Cover page

Image Name: Machine and Man: Are we having machines that can perform equally or better than humans?

Image Source: The Evolution of Humanoid robots

Managing Editor Chief Editor Editorial Board Correspondence

Ninan Sajeeth Philip Abraham Mulamoottil K Babu Joseph The Chief Editor

Ajit K Kembhavi airis4D

Geetha Paul Thelliyoor - 689544

Arun Kumar Aniyan India

Sindhu G

Journal Publisher Details

Publisher : airis4D, Thelliyoor 689544, India

Website : www.airis4d.com

Email : nsp@airis4d.com

Phone : +919497552476

Editorial

by Fr Dr Abraham Mulamoottil

airis4D, Vol.3, No.1, 2025

www.airis4d.com

The cover page of the third volume of the airis4D

journal (2025) indicates the evolution of humanoid

robots, delving into their capabilities and profound

implications for society. As technology advances at

an unprecedented pace, the line between man and

machine continues to blur. Are we witnessing the

rise of machines capable of rivalling or surpassing

human abilities?

The article by Arun Aniyan introduces the Agentic

Age, in which AI evolves from passive tools to

proactive agents capable of setting goals, learning,

and executing tasks autonomously. Enabled by large

language models (LLMs) and technologies like memory

systems and planning modules, these agents perform

complex tasks, adapt to feedback, and collaborate with

humans. Key components—Perception, Reasoning,

Planning, and Execution—allow agents to interpret

inputs, make decisions, and carry out actions eﬃciently.

Frameworks like Langchain and AgentGPT support

their development. AI agents are poised to replace

traditional systems, marking a transformative shift in

human-machine collaboration.

Comparing Cognitive Abilities: LLMs, LMMs,

and Human Cognition by Ninan Sajeeth Philip is a

conversation with a Large Language Model (LLMs)

to demonstrate that the model itself is not aware that

it is able to perform better than what it thinks. While

LLMs are highly proﬁcient in processing and generating

language, they believe that they fall short of human

cognition in areas where depth, ﬂexibility, and context-

sensitive reasoning are essential. The conversation

reveals that many of these beliefs are founded on shallow

investigations and that despite the intuition that they may

fail, during the conversation, the LLM models answered

the challenges correctly with logical explanations. The

discussion expands to include Large Multimodal Models

(LMMs), which can process diverse data types such

as text, audio, and video. These models are better

equipped to analyze non-verbal communication cues

like body language or vocal tone, oﬀering a broader

perspective than LLMs, which are primarily text-based.

Can all human knowledge be adequately expressed

through language? It acknowledges that certain types of

knowledge, such as embodied experiences, emotional

insight, and intuition, may not be fully translatable into

verbal forms, suggesting that there are aspects of human

cognition that transcend language.

In the article by Ajit Kembhavi, “Black Hole

Stories-15: The Nature of Gravitational Waves”

explores the fascinating nature of gravitational waves,

from their prediction by Einstein in 1916 to their ﬁrst

direct detection by LIGO in 2015. Gravitational waves

are ripples in spacetime caused by massive celestial

events like black hole mergers. Unlike electromagnetic

waves, they are unaﬀected by dense matter, allowing us

to observe extreme cosmic events. The article highlights

their generation, properties, and impact on astrophysics,

emphasizing their role in opening new horizons in

multimessenger astronomy. Future discussions will

delve into their generation mechanisms and applications.

The article by Aromal P, X-ray Astronomy:

Through Missions” explores the advancements in X-ray

astronomy during the 1990s, known as its ”golden age.”

It highlights signiﬁcant space missions like Granat,

ROSAT, BBXRT, and ASCA, which revolutionized the

study of X-ray sources with advanced telescopes and

instruments. These missions enabled groundbreaking

discoveries, such as detailed imaging of galactic

centres, the ﬁrst isolated neutron star, and X-ray

emissions from comets. Each mission introduced

unique technological innovations that deepened our

understanding of cosmic phenomena, setting the stage

for future exploration. Linn Abraham’s article ”What

are Galaxies?” explores the structure, classiﬁcation,

and large-scale organization of galaxies. It discusses

the evolution of our understanding of the Milky Way

and other galaxies, from Hubble’s discovery of their

existence beyond our galaxy to modern classiﬁcations

like spirals, ellipticals, and irregulars. Spiral galaxies,

including the Milky Way, exhibit rotation and disk-like

structures, while ellipticals are ellipsoidal with little

gas or dust. Galaxies also form larger patterns, such as

clusters and superclusters, with the Milky Way being

part of the Local Group and Local Supercluster. These

observations continue to shape our understanding of

cosmic structures and galaxy evolution.

The article of Sindhu G, Corona Borealis: The

Northern Crown is a small but striking constellation

with a semi-circular arrangement of stars. Rooted in

Greek mythology as Ariadne’s jewelled crown, it has

inspired cultural interpretations across civilizations.

