What children can teach us about building AI

Summary

Dieses Video untersucht die Geheimnisse des menschlichen Lernprozesses, insbesondere die Fähigkeit von Säuglingen, komplexe Sprachen und Motorik schnell zu erlernen, was AI-Algorithmen weit übertrifft. Professorin Elizabeth Spels Forschung in der Kindpsychologie zeigt, dass Säuglinge bereits im Schlaf lernen und sechs grundlegende Wissensarten teilen, die auch bei Tieren zu finden sind, sowie eine siebte, die Sprache, die einzigartig für den Menschen ist und kognitive Fähigkeiten katalysiert. Die Diskussion um AI-Forschung und das Potential, diese Wissensarten mit künstlicher Intelligenz zu verbinden, um KI-Systeme zu verbessern, wird ebenso behandelt.

Takeaways

• 🧠 Kinder lernen komplexe Fähigkeiten wie Sprache und Motorik schneller als kürzlich erkannt, was für KI-Forschung relevant ist.
• 👶 Elizabeth Spelke, eine Harvard-Professorin, untersucht, wie Kinder Wissen aufbauen und welche kognitiven Fähigkeiten sie von Geburt an mitbringen.
• 🔍 Spelke verwendet das Infant Gaze Tracking, um zu erforschen, was Säuglinge gelernt haben, indem sie die Dauer ihrer Aufmerksamkeit auf Objekte misst.
• 🌟 Säuglinge lernen bereits im Schlaf und können z.B. die Anzahl von Objekten oder die Worttrennung erkennen, bevor sie die Bedeutung der Wörter kennen.
• 🗣️ Die Fähigkeit zur Worttrennung ist ein Kernaspekt der Sprachentwicklung und wird von Säuglingen intuitiv erlernt, bevor sie die Grammatik oder die Bedeutung der Wörter kennen.
• 🧩 Professor Spelke identifizierte sechs Kernwissensbereiche, die sowohl bei Säuglingen als auch bei Erwachsenen und Tieren zu finden sind: Orte, Objekte, lebende Wesen, soziale Wesen, Zahlen und Geometrie.
• 🌐 Die siebte kognitive Fähigkeit, die nur beim Menschen vorkommt, ist die Sprache, die als kognitives Katalysator für das menschliche Denken gilt.
• 🤖 KI-Forschung konzentriert sich oft auf die Sprachfähigkeit, aber um KI-Systeme intelligenter zu machen, könnte es notwendig sein, auch die anderen sechs Kernkompetenzen zu integrieren.
• 🔢 Ein Beispiel für KI-Forschung, die auf Spelkes Forschung aufbaut, ist die SLM, die einen kleinen Sprachmodell mit einem symbolischen Löser kombiniert, um arithmetische Probleme zu lösen.
• 📈 KI-Systeme wie große Sprachmodelle sind derzeit nicht gut darin, hierarchisches Planen durchzuführen, was für KI-Systeme, die kompetent kodieren können, notwendig sein könnte.

Q & A

Was könnte das Geheimnis sein, um KI auf menschliches Niveau zu bringen?

Das Geheimnis könnte in den Gehirnen von Kindern liegen, da ihre Fähigkeit, komplexe motorische Fertigkeiten und Sprachen zu erlernen, weit über die Fähigkeiten der fortschrittlichsten maschinellen Lernalgorithmen hinausgeht.

Wie unterscheidet sich die kognitive Entwicklung von Säuglingen von der von KI-Systemen?

Säuglinge können in ihren ersten Lebensjahren ohne Anweisungen natürliche Sprachen erlernen, sich und Objekte bewegen, soziale Beziehungen aufbauen und Konzepte wie Zahl, Geometrie und Kausalität anwenden, während KI-Systeme dies oft nicht können oder es sehr langsam lernen.

Was ist der Unterschied zwischen dem Lernen von Säuglingen und dem von KI-Systemen?

Säuglinge können im Schlaf lernen und Informationen aufnehmen, während KI-Systeme normalerweise aktiv trainiert werden müssen. Säuglinge lernen auch, während sie schlafen, was KI-Systemen nicht möglich ist.

Wie verwendet Professor Spel die Methode der Säuglingsbeobachtung in ihren Studien?

Professor Spel verwendet die Säuglingsbeobachtung, um zu bestimmen, ob ein Säugling etwas gelernt hat, indem sie beobachtet, wie lange ein Säugling auf ein Objekt schaut, nachdem es ihm präsentiert wurde.

