Dunn, J., & Brophy, M. (2020). Mind-
mindedness in parents, social cognion in
children: Understanding aachment and
reecve funconing. Developmental
Psychology, 56(4), 752-764.
Fergusson, D. M., & Horwood, L. J. (1997).
"Early disrupve behavior, IQ, and
educaonal achievement." Journal of
Abnormal Child Psychology, 25(3), 241-253
Ferreira D. et al. (2016). Cognive reserve
and the dynamic genome: The role of
lifestyle in Alzheimer’s disease. Froners in
Aging Neuroscience, 8, 142. DOI:
10.3389/fnagi.2016.00142.
Fitch, W. T. (2020). The evoluon of
language: A comparave review. Froners
in Psychology, 11, 517.
Floridi, L. (2020). AI and Its New Winter:
From Myths to Realies. Philosophy &
Technology, 33(1), 1-3.
Flanagan, D. P., & Harrison, P. L. (2012).
Contemporary Intellectual Assessment:
Theories, Tests, and Issues (3rd ed.). Guilford
Press.
Flynn, J. R. (1987). "Massive IQ gains in 14
naons: What IQ tests really measure."
Psychological Bullen, 101(2), 171-191.
Friederici, A. D. (2017). Language in our
brain: The origins of a uniquely human
capacity. MIT Press
Friston, K. (2010). The free-energy principle:
A unied brain theory? Nature Reviews
Neuroscience, 11(2), 127–138.
Gabriel Béna, Dan F. M. Goodman
(2021/2024) Dynamics of specializaon in
neural modules under resource constraints
arXiv:2106.02626
Gale, C. R., Hatch, S. L., Bay, G. D., & Deary,
I. J. (2017). Intelligence in childhood and risk
of major depression in adulthood: The 1958
Naonal Child Development Survey.
Psychological Medicine, 47(8), 1357-1365.
Gallese, V., & Lako, G. (2005). The Brain's
Concepts: The Role of the Sensory-Motor
System in Reason and Language. Cognive
Neuropsychology, 22(3/4), 455-479.
Garcez, A. D., Lamb, L. C., & Gabbay, D. M.
(2009). Neural-symbolic cognive
reasoning. Springer Science & Business
Media.
Goertzel, B., & Pennachin, C. (Eds.). (2007).
Arcial general intelligence (Vol. 2).
Springer.
Goredson, L. S. (2002). "g: Highly general
and highly praccal." In R. J. Sternberg & E.
L. Grigorenko (Eds.), The General Factor of
Intelligence: How General Is It? Lawrence
Erlbaum Associates.
Goredson, L. S. (2004). Intelligence: Is it
the epidemiologists' elusive 'fundamental
cause' of social class inequalies in health?
Journal of Personality and Social Psychology,
86(1), 174-199.
Hadi Esmaeilzadeh, Reza Vaezi (2021)
Conscious AI arXiv:2105.07879
Haier, R. J., Jung, R. E., Yeo, R. A., Head, K., &
Alkire, M. T. (2009). The neuroanatomy of
general intelligence: Sex maers.
NeuroImage, 45(4), 1174-1182.
Haier, R. J., Siegel, B. V., Tang, C., Abel, L., &
Buchsbaum, M. S. (1992). "Intelligence and
changes in regional cerebral glucose
metabolic rate following learning."
Intelligence, 16(3-4), 415-426.
Hampshire, A., Higheld, R. R., Parkin, B. L.,
& Owen, A. M. (2012). Fraconang human
intelligence. Neuron, 76(6), 1225-1237.
Hassabis, D., Kumaran, D., Summereld, C.,
& Botvinick, M. (2017). Neuroscience-
inspired arcial intelligence. Neuron,
95(2), 245-258.
Hendler, J. (2008). Avoiding another AI
winter. IEEE Intelligent Systems, 23(2), 2-4.