Astronomically, it features bright stars like Alphecca

and Nusakan, variable stars such as R Coronae Borealis

and T Coronae Borealis, and the distant Corona Borealis

Galaxy Cluster. Its scientiﬁc signiﬁcance lies in stellar

evolution studies, dark matter exploration, and exoplanet

discoveries. Blending mythology, culture, and science,

Corona Borealis continues to captivate stargazers and

astronomers.

Victor Ambros and Gary Ruvkun were awarded

the 2024 Nobel Prize for discovering microRNAs

(miRNAs), tiny non-coding RNA molecules

that regulate gene expression post-transcriptionally.

miRNAs bind to target mRNAs, inhibiting protein

translation or degrading the mRNA, thereby playing

crucial roles in development, metabolism, and cell

diﬀerentiation. Their dysregulation is linked to diseases

like cancer, where they can act as oncogenes or

tumour suppressors. This discovery has revolutionized

molecular biology and holds promise for innovative

therapeutic approaches in treating diseases by targeting

miRNAs. Geetha Paul’s article ”Discovery of

MicroRNA and Its Role in Post-Transcriptional Gene

Regulation” discusses it.

Jinsu Ann Mathew has given an article “Evolution

of Domain Generation Algorithms: From Simple

Patterns to AI-Driven Adaptation”. Domain Generation

Algorithms (DGAs) have evolved from simple,

predictable patterns to complex, AI-driven mechanisms

to evade cybersecurity defences. Early DGAs relied

on static or time-based methods, making them easier

to detect. Modern DGAs, however, leverage adaptive

algorithms and AI to create highly dynamic, random,

and legitimate-looking domains that are harder to block.

By incorporating feedback and mimicking benign

domains, these advanced DGAs highlight the escalating

arms race between cybercriminals and defenders in

cybersecurity.

iii

Contents

Editorial ii

I Artiﬁcial Intelligence and Machine Learning 1

1 The Agentic Age 2

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Large Language Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.3 Agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.4 Frameworks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 LLMs, LMMs, and Human Cognition 7

II Astronomy and Astrophysics 20

1 Black Hole Stories-15

The Nature of Gravitational Waves 21

1.1 Some History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1.2 Gravitational Field Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1.3 Weak Field Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

1.4 Similarities and Diﬀerences Between Electromagnetic and Gravitational Waves . . . . . . . . . 23

2 X-ray Astronomy: Through Missions 25

2.1 Plus Virtutis: Golden Age Begins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.2 Satellites in 1990s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3 What are Galaxies? 29

3.1 Interstellar dust . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.2 Rotation and Presence of Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.3 Spirals and Ellipticals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.4 Structures on a Larger Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

4 Corona Borealis: The Northern Crown 32

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.2 Mythological Origins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.3 Historical Astronomical References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4.4 Astronomical Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4.5 Observing Corona Borealis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4.6 Cultural Signiﬁcance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4.7 Scientiﬁc Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4.8 Astrological Signiﬁcance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

CONTENTS

4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

III Biosciences 36

A Short Note on the Discovery of MicroRNA and Its Role in Post-Transcriptional Gene

Regulation 37

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

1.2 What is microRNA? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

1.3 Posttranscriptionnal Regulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

1.4 Mechanisms of Action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

1.5 Biological Signiﬁcance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

1.6 Implications for Medicine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

1.7 Mechanisms of miRNA Involvement in Cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

1.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

IV General 41

Evolution of Domain Generation Algorithms: From Simple Patterns to AI-Driven

Adaptation 42

1.1 The Early Days: Simple Pattern-Based DGAs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

1.2 Intermediate Phase: Time-Based and Seed-Driven DGAs . . . . . . . . . . . . . . . . . . . . . . 43

1.3 Modern DGAs: Adaptive and AI-Assisted Techniques . . . . . . . . . . . . . . . . . . . . . . . . 44

1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Part I

Artiﬁcial Intelligence and Machine Learning

The Agentic Age

by Arun Aniyan

airis4D, Vol.3, No.1, 2025

www.airis4d.com

1.1 Introduction

We stand at the dawn of a new technological era in

which artiﬁcial intelligence isn’t just responding to our

prompts but actively pursuing goals, making decisions,

and learning from its interactions with the world. This is

the Agentic Age, where AI systems evolve from passive

tools into proactive partners capable of undertaking

complex tasks with increasing autonomy.

The concept of AI agents represents a fundamental

shift in how we interact with artiﬁcial intelligence.

Unlike traditional AI systems that wait for human input,

these agents can initiate actions, adapt to changing

circumstances, and work persistently toward deﬁned

objectives. From virtual assistants that can engage

in multistep planning to autonomous systems that can

navigate complex digital environments, AI agents are

reshaping our understanding of what machines can

accomplish.