Was zeigte eine Studie über Säuglinge, die während des Schlafes Geräusche hören ließ?

Es zeigte, dass Säuglinge visuelle Interesse an Punkten zeigen, die einer Anzahl von Schallimpulsen entsprechen, die ihnen während des Schlafes gespielt wurden, was auf das Lernen während des Schlafes hindeutet.

Wie sind die sechs Kerntypen von Wissen, die Professor Spel in Säuglingen identifizierte, definiert?

Die sechs Kerntypen von Wissen sind Orte, Objekte, bewegte Wesen, soziale Wesen, Zahl und Geometrie. Diese Typen von Wissen interagieren nicht miteinander und scheinen unabhängig voneinander entwickelt worden zu sein.

Warum ist Sprache nach Ansicht von Professor Spel ein kognitives Katalysator für menschliches Denken?

Sprache ist ein kognitiver Katalysator, weil sie es ermöglicht, die sechs anderen Kernsysteme des Gehirns zusammenzuarbeiten, indem sie sie in Harmonie bringen und Ideen austauschen kann, was eine höhere kognitive Fähigkeit ermöglicht.

Was ist die Bedeutung der Kombinatorik und Rekursion in der Sprache für das menschliche Denken?

Die Kombinatorik und Rekursion in der Sprache ermöglichen es uns, einfache Konzepte zu kombinieren, um komplexe, hierarchisch strukturierte Gedanken zu bilden, was das abstrakte Denken ermöglicht.

Wie unterscheidet sich die Sprachfähigkeit von Tieren von der menschlichen Sprachfähigkeit?

Obwohl Tiere Kommunikationssysteme haben, die ein begrenztes Vokabular umfassen, können sie nur einen kleinen Teil der Konzepte ausdrücken, die sie in ihrem Kopf haben, im Gegensatz zur menschlichen Sprache, die nahezu jedes Konzept ausdrücken kann.

Was ist das Hauptziel der sLM-Forschung, die im Skript erwähnt wird?

Das Hauptziel der sLM-Forschung ist es, eine kleine Sprachmodell, das nicht in der Lage ist, numerische Probleme zu lösen, mit einem symbolischen Solver zu kombinieren, um die Fähigkeit des Systems zu verbessern, arithmetische Probleme in natürlicher Sprache zu lösen.

Warum sind hierarchische Planungsfähigkeiten für KI-Systeme wichtig?

Hierarchische Planungsfähigkeiten sind wichtig, weil sie es ermöglichen, über verschiedene Ebenen zu planen und zu denken, was natürlichen menschlichen Denkprozessen entspricht und für KI-Systeme zur Entwicklung von KI auf menschlichem Niveau notwendig ist.

Outlines

00:00

🧠 Kindliche Psychologie und KI-Forschung

Dieser Absatz stellt die Verbindung zwischen der kindlichen Psychologie und der KI-Forschung her. Es wird besprochen, dass die Fähigkeit von Kindern, komplexe Motorfertigkeiten und Sprachen zu erlernen, schneller ist als die von KI-Algorithmen. Die Rede ist von der Arbeit von Professorin Elizabeth Spelke, die sich mit der kindlichen Psychologie auseinandersetzt und versucht, herauszufinden, was Menschen von anderen Tieren unterscheidet. Ihre Forschung konzentriert sich auf das frühkindliche Erlernen von Sprachen und die Entwicklung von Wissen in verschiedenen Kategorien. Sie verwendet die Infant Gaze Tracking-Methode, um zu erforschen, wie Kinder lernen, und fand heraus, dass Kinder bereits im Schlaf lernen können. Diese Forschung könnte KI-Forschung inspirieren, indem sie zeigt, wie das menschliche Gehirn Wissen verarbeitet und wie es möglicherweise KI verbessern kann.

05:00

🔢 Sechse Kernwissensbereiche und die Sprache

In diesem Absatz werden sechs grundlegende Wissensbereiche beschrieben, die sowohl Menschen als auch Tiere teilen: Orte, Objekte, lebende Wesen, soziale Wesen, Zahlen und Geometrie. Diese Wissensbereiche interagieren nicht miteinander und scheinen unabhängig voneinander entwickelt worden zu sein. Menschen haben jedoch ein siebtes Wissen, das Sprache ist, das einzigartig für den Menschen ist und das Fundament für alle Sprachentwicklung im ersten Lebensjahr bildet. Sprache verbessert das Objekterkennen und die Zahlendarstellung und ist ein kognitives Katalysator, der es ermöglicht, dass verschiedene Teile des Gehirns zusammenarbeiten. Sprache hat rekursive Eigenschaften, die es ermöglichen, komplexe, hierarchisch strukturierte Gedanken zu bilden. Tiere haben zwar auch Kommunikationssysteme, diese sind jedoch eingeschränkt und können nicht alle Konzepte, die sie im Kopf haben, ausdrücken.