Henning Zakrisson (2023/2024) Trinary
Decision Trees for handling missing data
arXiv:2309.03561v2
Herbert Simon, Cockburn, Ransbotham,
Gerbert, Bughin, Sudarshan, Herweijer
(2021) The Evoluonary Dynamics of
Arcial Intelligence Ecosystems ETH Zurich
Hinton, G. E., & Salakhutdinov, R. R. (2006).
Reducing the dimensionality of data with
neural networks. Science, 313(5786), 504-
507.
Ho, J., Jain, A., & Abbeel, P. (2020). Denoising
diusion probabilisc models. Advances in
Neural Informaon Processing Systems, 33,
6840-6851.
Hogan, R., & Kaiser, R. B. (2005). What we
know about leadership. Review of General
Psychology, 9(2), 169-180.
Horn, J. L., & Caell, R. B. (1966).
"Renement and test of the theory of uid
and crystallized general intelligences."
Journal of Educaonal Psychology, 57(5),
253-270.
Humphrey, N. K. (1976). "The social funcon
of intellect." In P. P. G. Bateson & R. A. Hinde
(Eds.), Growing Points in Ethology (pp. 303-
317). Cambridge University Press.
Hochreiter, S., & Schmidhuber, J. (1997).
Long short-term memory. Neural
computaon, 9(8), 1735-1780.
Insel, T. R., Landis, S. C., & Collins, F. S.
(2013). The NIH BRAIN Iniave. Science,
340(6133), 687-688. DOI:
10.1126/science.1239276
Jackson, T., & Wang, Y. (2013). "Cognive
competence and the use of social media:
Predictors of online engagement and
behavioral outcomes." Journal of Applied
Developmental Psychology, 34(3), 166-174.
Jesus Garcia Fernandez, Nasir Ahmad,
Marcel van Gerven (2024) Ornstein-
Uhlenbeck Adaptaon as a Mechanism for
Learning in Brains and Machines
arXiv:2410.13563
John Achterberg, Duncan Astle, D. Akarca
(2024) Exploring Self-Awareness in AI
Systems Restack.io
Jinheng Wang, Hansong Zhou, Ting Song,
Shaoguang Mao, Shuming Ma, Hongyu
Wang, Yan Xia, Furu Wei (2024) 1-bit AI
Infra: Part 1.1, Fast and Lossless BitNet
b1.58 Inference on CPUs,
arXiv:2410.16144v2
Jensen, A. R. (1998). The g factor: The
science of mental ability. Praeger.
Jung, R. E., & Haier, R. J. (2007). The Parieto-
Frontal Integraon Theory (P-FIT) of
intelligence: Converging neuroimaging
evidence. Behavioral and Brain Sciences,
30(2), 135-154.
Javadi, A. H., Emo, B., Howard, L. R., Zisch, F.
E., Yu, Y., Knight, R., & Spiers, H. J. (2017).
Hippocampal and prefrontal processing of
network topology to simulate the future.
Nature Communicaons, 8(1), 14652.
Jung, R. E., Segall, J. M., Bockholt, H. J.,
Flores, R. A., Smith, S. M., Chavez, R. S., &
Haier, R. J. (2010). Neuroanatomy of
creavity. Human Brain Mapping, 31(3),
398-409.
Jouppi, N. P., Young, C., Pal, N., Paerson,
D., Agrawal, G., Bajwa, R., ... & Yoon, D. H.
(2017). In-datacenter performance analysis
of a tensor processing unit. In Proceedings of
the 44th annual internaonal symposium on
computer architecture (pp. 1-12).
Kaufman, J. C., & Sternberg, R. J. (2010). The
Cambridge Handbook of Creavity.
Cambridge University Press.
Kingma, D. P., & Ba, J. (2014/2017). Adam: A
method for stochasc opmizaon. arXiv
preprint arXiv:1412.6980v9.
Kingma, D. P., & Welling, M. (2013/2022).
Auto-encoding variaonal bayes. arXiv
preprint arXiv:1312.6114v11.
Kipf, T. N., & Welling, M. (2017). Semi-
supervised classicaon with graph
convoluonal networks. arXiv preprint
arXiv:1609.02907v4.
Krizhevsky, A., Sutskever, I., & Hinton, G. E.