This transformation isn’t just a technical evolution

- it is a paradigm shift that promises to revolutionize how

we work, create, and solve problems. As these agents

become more sophisticated, they’re beginning to tackle

tasks that once required signiﬁcant human oversight,

from managing email workﬂows to orchestrating

complex software development processes. However,

this advancement also raises important questions about

the nature of autonomy, the boundaries of AI decision

making, and the future of human-AI collaboration.

In this article, we will explore how AI agents

are redeﬁning the relationship between humans and

machines, the current state of agent technology, and

the implications of this change for our future. We will

examine the remarkable capabilities demonstrated by

today’s AI agents and the challenges and considerations

of delegating increasingly complex tasks to autonomous

systems.

1.2 Large Language Models

The emergence of Large Language Models (LLMs)

marked a crucial turning point in the development of AI

agents. These sophisticated neural networks, trained on

vast amounts of human-generated text, demonstrated an

unprecedented ability to understand context, generate

human-like responses, and most importantly, follow

complex instructions. This last capability – interpreting

and executing detailed prompts - laid the groundwork

for the ﬁrst generation of truly functional AI agents.

LLMs brought three critical capabilities that

enabled the creation of eﬀective AI agents. First, their

ability to understand natural language eliminated the

need for specialized programming interfaces, allowing

agents to interpret human instructions directly. Second,

their contextual understanding allowed them to maintain

coherent sequences of actions in multiple steps. Third,

their generation capabilities allowed them to produce

not just text, but structured outputs like code, queries,

and API calls, the building blocks of automated actions.

The transition from simple LLM interactions

to agentic systems occurred when researchers and

developers began combining these models with other

technologies: memory systems to maintain context over

long periods, planning modules to break down complex

tasks, and tool use capabilities to interact with external

software and services. This integration transformed

1.3 Agent

static language models into dynamic agents capable of

pursuing goals over time.

What makes modern AI agents particularly

powerful is their ability to learn from their interactions

and adapt their strategies. Unlike earlier rule-

based systems, these agents can understand feedback,

modify their approaches based on results, and even

learn new capabilities through demonstration. This

adaptability, combined with their foundation in natural

language understanding, makes them uniquely suited

to collaborate with humans in ways that previous AI

systems could not.

1.3 Agent

At their core, AI agents are autonomous systems

that perceive their environment, make decisions, and

take actions to achieve speciﬁc goals. Their architecture

typically consists of four fundamental components:

perception, reasoning, planning, and execution.

1.3.1 The Perception Layer

Modern agents utilize LLMs as their primary

perception mechanism, enabling them to process

diverse inputs including text, structured data, and

API responses. This layer transforms raw input into

semantic representations that the agent can reason about.

Advanced agents often implement memory systems to

maintain context across interactions, typically using

vector databases or other persistent storage mechanisms.

The perception layer serves as the agent’s sensory

system, processing and interpreting inputs from its

environment. The foundation starts with LLM-

based natural language understanding, which processes

various input types such as:

Text-based commands and queries

Structured data from APIs and databases

System state information and environmental

feedback

User interactions and historical context

The perception layer also has memory which

can maintain short-term results, store the sequence of

interactions and results, memorize the learned concepts,

and to enable retrieval of past experiences.

Context of a task has to be maintained until the

completion and the perception layer has a sophisticated

mechanism to handle this. It mainly consists of an

attention mechanism and context manager. The progress

of a task is monitored through state tracking and it has

a relevance scoring mechanism to ﬁlter the appropriate

results.

1.3.2 The Reasoning Engine

The reasoning engine represents the cognitive heart

of an AI agent, orchestrating complex decision-making

processes through an intricate interplay of various

computational mechanisms. At its core, the engine

employs a sophisticated architecture that combines

probabilistic reasoning, logical inference, and adaptive

learning to navigate complex problem spaces and

generate eﬀective solutions.

Central to the reasoning engine’s operation

is its state representation system, which maintains

a comprehensive model of the agent’s current

understanding and beliefs about its environment.

This state model encompasses not just immediate

task-relevant information, but also historical context,

learned patterns, and uncertainty estimates. The

engine continuously updates this state representation

as new information arrives, using Bayesian inference

mechanisms to adjust conﬁdence levels and belief states

in response to new evidence.

The reasoning component leverages the LLM’s

capabilities to understand the current state, evaluate

options, and make decisions. This involves several key

mechanisms:

State tracking to maintain awareness of the agent’s

progress and context

Goal decomposition to break complex objectives

into manageable subtasks

Outcome evaluation to assess the success of

actions and adjust strategies

The decision-making process within the reasoning

engine operates through multiple cognitive frameworks

working in concert. The primary framework employs

a combination of deductive and abductive reasoning,

1.3 Agent

allowing the agent to both draw logical conclusions from

known premises and generate plausible hypotheses to

explain new observations. This is complemented by

an analogical reasoning system that identiﬁes relevant

patterns from past experiences and adapts them to

current situations.