10:02

🤖 KI-Forschung und Kombination von Wissensbereichen

Dieser Absatz behandelt die KI-Forschung und wie man kleine Sprachmodelle mit symbolischen Lösern kombiniert, um Probleme zu lösen, die einen arithmetischen Anteil in natürlicher Sprache haben. Die Idee ist, Probleme in eine formale Sprache umzuwandeln, die dann an einen symbolischen Löser übergeben werden kann. Die Forschung konzentriert sich auf kleine Sprachmodelle mit weniger als 50 Milliarden Parametern, da numerische Fähigkeiten erst ab diesem Punkt erscheinen. Die Kombination aus Sprachmodell und symbolischem Löser ermöglicht es, arithmetische Probleme zu lösen, was direkt analog zu einer Kombination aus einem der sechs Kernsysteme und einem Sprachverarbeitungsmodul ist. Die Forschung zeigt, dass durch die Feinabstimmung des Modells anstatt eines erneuten Trainings eine tiefere Integration zwischen dem Sprachmodell und dem symbolischen Ausführungsmodul erreicht werden kann.

15:03

🏡 Hierarchische Planung und Weltmodelle

Hierarchische Planung ist ein menschliches Denkvermögen, das KI-Systeme derzeit nicht sehr bewältigen können. Um hierarchische Planung zu ermöglichen, ist ein hierarchisches Weltmodell notwendig, das verschiedene Ebenen der Realität repräsentiert. Das menschliche Gehirn kann aufgrund seiner verschiedenen kognitiven Fähigkeiten, wie räumliche Verarbeitung, Langzeitgedächtnis und numerische Repräsentation, automatisch hierarchische Pläne bilden. KI-Systeme, die nur über Sprachfähigkeiten verfügen, haben Schwierigkeiten, solche hierarchischen Pläne zu entwickeln. Die Integration von anderen kognitiven Fähigkeiten in KI-Systeme könnte zu einer verbesserten Fähigkeit zur hierarchischen Planung führen und zu KI-Systemen, die ähnlich kompetent sind wie der menschliche Verstand.

20:04

👋 Schlussfolgerung und Ausblick

Die Schlussfolgerung des Videos betont die Bedeutung der Forschung von Professor Spelke, die zeigt, wie Kinder lernen und welche sechs Kernwissensbereiche Menschen und Tiere teilen. Sie identifizierte auch das siebte System, die Sprache, als kognitives Katalysator, der es ermöglicht, dass verschiedene Teile des Gehirns zusammenarbeiten. In Bezug auf KI-Forschung wurde die SLM-Forschung vorgestellt, die kleine Sprachmodelle mit symbolischen Lösern kombiniert, um numerische Probleme zu lösen. Schließlich wurde die Notwendigkeit von hierarchischer Planung und Weltmodellen diskutiert, um KI-Systeme zu verbessern. Der Sprecher regt dazu an, über die verschiedenen Definitionen von Intelligenz nachzudenken und wie KI-Systeme intelligent sein könnten, unabhängig von menschlichen Perspektiven.

Keywords

💡Kinderpsychologie

Kinderpsychologie ist das Studium des Denkens, Fühlens und Verhaltens von Kindern. Im Video wird diese Disziplin genutzt, um zu untersuchen, wie Kinder schnell komplexe Fähigkeiten erlernen und wie dies für die Entwicklung von KI von Vorteil sein könnte. Die Forscherin Elizabeth Spelke untersucht, welche grundlegenden Kategorien des menschlichen Wissens existieren und wie das Wissen in jeder Kategorie von Anfang an entwickelt wird, was zentral für das Verständnis der kognitiven Fähigkeiten von Kindern und möglicherweise auch für KI ist.

💡Sprachentwicklung

Sprachentwicklung bezieht sich auf den Prozess, wie Kinder natürliche Sprachen erlernen. Im Video wird erwähnt, dass Kinder ohne Anleitung eine oder mehrere Sprachen erlernen und dies als Teil ihrer kognitiven Entwicklung. Die Fähigkeit, Sprache zu erlernen, wird als ein Schlüsselmerkmal der menschlichen kognitiven Fähigkeiten betrachtet und könnte für KI-Systeme, die nach menschlicher Intelligenz streben, von großer Bedeutung sein.