(2012). ImageNet classicaon with deep
convoluonal neural networks. Advances in
neural informaon processing systems, 25.
Kuncel, N. R., Ones, D. S., & Sacke, P. R.
(2014). Individual dierences as predictors
of work, educaonal, and broad life
outcomes. Personality and Individual
Dierences, 67, 3-14.
Lake, B. M., Ullman, T. D., Tenenbaum, J. B.,
& Gershman, S. J. (2017). Building machines
that learn and think like people. Behavioral
and brain sciences, 40.
LeCun, Y., Boser, B., Denker, J. S., Henderson,
D., Howard, R. E., Hubbard, W., & Jackel, L.
D. (1989). Backpropagaon applied to
handwrien zip code recognion. Neural
computaon, 1(4), 541-551.
LeCun, Y., Boou, L., Bengio, Y., & Haner, P.
(1998). Gradient-based learning applied to
document recognion. Proceedings of the
IEEE, 86(11), 2278-2324.
LeCun, Y., Boou, L., Orr, G. B., & Müller, K.
R. (1998). Ecient backprop. In Neural
networks: Tricks of the trade (pp. 9-50).
Springer, Berlin, Heidelberg.
LeCun, Y., Bengio, Y., & Hinton, G. (2015).
Deep learning. Nature, 521(7553), 436-444.
Lishuang Wang et al (2024) Tensorized
NeuroEvoluon of Augmenng Topologies
for GPU Acceleraon, arXiv:2404.01817
Livingston G. et al. (2020). Demena
prevenon, intervenon, and care: 2020
report of the Lancet Commission. The
Lancet, 396(10248), 413-446. DOI:
10.1016/S0140-6736(20)30367-6.
Lusardi, A., & Mitchell, O. S. (2007).
"Financial literacy and rerement
preparedness: Evidence and implicaons for
nancial educaon." Business Economics,
42(1), 35-44.
Maturana, H., & Varela, F. (1980).
Autopoiesis and Cognion: The Realizaon
of the Living. Reidel.
Maximilian Beck et al. (2024). xLSTM:
Extended Long Short-Term Memory
arXiv:2405.04517
McCarthy, J., Minsky, M. L., Rochester, N., &
Shannon, C. E. (1955). A proposal for the
Dartmouth summer research project on
arcial intelligence. AI Magazine, 27(4),
12-12.
McGrew, K. S. (2009). "The Caell-Horn-
Carroll theory of cognive abilies: Past,
present, and future." In D. P. Flanagan & P. L.
Harrison (Eds.), Contemporary intellectual
assessment: Theories, tests, and issues (pp.
136-182). Guilford Press.
McCulloch, W. S., & Pis, W. (1943). A logical
calculus of the ideas immanent in nervous
acvity. The bullen of mathemacal
biophysics, 5(4), 115-133.
McClelland, J. L., & Rumelhart, D. E. (1986).
Parallel distributed processing: Exploraons
in the microstructure of cognion. Volume 1:
Foundaons. MIT Press.
McGurn, B., Starr, J. M., Topfer, J. A., Pae,
A., Whiteman, M. C., & Deary, I. J. (2008).
Childhood cognive ability and risk of late-
onset Alzheimer and vascular demena.
Neurology, 71(14), 1051-1056.
Messick, S. (1989). Validity. In R. L. Linn (Ed.),
Educaonal measurement (pp. 13-103).
American Council on Educaon and
Macmillan.
Mingrui Chen, Weizhi Lu (2021) Deep
Learning to Ternary Hash Codes by
Connuaon arXiv:2107.07987
Mills, M. C., & Rahal, C. (2019). The GWAS
diversity monitor tracks diversity by disease
in real me. Nature Genecs, 51(2), 19-20.
Mineault, P., Zanichelli, N., Peng, J. Z., et al.
(2024). NeuroAI for AI Safety. arXiv preprint
arXiv:2411.18526.
Mo, T. E., Caspi, A., Harrington, H., &
Milne, B. J. (2002). "Males on the life-course-
persistent and adolescence-limited
ansocial pathways: Follow-up at age 26
years." Development and Psychopathology,
14(1), 179-207.