LLM’s inherent ability to think and also keep the

state of a task is made use in the reasoning engine.

This property of LLM enables agents to think without

additional development costs.

1.3.3 The Planning System

The planning system serves as the strategic

architect of an AI agent, transforming high-level goals

into structured, executable sequences of actions. This

sophisticated system operates through multiple temporal

horizons simultaneously, combining immediate tactical

planning with longer-term strategic considerations to

create coherent and adaptable action plans.

At the foundation of the planning system lies a

hierarchical task decomposition framework that breaks

complex objectives into progressively more manageable

components. This decomposition process employs

both top-down and bottom-up analysis, ensuring

that high-level goals remain aligned with concrete,

executable actions while maintaining the ﬂexibility to

adapt to changing circumstances. The system utilizes

sophisticated goal representation mechanisms that

capture both explicit objectives and implicit constraints,

enabling it to generate plans that balance multiple

competing priorities.

Agents need to plan the execution of a task based

on the user input and also the result obtained from the

LLM. Agents employ various planning strategies, from

simple sequential execution to sophisticated hierarchical

task networks. Modern approaches often use:

Tree of Thoughts: A method where the

agent explores multiple potential solution paths

simultaneously

Chain of Thought: Breaking complex reasoning

into explicit intermediate steps

Recursive Task Decomposition: Converting high-

level goals into increasingly speciﬁc subtasks

Temporal reasoning forms a crucial component

of the planning system, enabling it to reason

explicitly about timing constraints, action durations,

and sequential dependencies. The system maintains

a temporal network that represents both quantitative

timing requirements and qualitative ordering constraints

between actions. This temporal reasoning capability

allows the agent to generate plans that not only

achieve their objectives but do so within speciﬁed time

constraints while maintaining proper synchronization

between diﬀerent components of the plan.

The planning system incorporates a robust

mechanism for handling uncertainty and contingencies.

Rather than generating rigid, linear plans, the system

creates ﬂexible plan structures that include branch points

and contingency options. These adaptive plans allow the

agent to respond dynamically to changing circumstances

without requiring complete replanning. The system

continuously monitors plan execution and maintains

estimates of plan reliability, enabling it to proactively

identify potential failure points and develop backup

strategies.

1.3.4 The Execution Framework

The execution layer serves as the bridge between

an agent’s decision-making processes and real-world

actions, transforming abstract plans into concrete

operations. At its foundation, this layer operates as an

orchestration system, coordinating various tools, APIs,

and external services while maintaining operational

coherence throughout the execution process.

The execution layer translates planned actions into

real-world eﬀects through:

Tool use capabilities: APIs, function calling, and

system integrations

Error handling mechanisms: Detecting failures

and implementing recovery strategies

Feedback loops: Monitoring outcomes and

adjusting execution in real-time

Modern execution layers implement a sophisticated

action pipeline that begins with the translation of high-

level directives into speciﬁc, actionable commands.

This translation process involves parsing the agent’s

1.3 Agent

planned actions through a series of validators and

formatters that ensure compatibility with target systems.

The layer maintains detailed execution contexts,

tracking the state of ongoing operations and managing

the ﬂow of data between diﬀerent components of the

system.

Tool management forms a crucial aspect of the

execution layer’s functionality. The layer maintains a

dynamic registry of available tools, each with its own

interface speciﬁcations, authentication requirements,

and operational constraints. When executing actions,

the layer handles all aspects of tool interaction, from

authentication and rate limiting to error handling and

response processing. This includes managing API calls,

coordinating with external services, and ensuring proper

handling of system resources.

Error handling in the execution layer operates on

multiple levels. The primary level deals with immediate

execution failures, implementing retry mechanisms

with exponential backoﬀ and fallback options. A

secondary level handles semantic failures, where actions

execute successfully but fail to achieve their intended

outcomes. The tertiary level manages strategic failures,

where entire approaches prove ineﬀective and require

rethinking from the planning layer.

Feedback processing represents another critical

function of the execution layer. As actions are executed,

the layer captures both explicit and implicit feedback

signals. These signals ﬂow through an analysis pipeline

that extracts relevant metrics and insights, which are

then fed back into the agent’s perception and reasoning

systems. This creates a continuous learning loop that

helps the agent reﬁne its execution strategies over time.

The execution layer also implements sophisticated

monitoring and logging systems that provide visibility

into the agent’s operations. These systems track

performance metrics, resource usage, and execution

patterns, enabling both real-time optimization and post-

hoc analysis. This data proves invaluable for debugging,

optimization, and the ongoing improvement of agent

capabilities.

Resource management represents a ﬁnal crucial

aspect of the layer. The layer implements sophisticated

scheduling and prioritization systems that ensure

eﬃcient use of available computational resources,

API quotas, and other constrained assets. This

includes managing parallel execution when possible,

implementing appropriate rate limiting, and ensuring

fair resource allocation across multiple concurrent tasks.