💡Kognitive Fähigkeiten

Kognitive Fähigkeiten umfassen die verschiedenen Aspekte des Denkprozesses, wie das Lernen, Behalten von Informationen und das Problemlösen. Im Video werden sechs grundlegende kognitive Fähigkeiten identifiziert, die sowohl von Menschen als auch von Tieren geteilt werden, und eine siebte, die spezifisch für Menschen ist: die Sprache. Diese Fähigkeiten sind für die Interaktion mit der Welt und die Entwicklung von KI-Systemen entscheidend.

💡Rekursive Kompositionalität

Rekursive Kompositionalität bezieht sich auf die Fähigkeit, einfache Konzepte zu kombinieren, um komplexere, hierarchisch strukturierte Gedanken zu bilden. Im Video wird dies als ein Schlüsselmerkmal der menschlichen Sprache und kognitiven Fähigkeiten dargestellt, das es Menschen ermöglicht, abstrakte Gedanken zu denken und sich von Tieren abzuheben. Diese Eigenschaft könnte für KI von entscheidender Bedeutung sein, die nach menschlicher Intelligenz strebt.

💡Kombinatorische Macht

Die kombinatorische Macht ist die Fähigkeit, Elemente zu verschiedenen Kombinationen zu bringen, um neue Ideen oder Lösungen zu schaffen. Im Video wird dies als eine Eigenschaft der Sprache beschrieben, die es ermöglicht, Konzepte flexibel zu verbinden und komplexe Gedanken zu bilden. Diese Fähigkeit ist für die kognitive Flexibilität und die Fähigkeit zur abstrakten Denkweise von zentraler Bedeutung.

💡Infant Gaze Tracking

Infant Gaze Tracking ist eine Methode, die verwendet wird, um zu bestimmen, ob ein Säugling etwas gelernt hat, indem man beobachtet, auf welche Objekte sie länger schauen. Im Video wird dies als eine Methode beschrieben, die von Elizabeth Spelke verwendet wird, um zu untersuchen, wie Säuglinge lernen und welche Wissenskategorien sie entwickeln. Dies ist ein wichtiger Aspekt der Forschung zur kognitiven Entwicklung von Säuglingen.

💡Symbolische Lösung

Symbolische Lösung bezieht sich auf die Verwendung formeller Algorithmen, um mathematische Probleme zu lösen. Im Video wird dies als eine Methode beschrieben, die in der KI-Forschung verwendet wird, um KI-Systemen die Fähigkeit zu geben, arithmetische Probleme zu lösen, indem sie kleine Sprachmodelle mit symbolischen Lösern kombinieren. Dies ist ein Beispiel für die Integration verschiedener kognitiver Fähigkeiten in KI-Systeme.

💡Hierarchical Planning

Hierarchical Planning ist der Prozess, bei dem Ziele und Aufgaben in einer Hierarchie organisiert werden, um komplexe Aufgaben effizienter zu planen und auszuführen. Im Video wird dies als eine Fähigkeit betrachtet, die KI-Systemen fehlt, und die für die Entwicklung von KI von hoher Intelligenz und Autonomie notwendig ist. Die Fähigkeit zur Hierarchischen Planung erfordert eine Hierarchie von Weltmodellen, die verschiedene kognitive Fähigkeiten integrieren.

💡Kognitive Katalysator

Ein kognitiver Katalysator ist ein Element, das die Interaktion und Integration verschiedener kognitiver Fähigkeiten fördert. Im Video wird die Sprache als einen kognitiven Katalysator beschrieben, der es Menschen ermöglicht, verschiedene kognitive Module wie numerische, räumliche und soziale Fähigkeiten zu verbinden und zusammenzuarbeiten, um komplexere Denkprozesse zu unterstützen.

💡Künstliche Intelligenz (KI)

Künstliche Intelligenz, oder KI, bezieht sich auf Systeme, die nach menschlicher Intelligenz gestaltet sind und in der Lage sind, Aufgaben zu lösen, die normalerweise intelligente Entscheidungen erfordern. Im Video wird KI als ein Ziel beschrieben, das durch das Verständnis und die Integration der kognitiven Fähigkeiten von Säuglingen und Kindern erreicht werden kann, was zu KI von noch höherer Komplexität und Autonomie führen könnte.

Highlights

The secret to unlocking human-level AI might lie in understanding how children learn.

Children's rapid mastery of complex skills outpaces current machine learning algorithms.