Mo, T. E. (1993). "Adolescence-limited
and life-course-persistent ansocial
behavior: A developmental taxonomy."
Psychological Review, 100(4), 674-701.
Münzer, S., Zimmer, H. D., Schwalm, M.,
Baus, J., & Aslan, I. (2020). Computer-
assisted navigaon and the acquision of
route and survey knowledge. Journal of
Environmental Psychology, 28(1), 17-30
Nassim Dehouche, Richard Blythman (2023)
A Blockchain Protocol for Human-in-the-
Loop AI arXiv:2211.10859
Neubauer, A. C., & Fink, A. (2009).
Intelligence and neural eciency: Measures
of brain acvaon versus measures of
funconal connecvity in the brain.
Intelligence, 37(2), 223-229.
Needham, A., Libertus, K., & Heider, L.
(2014). What do infants learn about
objects? The role of self-directed acon.
Infancy, 19(3), 372-392.
Newell, A., & Simon, H. A. (1976). Computer
science as empirical inquiry: Symbols and
search. Communicaons of the ACM, 19(3),
113-126.
Nie, N. H., Junn, J., & Stehlik-Barry, K. (1996).
Educaon and democrac cizenship in
America. University of Chicago Press.
Nilsson, N. J. (1983). Arcial intelligence
prepares for 2001. AI Magazine, 4(4), 7.
Nisbe, R. E., Aronson, J., Blair, C., Dickens,
W., Flynn, J., Halpern, D. F., & Turkheimer, E.
(2012). Intelligence: New ndings and
theorecal developments. American
Psychologist, 67(2), 130–159.
Northo, G., & Bermpohl, F. (2004). Corcal
midline structures and the self. Trends in
Cognive Sciences, 8(3), 102-107.
Northo, G. (2014). How do resng state
changes in the brain impact self-
consciousness? Froners in Human
Neuroscience, 8, 1-16.
Northo, G. (2018). The Spontaneous Brain:
From the Mind–Body to the World–Brain
Problem. MIT Press.
Northo, G. (2011). Self and others: A
neurophilosophical approach. Philosophy,
Ethics, and Humanies in Medicine, 6(1), 1-
19.
Nowak, M. A., & Sigmund, K. (2005).
"Evoluon of indirect reciprocity." Nature,
437(7063), 1291-1298.
Pahor, A., Jausovec, N., & Jausovec, K.
(2019). The relaonship between brain
eciency and intelligence: The role of
resng-state EEG and structural
connecvity. Brain and Cognion, 135,
103581.
Peterson, J. & Homan, M. (2022). Ethical
challenges in AI: A review of global policy
implicaons. AI & Society.
Planning for AGI and beyond (2023) OpenAI
Plomin, R., & von Stumm, S. (2018). The new
genecs of intelligence. Nature Reviews
Genecs, 19(3), 148-159.
Poulton, R., Mo, T. E., & Silva, P. A.
(2015). "The Dunedin Muldisciplinary
Health and Development Study: Overview of
the rst 40 years, with an eye to the future."
Social Psychiatry and Psychiatric
Epidemiology, 50(5), 679-693.
Premack, D., & Woodru, G. (1978). "Does
the chimpanzee have a theory of mind?"
Behavioral and Brain Sciences, 1(4), 515-
526.
Rao, R. P. N., & Ballard, D. H. (1999).
Predicve coding in the visual cortex: A
funconal interpretaon of some extra-
classical recepve-eld eects. Nature
Neuroscience, 2(1), 79–87.
Raven, J. C. (2000). Manual for Raven’s
Progressive Matrices and Vocabulary Scales.
Oxford Psychologists Press.
Ree, M. J., & Earles, J. A. (1991). Predicng
training success: Not much more than g.
Personnel Psychology, 44(2), 321-332.
Reise, S. P., Scheines, R., Widaman, K. F., &
Haviland, M. G. (2013). Muldimensionality
and Structural Coecient Bias in Structural
Equaon Modeling: A Bifactor Perspecve.