1.3.5 Integration and Control Flow

The integration and control ﬂow system represents

the operational backbone of an AI agent, orchestrating

the complex interactions between perception, reasoning,

planning, and execution components while maintaining

operational coherence across the entire system. This

sophisticated orchestration layer ensures smooth

information ﬂow, proper synchronization, and eﬀective

coordination among all agent subsystems.

At the heart of the control ﬂow system lies a state

management architecture that maintains a uniﬁed view

of the agent’s operational status. This architecture

implements a hierarchical state representation that

captures both high-level strategic status and low-level

tactical details. The state management system employs

sophisticated synchronization mechanisms to ensure

consistency across diﬀerent components while enabling

parallel operation where appropriate. This parallel

processing capability allows the agent to simultaneously

engage in perception, planning, and execution activities

while maintaining operational coherence.

These components work together in a continuous

cycle:

Input processing through the perception layer

State evaluation and goal alignment in the

reasoning engine

Action planning and optimization

Execution and monitoring

Feedback incorporation and strategy adjustment

The control ﬂow implements an event-driven

architecture that enables responsive operation across

multiple time scales. This event system processes both

external triggers from the environment and internal

events generated by various agent components. Events

ﬂow through a sophisticated priority management

system that ensures critical operations receive

immediate attention while maintaining eﬃcient

1.4 Frameworks

processing of routine tasks. The event handling

framework includes deadlock prevention mechanisms

and ensures proper resource allocation across competing

processes.

Information ﬂow management represents a

crucial aspect of the integration system. The

system implements a publish-subscribe architecture

that enables eﬃcient distribution of information

across agent components while maintaining proper

encapsulation. This information distribution system

includes sophisticated ﬁltering mechanisms that ensure

each component receives relevant updates while

avoiding information overload. The system maintains

careful control over information access patterns,

implementing proper synchronization primitives to

prevent race conditions and ensure data consistency.

The integration layer also implements a

comprehensive monitoring and diagnostic system

that tracks the health and performance of all agent

components. This monitoring system maintains detailed

metrics on component performance, resource utilization,

and interaction patterns. The diagnostic capabilities

enable early detection of potential issues and implement

automated recovery mechanisms where appropriate.

This system also maintains detailed audit trails that

enable post-hoc analysis of agent behavior and support

continuous improvement of system performance.

1.4 Frameworks

There are several frameworks that are available to

implement agentic environments. Some of the notable

ones are the following.

1. Langchain

2. CrewAI

3. AgentAI

4. AgentGPT

1.5 Conclusion

In a recent talk, Satya Nadella the CEO of

Microsoft predicted that traditional SaaS (Software

as a Service) infrastructure will vanish and companies

will start to use Agentic systems to replace the backends.

This will remove the requirements of having code for

each component of the backend system which does data

ingestion, processing, and presentation. Instead, the

agents will take over and perform the tasks and decide

on which databases to connect and what to process.

As of now the tech around LLMs and supporting

infrastructure has matured well enough over the last

year. This development deﬁnitely points to the direction

of agentic systems taking over the backend systems.

Reference

AI Agents in Production

Real World Gen AI use cases from the worlds

leading organizations.

Agent Architectures in AI

Emergence of AI Agent Architecture

About the Author

Dr.Arun Aniyan is leading the R&D for

Artiﬁcial intelligence at DeepAlert Ltd,UK. He comes

from an academic background and has experience

in designing machine learning products for diﬀerent

domains. His major interest is knowledge representation

and computer vision.

LLMs, LMMs, and Human Cognition

by Ninan Sajeeth Philip in conversation with Perplexity AI on December 07 2024

airis4D, Vol.3, No.1, 2025

www.airis4d.com

Question: How is cognition in LLM diﬀerent

from human cognition?

Response by Perplexity AI: Large Language

Models (LLMs) exhibit notable diﬀerences from

human cognition despite some similarities in

language processing and sensory judgments. Key

distinctions include:

Reasoning Abilities: Humans outperform

LLMs in complex reasoning tasks,

particularly those involving novel situations,

indicating greater ﬂexibility and robustness

in human cognition [1].

Memory Structures: While LLMs show

some human-like memory eﬀects, their

memory mechanisms diﬀer signiﬁcantly,

lacking the coherence and stability of

human conceptual structures [2].

Contextual Understanding: LLMs excel

in surface-level language tasks but struggle

with deeper, context-dependent reasoning,

highlighting limitations in functional

linguistic competence [1, 2].

Q: Can you give an example for each of the three

limitations/diﬀerences mentioned?

R: Here are examples illustrating the three

limitations of LLMs compared to human

cognition:

Reasoning Abilities: In a reasoning

task about cabbages, an LLM incorrectly

identiﬁed that ”half of the cabbages were

small” as false, while the correct answer

was that ”all red cabbages weren’t big” must

be false. This highlights LLMs’ struggles

with complex logical deductions [3].