Elizabeth Spelke's research focuses on cognition in babies to understand human uniqueness.

Infant gaze tracking is used to study what infants have learned without their ability to speak.

Infants show interest in numbers and can learn concepts even while sleeping.

Babies are familiar with word segmentation before birth, a skill that takes adults much longer to learn.

Infants prefer listening to their native language and regional accents.

Professor Spelke identified six core types of knowledge shared by humans and animals.

Humans possess a seventh knowledge type, language, which is unique and a cognitive catalyst.

Language learning in humans enhances object recognition and potentially other cognitive functions.

Animal communication systems lack the complexity and combinatorial power of human language.

The paper 'sLM' discusses integrating a symbolic solver with a small language model for arithmetic reasoning.

Symbolic reasoning is incorporated into the fine-tuning of models, not just as a tool.

Hierarchical planning is a capability that current AI systems lack.

Integrating multiple core knowledge systems may lead to more advanced AI capabilities.

The video concludes by emphasizing the importance of hierarchical world models for AI planning.

Transcripts

hi everyone the secret to unlocking

human level AI might very well lie

within the minds of children their rapid

Mastery of complex motor skills and

languages far outpaces the capabilities

of our most advanced machine learning

algorithms what could be the key to this

remarkable learning prowess can the core

knowledge and brain architecture

observed in infants Inspire

unprecedented advancements in AI keep

watching to learn more this video has

three parts child psychology the special

soft

and relation to AI research part one

child psychology I recently attended an

academic AI conference in Vancouver

called triple AI it was a massive

conference with I think 2,200 accepted

papers I can't believe the scale of AI

conferences and it had some very

interesting invited talks one of those

talks was by someone from a slightly

different field Elizabeth S spel is a

professor at Harvard and she studies

child psychology essentially she's the

author of a few books including what

babies know and her research focuses on

studying cognition in babies to figure

out what makes humans different from

other animals and in particular what are

the Bedrock categories of human

knowledge and how does knowledge in each

category develop from its very

Beginnings here's a quote from her talk

abstract in their first years and with

no instruction children learn one or

more natural languages they develop a

host of skills for moving themselves and

manipulating objects they learn the

paths connecting places in their homes

and the bonds connecting people in their

social networks and they deploy concepts

of number geometry and causality that

continue to guide the reasoning of

adults how children do this is a great

unanswered question so how does

Professor spel study this problem and

apologies if I'm mispronouncing her name

there I don't remember the pronunciation

the experimental method that she uses in

her studies is to do infant gaze

tracking basically she figured out that

if an infant has learned something then

they're going to look at that thing for

a longer time period they're going to be

more interested in it because we're

talking about infants that are much too

young to talk at this phase so you have

to use an indirect method to figure out

if they've learned something one

experiment that the professor performed

was to try to get infants to learn the

difference between four objects and 12

objects she had them look at four dots

for example and then later was able to

show that they had learned that because

they would look more at four dots than

they would at 12 and it's even

multimodal an infant can listen to a

sound that occurs four times and then

later they'll be interested more in four

dots basically their brain is learning

the concept of number but what's really

fascinating is that the researchers

actually played The Sounds the four

sounds for example or the 12 sounds in a

row to infants while the infants were

sleeping and even in that case they

actually were more interested in four or

12 dots visually at a later date so in

other words even while they were

sleeping the infants were picking up

information and learning it there was

another study that showed that babies

are familiar with word segmentation

before their birth word segmentation is

basically the the problem of you listen

to a spoken language and you can figure

out where the word breaks are supposed

to be adults tend to learn word

boundaries very slowly if you've learned

a second language as I have been doing

with Japanese you'll know how difficult

it can sometimes be to go was that one

word or was that three but babies and

native speakers know word boundaries

intuitively this is even before they

know the meanings of any of the words

and before they know the grammar of the

words as well but there's enough

patterns in the speech that they're able

to figure it out I actually have

secondhand confirmation of babies being

able to to learn word segmentation I

have a friend who has English as her

first language but her mother spoke

Spanish while she was a really young kid

I don't know if this was before or after

her birth but regardless my friend was

never taught any Spanish at all however

when she tried to learn Spanish as an

adult her teacher said you have the ear

of a native speaker she was very good at

reproducing exactly the right sounds and

I would guess had very little trouble

with word segmentation well again this

is probably because both Spanish and

English were being spoken spoken around

her before she was born and as she was a

very young child but I suppose the

Spanish speaking must have stopped at

some point and then she just picked up

English as her only native language

Anyway by using this infant gaze

tracking metric researchers were able to

show that babies pay more attention to

someone speaking their native language