Educaonal and Psychological
Measurement, 73(1), 5–26.
Rezende, D. J., Mohamed, S., & Wierstra, D.
(2014). Stochasc backpropagaon and
approximate inference in deep generave
models. Internaonal conference on
machine learning (pp. 1278-1286). PMLR.
Rizzola, G., & Craighero, L. (2004). The
Mirror-Neuron System. Annual Review of
Neuroscience, 27, 169-192.
Roberts, B. W., & Kuncel, N. R. (2007). "The
power of personality: The comparave
validity of personality traits, socioeconomic
status, and cognive ability for predicng
important life outcomes." Perspecves on
Psychological Science, 2(4), 313-345.
Roid, G. H. (2003). Stanford-Binet
Intelligence Scales, Fih Edion (SB5).
Riverside Publishing.
Rosenbla, F. (1958). The perceptron: a
probabilisc model for informaon storage
and organizaon in the brain. Psychological
review, 65(6), 386.
Rumelhart, D. E., Hinton, G. E., & Williams,
R. J. (1986). Learning representaons by
back-propagang errors. Nature,
323(6088), 533-536.
Rutwik Jain, Brandon Tran, Keng Chen,
Mahew D. Sinclair and Shivaram
Venkataraman (2024) PAL: A Variability-
Aware Policy for Scheduling ML Workloads
in GPU Clusters arXiv:2408.11919
Salgado, J. F., Anderson, N., & Viswesvaran,
C. (2020). Meta-analyc examinaon of the
predicve validity of general mental ability
for remote job performance. Journal of
Occupaonal and Organizaonal
Psychology, 93(2), 343-367.
Samuel Schmidgall, Rojin Ziaei, Jascha
Achterberg, Louis Kirsch, S Hajiseyedrazi,
Jason Eshraghian (2023), Brain-inspired
learning in arcial neural networks: a
review arXiv:2305.11252.
Savage, J. E., Jansen, P. R., Stringer, S.,
Watanabe, K., Bryois, J., de Leeuw, C. A., ...
& Posthuma, D. (2018). Genome-wide
associaon meta-analysis in 269,867
individuals idenes new genec and
funconal links to intelligence. Nature
Genecs, 50(7), 912-919.
Schmidt, F. L., & Hunter, J. E. (1998). "The
validity and ulity of selecon methods in
personnel psychology: Praccal and
theorecal implicaons of 85 years of
research ndings." Psychological Bullen,
124(2), 262-274.
Schmidhuber, J. (2015). Deep learning in
neural networks: An overview. Neural
Networks, 61, 85-117.
Schneider, W. J., & McGrew, K. S. (2012).
"The Caell-Horn-Carroll model of
intelligence." In D. P. Flanagan & P. L.
Harrison (Eds.), Contemporary intellectual
assessment: Theories, tests, and issues (pp.
99-144). Guilford Press.
Schrank, F. A., McGrew, K. S., & Mather, N.
(2014). Woodcock-Johnson IV.
Shane, S. (2003). A General Theory of
Entrepreneurship: The Individual-
Opportunity Nexus. Edward Elgar
Publishing.
Sidney Pontes-Filho, Stefano Nichele
(2019/2020) A Conceptual Bio-Inspired
Framework for the Evoluon of Arcial
General Intelligence arXiv:1903.10410v4
Smith, L. B., & Gasser, M. (2005). The
development of embodied cognion: Six
lessons from babies. Arcial Life, 11(1-2),
13-29.
Soldan A. et al. (2017). Cognive reserve
and Alzheimer's biomarkers. Neurology,
88(12), 1098-1106. DOI:
10.1212/WNL.0000000000003692.
Spearman, C. (1904). "General intelligence,
objecvely determined and measured."
American Journal of Psychology, 15, 201-
292.
Srivastava, N., Hinton, G., Krizhevsky, A.,
Sutskever, I., & Salakhutdinov, R. (2014).
Dropout: a simple way to prevent neural
networks from overng. The journal of
machine learning research, 15(1), 1929-
1958.