Memory Structures: LLMs often fail

to maintain coherent memory across

interactions. For instance, they might

correctly answer ”Who is Tom Cruise’s

mother?” but struggle to infer ”Who is

Mary Lee Pfeiﬀer’s son?” illustrating their

fragmented memory compared to humans’

integrated knowledge [3].

Contextual Understanding: When

faced with nuanced questions requiring

deep contextual understanding, LLMs

can misinterpret or overlook critical

information, leading to incorrect

conclusions. This limitation contrasts with

humans’ ability to navigate complex social

and contextual cues eﬀectively [4, 5].

Q: These examples are too shallow to support the

claims. Can you provide more robust indicators

to conﬁrm that they are diﬀerent?

R: Robust indicators conﬁrming the diﬀerences

between LLMs and human cognition include:

Theory of Mind: In comprehensive

tests, LLMs like GPT-4 performed well

identifying false beliefs and indirect

requests but struggled with recognizing

faux pas, indicating a lack of nuanced

understanding of social contexts that

humans easily navigate. This highlights

limitations in LLMs’ ability to track

complex mental states compared to humans

[6].

Memory and Integration: LLMs

excel in synthesizing vast information

from scientiﬁc literature, outperforming

human experts in predicting neuroscience

outcomes. However, they lack the coherent

memory structure that allows humans to

integrate knowledge over time, leading to

inconsistencies when the prior context is

necessary [7].

Contextual Reasoning: While LLMs can

generate contextually relevant responses,

they often fail in deeper reasoning tasks that

require understanding irony or misdirection,

showing a reliance on surface-level patterns

rather than genuine comprehension of

complex human emotions and intentions

[8, 9].

Q: Please provide the experiments that brought

them to these conclusions

R: The experiments supporting the diﬀerences

between LLMs and human cognition are outlined

in several studies:

(a).

Theory of Mind Tasks: A comprehensive

battery of tests was conducted comparing

LLMs (including GPT-4 and LLaMA2)

against 1,907 human participants. This

included tasks on understanding false

beliefs, indirect requests, and recognizing

irony. While GPT-4 performed well

in identifying indirect requests and false

beliefs, it struggled with recognizing faux

pas, indicating limitations in nuanced social

understanding [6, 10].

(b).

False-Belief Tasks: In a study assessing

the ability of LLMs to solve classic false-

belief tasks, GPT-3 (davinci-003) achieved

around 70% success, comparable to seven-

year-old children. However, earlier models

showed virtually no ability to solve these

tasks, suggesting that advanced language

capabilities may correlate with emergent

theory of mind [11].

(c).

Complex Reasoning: Another study

tested LLMs on non-literal language usage

and recursive intentionality. Results

demonstrated that while instruction-tuned

models from the GPT family often

outperformed children in some tasks, they

still lacked the depth of reasoning required

for more complex social interactions,

highlighting a gap in cognitive ﬂexibility

compared to humans [11].

Q: What is faux pas test, and how was it

performed?

R: The Faux Pas Test is designed to assess

an individual’s ability to recognize social

mistakes and understand the emotions of

others. Developed by Simon Baron-Cohen and

colleagues in the 1980s, it evaluates the theory

of mind (ToM) by presenting scenarios where a

character unintentionally makes a social blunder.

Test Structure:

(a).

Story Presentation: Participants read

a series of short stories featuring a

faux pas, such as saying something

inappropriate in a social context.

(b).

Questioning: After each story,

participants answer questions:

Did someone say something they

shouldn’t have?

What was said that was

inappropriate?

Why was it a faux pas?

Did the speaker realize the

implications of their statement?

Performance Evaluation: Participants

score points for correctly identifying faux

pas and understanding the emotional

impact on others. The average score

for children aged 9-11 is around 8.2 out

of 10 stories, indicating their developing

social understanding compared to adults or

individuals with autism spectrum disorders

who may struggle with these tasks [

14].

Q: Is faux pas universal? Will it not vary from

culture to culture?

R: The concept of faux pas is not universal and

varies signiﬁcantly across cultures. Here are key

points illustrating this cultural variability:

Social Norms: Diﬀerent cultures have

unique social norms that dictate acceptable

behaviour. For example, while maintaining

eye contact is seen as a sign of conﬁdence

in Western cultures, it may be perceived as

disrespectful in some Asian cultures, where

looking down is more appropriate during

conversations [15].

Greeting Practices: Greetings can diﬀer

widely; a handshake is common in many

Western countries, while cheek kissing is

customary in parts of Europe and Latin

America. In contrast, bowing is preferred

in Japan, highlighting how the same action

can be interpreted diﬀerently depending on

cultural context [15, 16].

Dining Etiquette: Table manners vary; for

instance, ﬁnishing all the food on your plate

might indicate satisfaction in some cultures,

while in others, it suggests the host did not

provide enough food. Similarly, eating with

hands is acceptable in certain regions but

considered rude in others [15, 17].