and not only that but they pay even more

attention to someone who's speaking with

their own Regional accent slightly older

kids would rather make friends with

someone who sounds like them than looks

like them if someone is of the same race

but has a foreign accent for example

then they would not really prefer that

person as a friend compared to someone

of A different race but that sounds like

they came from the same neighborhood so

word segmentation and being able to

recognize your native language are

really core parts of how infants and

children interact with the world and all

of that learning starts even before the

baby is born part two the Special Sauce

in her research and in her book what

babies know Professor spel identified

six core types of knowledge these six

core types of knowledge that are

represented in babies brains and adult

brains are also shared with animals they

are places objects animate beings social

beings number and geometry these six

types of core knowledge don't really

interact with each other and therefore

potentially evolved independently when

an animal or human is engaging with one

type of knowledge then they don't use

the others and as soon as they switch to

something else their brain completely

switches focus and yes plenty of animals

exhibit all the signs of these six

knowledge types as well presumably they

are what evolution deemed necessary to

survive in the physical world however

humans have a seventh knowledge type

this seventh type is language and it's

Unique to humans and it underlies all

the language learning that happens

within a baby's first year of Life we'll

talk more in a minute about why animal

systems of communication don't really

count as language in this sense when

humans are learning language doing so

actually improves their object

recognition if you learn the name for

something it's easier for you to recogn

ize it it might improve number

representation as well this is my

supposition but I think the reverse is

true as well if you learn a different

representation for a word or a concept

that exists in your language then that

will result in a more solid

understanding it only seems natural for

example recently I was looking at a map

of Southeast Asia which is the first

time I've looked at it in great detail

in a while and I spotted where Myanmar

was which I didn't really know and then

when I started reading news stories

about Myanmar it was easier for me to

remember because I had a visualization

or representation of what that country

was in fact Professor spel talks about

this seventh knowledge type or language

learning as an actual cognitive Catalyst

she thinks it's the secret ingredient

for human cognition and what allows us

to be so different from animals

specifically she thinks that human

language is the Catalyst that allows our

numeric architectonic and social modules

of our brain to all act in harmony it

allows them to join forces swap ideas

and essentially engage more multiple

parts of your brain at once instead of

just switching from one completely to

the other here's a quote what's special

about language is its combinatorial

power we can use it to combine anything

with anything apparently for example

children start integrating what they

know about their physical environment

and their navigational sense at around

the same time as the age when they begin

to master spatial language and think

about what's left and what's right

another paper I read points out that

language has recursive machinery and

this recursiveness allows us to flexibly

combine any concepts any concepts we

like into complex hierarchically

structured thoughts yes language as in

speech itself is also hierarchical and

contains very formalized structure but

the idea here is that that structure

that recursive property can actually be

applied to thoughts as well as the paper

puts it compositionality is a key

component of linguistically structured

thought so basically the ability to

think about something very small and

very simple and something else very

small and simple and combine them

together into a larger composite thought

is really what makes abstract thought

possible the legendary Chomsky even

called this the basic property and

argues that it evolved in humans because

of the way that it allows us to

structure thoughts rather than the way

that it allows us to structure speech

although that was definitely a desirable

side effect so although Dr spel is

basically calling the seventh system the

language system I guess you could also

call it the recursive ly structured

system or the recursion and composition

system or something like that so what

about language in animals after all some

animals seem pretty smart like you can

train a dog to push a button when it

wants to get the humans to do something

and you have monkeys and chimpanzees and

birds that all communicate with

languages that seem to have reasonably

large vocabularies but this is the

difference pretty much any concept that

a human might want to think about can be

expressed in language but for all known

animal communication systems the animals

can only express a small subset of the

concepts that they can actually

represent in their head for example

honeybees have color vision they can

remember the colors of flowers that they

visit but the honeybee language where

they dance to communicate to other

honeybees where to find food or whatever

allows them to communicate only the

spatial location the spatial direction

and distance of that food rather than

the color of the flower that they're

sending them to and that's all the

information that's actually needed for

the honey bees to survive which is good

but it means that the bees will never be

able to scaffold thoughts upon thoughts

and communicate to one another and build

up a culture over time every honeybee

has to start from scratch so the Special

Sauce is language or recursive

compositionality or whatever you want to

call it therefore if you're trying to

create really smart like human level AI

it makes sense to concentrate on those

six core types of knowledge that other