Srinivasa Deshmukh, Panait Luke (2022) AI
and the Sense of Self arXiv:2201.05576
Stanley & Miikkulainen (2002) Evolving
Neural Networks Through Augmenng
Topologies, Evoluonary Computaon
Stern Y. et al. (2019). Cognive reserve and
Alzheimer's disease. Neurology, 92(1), 10-
12. DOI: 10.1212/WNL.0000000000007122.
Sternberg, R. J. (2004). Successful
intelligence: Finding a balance. Yale
University Press.
Sparrow, B., Liu, J., & Wegner, D. M. (2011).
Google Eects on Memory: Cognive
Consequences of Having Informaon at Our
Fingerps. Science, 333(6043), 776-778.
Suzuki, S., et al. (2022). Dynamic Modulaon
of Sensory Processing by Motor State.
Journal of Neuroscience, 42(10), 2001-2016.
Taddeo, M., & Floridi, L. (2018). Regulate
arcial intelligence to avert cyber arms
race. Nature, 556(7701), 296-298.
Tatjana Wingarz, Anne Lauscher, Janick
Edinger, Dominik Kaaser, Stefan Schulte,
Mathias Fischer (2024) SoK: Towards
Security and Safety of Edge AI
arXiv:2410.05349
Thompson, N. C., Greenewald, K., Lee, K., &
Manso, G. F. (2022). The computaonal
limits of deep learning. arXiv preprint
arXiv:2007.05558v2.
Touvron, H., Lavril, T., Izacard, G., Marnet,
X., Lachaux, M. A., Lacroix, T., ... & Lample,
G. (2023). LLaMA: Open and Ecient
Foundaon Language Models. arXiv
preprint arXiv:2302.13971.
Tomasello, M. (2014). A Natural History of
Human Thinking. Harvard University Press.
Thiele, J. A., Faskowitz, J., Sporns, O., &
Hilger, K. (2024). Choosing explanaon over
performance: Insights from machine
learning-based predicon of human
intelligence from brain connecvity. PNAS
Nexus, 3, pgae519.
hps://doi.org/10.1093/pnasnexus/pgae51
9
Thurstone, L. L. (1938). Primary mental
abilies. University of Chicago Press.
Vahdat, A., & Kautz, J. (2020). NVAE: A deep
hierarchical variaonal autoencoder. In
Advances in Neural Informaon Processing
Systems (pp. 19667-19679).
Vaswani, A., Shazeer, N., Parmar, N.,
Uszkoreit, J., Jones, L., Gomez, A. N., ... &
Polosukhin, I. (2017/2023). Aenon is all
you need. In Advances in neural informaon
processing
Vincent, P., Larochelle, H., Bengio, Y., &
Manzagol, P. A. (2008). Extracng and
composing robust features with denoising
autoencoders. In Proceedings of the 25th
internaonal conference on Machine
learning (pp. 1096-1103).
Wechsler, D. (1997). WAIS-III: Wechsler
Adult Intelligence Scale (3rd ed.).
Psychological Corporaon.
Wechsler, D. (2008). Wechsler Adult
Intelligence Scale - Fourth Edion (WAIS-IV).
Pearson.
Wechsler, D. (2014). Wechsler Intelligence
Scale for Children - Fih Edion (WISC-V).
Pearson.
Whalley, L. J., & Deary, I. J. (2001).
Longitudinal cohort study of childhood IQ
and survival up to age 76. BMJ, 322(7290),
819.
Whiten, A., & Byrne, R. W. (1997).
"Machiavellian intelligence II: Extensions
and evaluaons." In A. Whiten & R. W. Byrne
(Eds.), Machiavellian Intelligence II:
Extensions and Evaluaons (pp. 1-23).
Cambridge University Press.
Xu, X., Weber, I., & Staples, M. (2021).
Blockchain and Distributed Ledger
Technology Use Cases: Applicaons and
Opportunies. Springer.
Yujian Li (2018) Theory of Cognive
Relavity: A Promising Paradigm for True AI
arXiv:1812.00136