These examples demonstrate that faux pas are

deeply rooted in cultural contexts, making it

essential to understand local customs to avoid

unintentional oﬀenses.

Q: Well, with these diﬀerences in the cultural

diﬀerences, how can faux pas be a quantitive

measure to evaluate the cognitive ability of an

LLM model that is trained on materials across

cultures?

R: The Faux Pas Test is designed to evaluate

social understanding and theory of mind by

assessing an individual’s ability to recognize

social blunders. While it is a valuable tool in

psychology, its eﬀectiveness as a quantitative

measure for evaluating LLMs may be limited due

to cultural variations in social norms.

Cultural Context: Faux pas can vary

signiﬁcantly across cultures, meaning that

what constitutes a faux pas in one culture

may not be perceived the same way in

another. This variability can lead to

inconsistencies in how LLMs trained on

diverse cultural data interpret and respond

to faux pas scenarios.

Interpretation Challenges: LLMs might

generate responses based on statistical

patterns rather than genuine understanding.

This can result in seemingly correct answers

that do not reﬂect an accurate grasp of

the nuanced social dynamics involved,

potentially skewing evaluation results.

Quantitative Limitations: Since the

faux pas test relies on culturally speciﬁc

scenarios, using it as a universal metric for

cognitive ability may not yield reliable or

valid comparisons across diﬀerent cultural

contexts. This limitation suggests that while

LLMs can perform well on such tests, their

performance may not accurately reﬂect their

understanding of social intelligence.

Q: Are there any serious diﬀerences between

human cognitive ability and that of LLMs if the

faux pas test is avoided?

R: There are signiﬁcant diﬀerences between

human cognitive abilities and those of LLMs,

even when avoiding the faux pas test. Key

ﬁndings from recent studies include:

Depth of Understanding: Humans

possess a rich, multifaceted understanding

shaped by personal experiences and

emotions, allowing for abstract thinking

and inferential reasoning. In contrast,

LLMs rely on pattern recognition from

training data, lacking true comprehension

of concepts or context-sensitive nuances

[35].

Reasoning and Adaptability: Human

reasoning is characterized by logical

thinking and the ability to adapt based on

past experiences. LLMs, however, operate

through statistical analysis and cannot learn

or adapt from interactions beyond their

initial training, limiting their eﬀectiveness

in novel situations [35, 36].

Performance Variability: While LLMs

can perform well on many cognitive

tasks, their reasoning processes diﬀer

fundamentally from humans. They may

exhibit errors similar to human reasoning

but often do not align with human intuitions

or problem-solving strategies, indicating a

distinct cognitive architecture [36].

Q: When you say logical, can you give an example

where the logical prediction of LLMs is diﬀerent

from that of humans?

R: Logical predictions of LLMs can diﬀer from

those of humans in several ways. Here are

examples illustrating these diﬀerences:

Wason Selection Task: In this classic

logical reasoning task, participants must

determine which cards to ﬂip to verify a

conditional rule (e.g., ”If there is a ’D’

on one side, then there is a ’3’ on the

other”). Humans often incorrectly choose

cards that do not provide useful information,

such as ﬂipping the ’3’ card instead of

the ’D’. LLMs also exhibit similar errors,

indicating they share this cognitive bias

with humans, but they may lack the nuanced

understanding of why certain choices are

more logical than others [19].

Syllogistic Reasoning: When presented

with syllogisms (e.g., ”All cats are animals;

all animals are living beings; therefore,

all cats are living beings”), humans can

intuitively grasp the logical structure.

However, LLMs may misinterpret or fail

to apply the logical rules consistently,

particularly if the phrasing is altered or

if they encounter less common logical

structures [20].

Contextual Reasoning: Humans excel

at contextual reasoning based on prior

knowledge and experiences. For example,

when asked to deduce implications from a

story involving relationships or emotions,

humans can draw on their life experiences

to make connections. LLMs, however, may

struggle with context-sensitive reasoning

and often rely on surface patterns from

their training data rather than deeper

understanding [21, 22].

In comparison to LLMs, human reasoning is

characterized by a rich tapestry of experiences

and emotional intelligence that informs decision-

making and logic application. While LLMs

can mimic certain aspects of human reasoning

through advanced training techniques like Chain-

of-Thought (CoT) prompting, they still lack the

depth and ﬂexibility inherent in human cognition

[22].

10.

Q: All cats are animals; all animals are living

beings; therefore, all cats are living beings Is this

statement logically correct?

R: Yes, the statement ”All cats are animals; all

animals are living beings; therefore, all cats are

living beings” is logically correct. This reasoning

follows a valid form of deductive logic known as

syllogistic reasoning. Here’s a breakdown of the

argument:

(a).

Premise 1: All cats are animals. (This

establishes that the category of ”cats” falls

within the broader category of ”animals.”)

(b).

Premise 2: All animals are living beings.