animals have and combine it with the

existing language expertise because so

far most models skip over all of those

six core competencies and they just head

straight for the language which is great

because that's the most complicated one

and the one that we interact with on a

day-to-day basis but it also means that

if you ask a language model to do some

reasoning it has only the language

capabilities there it can't fall back on

some spatial reasoning module to solve a

geometric problem or whatever perhaps

easier said than done though part three

relation to AI research one of the

papers I saw at triple AI was called s

LM which I think stands for symbolic

reasoning language model the idea of

this work is to take a small language

model they call it a Frugal language

model because it doesn't have many

parameters and get that model to offload

some computations to a symbolic solver

this would effectively be taking one of

those six core competencies that I

talked about and fusing it with the

language model in this work they focused

on small language models with less than

50 billion parameters because apparently

numeric reasoning capability only only

starts to emerge once you exceed that

amount the idea is to take in problems

that have an arithmetic component in

natural language and convert them into a

formal language which you can then pass

to a symbolic solver the symbolic solver

will use mathematically precise

algorithms to figure out the solution

and then pass it back up to the language

model which can answer the original

queries for example you could imagine

asking a language model to generate

python code which then calls a symbolic

execution engine they did experiments

with two main models GPT DJ which has 6

billion parameters and vuna which has 13

billion as the larger model vuna can

generate python code but gptj could only

generate what they called pseudo code

actually though what they call pseudo

code is essentially an Assembly Language

they can do additions subtractions and

other operations on variables that the

LM defined so the Assembly Language can

look up variables from the original

prompt basically or if you want to be

technical it looks a lot like an SSA

form inside a compiler but anyway how

does this paper work so first they take

the base model and they don't retrain it

at all everything done in this work is

based around fine-tuning there's two

main stages first they use Laura low

rank adaptation to fine-tune the model

to the formal language generation task

remember the formal language could be

just python or it could be their pseudo

code language second they fine-tuned the

model to actually use the solver and

they Incorporated a feedback loop to try

to judge the quality of the formal

language representation including for

example whether the python code actually

ran or not the results were pretty

impressive for example they showed a 30%

in terms of absolute percentage

Improvement on one of the tasks compared

to the base model that hadn't been

fine-tuned along with the symbolic

reasoning engine and that's a key Point

as well this isn't just a system where

they exposed a tool to an existing large

language model which is sort of how

internet search gets exposed for example

instead the symbolic reasoning is

actually included in the fine-tuning of

the model not quite during the training

phase days because they just used a

pre-trained model but it's absolutely

integral to the fine-tuning stage in

other words during fine tuning the model

would actually be generating some python

running it checking the output this is a

much deeper integration between the

language model and the symbolic

execution engine than just using it as a

tool would be and because all their work

was done on small models that did not

need to be retrained but just fine-tuned

this work was also very computationally

efficient and easy for researchers to

replicate so I'm a big fan of this work

and it's interesting because it does

represent combining one of those six

systems along with the language

reasoning capabilities and I suspect

there will have to be a lot more work

like this in the future as we start to

put together heterogeneous combinations

of different types of reasoning modules

in fact that was very specifically one

of the ways to build AGI that was called

out and I talked about that in a

previous video there was another talk at

the triple AI conference by the famous

deep learning Pioneer Yan Lun he called

out one of the main things that current

AI systems are lacking and that is

hierarchical planning based on

hierarchical world models so what does

that mean well I've touched on that in

the past as well by saying that large

language models are essentially

feedforward neural networks they can't

think on something deeply for a long

time because they don't really have

Cycles they're not recurrent neural

networks that's one of the reasons that

if you want an llm to solve a

complicated problem you should use chain

of thoughts and ask it to split its

reasoning into small steps because it's

a lot more likely that the difference

between one step and the next is going

to be feasible for a feed forward neural

network than trying to get it to do it

all in one shot another way of putting

this is that the llm doesn't have any

planning capabilities it can't plan

ahead and say oh in step five I'm going

to need this so I should do this now it

can figure that out if you ask it to

explicitly break down its reasoning into

steps but overall the planning abilities

are not very good even harder is a

hierarchical planning ability if you

have a hierarchical plan it means you

have a top level goal a slightly lower

level goal and even lower level goals

and furthermore the really lowlevel

goals don't have to be worked out in

detail in their entirety for example

let's say my goal is to make a sandwich

for myself but I don't have the

ingredients for the sandwich so a lower

level thing is I need to go to the

grocery store to get those ingredients

and lower level than that is I need to

figure out which direction I should go

in order to walk to the grocery store

like which path would be best to take

but in order to do that I need to figure

out what the weather is to decide if I

should wear a raincoat or just go out as

I am but in order to do that I need to

get up out of my chair and actually go

to that room to check the weather so as

soon as I think to myself wait I need to

make a sandwich then that hierarchical

plan can just form itself in my mind and

suddenly I'm standing up to go check the

weather and of course I didn't have to

plan out all the Tiny Steps in advance

once I step into that other room and

figure out what the weather is by I

don't know asking my Amazon device for

example I didn't have to map out in my

head all of the individual steps I'm

going to take all the way to the grocery

store or whatever because my brain can

Define the plan as time goes on also

important in this discussion is a

hierarchical world model you can only do

planning at different levels in a

hierarchy if you have a model of the

world that is actually accurate at

different levels hierarchical planning

seems very natural to humans but if you

have certain conditions then it can be

really hard for example if you have ADHD

you might go out to see a friend and

forget your jacket and get soaked by

rain along the way it's easy to get

fixated on a goal and Overlook the small

things or to live in the current moment

and not really plan ahead too much for

ADHD it can depend on the energy level

that you have whether or not you think

about all the necessary steps and I

think it's a skill we have to learn and

consciously practice as well in fact

when a professor first asked me a

problem in other words at the equivalent

of how do I get to the grocery store I

didn't know how to answer him I was just

like well you figure out the closest

grocery store and you go there but he

was like no no no in programming you

first have to figure out get up from

your chair take one step to the right

Etc and of course course now if I'm

doing a coding task or anything like

that I'm engaged in many many levels of

hierarchical planning so it will be

interesting to see when AI systems can

really engage in planning at similar

levels or even more advanced levels

because I think that's one of the key

components that's missing when you think

about can an AI system write code

competently for example and the way this

relates to everything else in this video

is that you want hierarchical planning

but that needs a hierarchical world

model and I think it's difficult to have

a hierarchical world model model if you

just have one reasoning component in

your system that only understands

language but if you have Incorporated

one or more of those other six core

knowledge systems then I think it's much

easier to have a hierarchical world

model my spatial reasoning capabilities

can tell me how to get up out of this

chair and over to the other room my

episodic or long-term memory can tell me

which path which turns I have to take to

get to the grocery store my memory of

what food tastes like can help me

remember which ingredients I need to

make my sandwich and my number

representation system can help me figure

out if the grocery store is overcharging

me for things or not so it will be very

interesting to see what AI systems will

be capable of once they start

integrating hierarchical World models

hierarchical planning and more of those

six core competencies that humans and

other animals have as a sort of

cognitive Baseline finally in conclusion

we talked about how Professor s spel

work on studying child psychology has

helped us figure out what babies know we

discovered that babies can start to

learn about the number of items they can

start to engage their social muscles

they start to engage language learning

and word segmentation even before

they're born and a lot of this learning

even happens while they're asleep the

professor also identified six core areas

in which humans seem to share a lot of

similarities with animals and those six

core knowledge systems do not interact

with each other if the brain is focusing

on one then the other types of reasoning

are shut off humans have a seventh

reasoning system which is what deal

deals with language and this seventh

system is what enables humans to think

abstractly to reason and to do far more

than other animals can in fact this

seventh system is a cognitive Catalyst

because it enables those six other core

systems to actually work together you

can use some of this bit of reasoning

along with some of that and combine it

together with language we're using

language somewhat Loosely here because

structured thought recursive thought and

compositional thought is actually the

mechanism that enables these different

types of reasoning to be used together

and then once you have thoughts in a

structured and whatever manner then by

using language processing parts of your

brain you can turn it into speech in

terms of AI research we talked about s

LM which takes a really small language

model that is not capable of numeric

reasoning and combines it with a

symbolic solver this enables the

combined system to solve arithmetic

problems it's directly analogous to one

of those six core systems the one that

deals with numbers getting combined with

a language processing module finally we

also talked about hierarchical planning

which is something that existing AI

systems like large language models are

not very proficient at hierarchical

planning seems possible once you have

hierarchical World representations that

could come about for example because

you're representing more of those six

core knowledge systems so next time you

walk out of the house without your

wallet or your keys you know that you

can blame your brain's hierarchical

planning system if you liked this video

please check out this previous one I

made where I talked about different

definitions of intelligence and how we

could think about whether an AI system

is intelligent or not outside of the

human lens well that's all I have for

today thank you very much for watching

bye