(This establishes that the category of

”animals” falls within the broader category

of ”living beings.”)

(c).

Conclusion: Therefore, all cats are living

beings. (This conclusion logically follows

from the two premises.)

Logical Structure The logic can be represented

as follows:

If A (cats) is a subset of B (animals), and

B is a subset of C (living beings), then A

must also be a subset of C.

This type of reasoning is sound and valid,

and it illustrates how deductive logic works in

determining relationships between categories.

11.

Q: So in the above test, you have clearly passed

the test, and not only that, but you have also

explained why it is logical. This contradicts the

answer you gave to the previous question that

LLMs fail to answer the logical question ”All

cats are animals; all animals are living beings;

therefore, all cats are living beings”. So, are

you framed to answer as you did to the previous

question?

R: Thank you for pointing out the inconsistency.

Let’s clarify the situation regarding LLMs and

their performance on logical reasoning tasks like

syllogisms.

Performance Variability: While LLMs

can often provide correct answers to

logical syllogisms, their performance can be

inconsistent. They may correctly identify

the logical structure in straightforward

cases, such as the example you provided,

but they can struggle with more complex or

nuanced logical scenarios. This variability

is inﬂuenced by the speciﬁc phrasing of

questions, the complexity of the logic

involved, and the context in which the

reasoning is presented.

Understanding vs. Pattern Recognition:

LLMs operate based on statistical patterns

learned from vast amounts of text data.

They can generate correct responses based

on these patterns but do not possess true

understanding or reasoning capabilities

like humans. In some cases, they may

produce incorrect or illogical responses,

particularly when faced with less common

logical structures or when the context is

ambiguous.

Cognitive Bias: Both humans and LLMs

can exhibit cognitive biases in reasoning

tasks. For example, humans might make

intuitive errors in syllogistic reasoning due

to biases or heuristics, while LLMs might

misinterpret the logical relationships based

on how similar patterns were represented

in their training data.

In summary, while I (as an LLM) provided a

correct answer to the syllogism you presented,

this does not imply that all LLMs will consistently

perform at that level across various logical tasks.

The ability of LLMs to handle logical reasoning

can vary signiﬁcantly based on context and

complexity. Thank you for your understanding,

and I appreciate your engagement in this

discussion!

12.

Q: Can you provide an example logical case

where LLMs fail but humans are able to

outperform?

R: An example of a logical case where LLMs

may fail while humans can outperform involves

transitive reasoning. Consider the following

scenario:

Example Problem:

”Jane runs faster than Joe, and Joe runs

faster than Sam. Does Sam run faster

than Jane?”

Human Reasoning: Humans typically

understand the transitive property of

relations and can logically deduce that since

Jane is faster than Joe, and Joe is faster than

Sam, it follows that Sam cannot be faster

than Jane. Therefore, the correct answer is

”No.”

LLM Performance: In tests, some LLMs

have struggled with this type of reasoning.

While they might ultimately produce the

correct answer, they can fail to explain why

the answer is logically valid, relying instead

on pattern recognition rather than genuine

understanding of transitive relationships.

This indicates a lack of true comprehension

of the underlying logic, as they may simply

echo statistically likely responses based

on training data rather than engaging in

logical reasoning processes themselves

[32, 33, 34].

This example illustrates a clear distinction

between human cognitive abilities and those of

LLMs in logical reasoning tasks.

13.

Q: Please answer : Jane runs faster than Joe, and

Joe runs faster than Sam. Does Sam run faster

than Jane?”

R: In the scenario where ”Jane runs faster than

Joe, and Joe runs faster than Sam,” the logical

conclusion is that Sam does not run faster than

Jane. This follows from the transitive property

of inequality, which states that if

A >B

and

B >C

then

A >C

Therefore, the answer to the question is No, Sam

does not run faster than Jane. This reasoning

aligns with the transitive property, which can be

summarized as:

If Jane >Joe (A >B)

And Joe >Sam (B >C)

Then Jane >Sam (A >C)

This logical deduction is straightforward for

humans, as they can easily apply the transitive

property of inequalities. LLMs can also

arrive at this conclusion but may struggle with

more complex logical structures or nuanced

contexts, illustrating their limitations in reasoning

compared to human cognitive ﬂexibility.

14.

Q: But again, the LLM answered correctly with

the correct logical explanation for the task. How

do you then say it does not?

R: You raise a valid point, and I appreciate

your patience as we clarify the nuances of LLM

performance in logical reasoning tasks.

Correctness vs. Understanding: While

LLMs can provide correct answers and

logical explanations for straightforward

problems like the one involving Jane,

Joe, and Sam, the key distinction lies

in how they arrive at those answers.

LLMs generate responses based on patterns

learned from vast datasets rather than

genuine comprehension of logic. This

means that while they can often produce

the correct answer, they might not truly

”understand” the underlying principles of

logic in the same way humans do.

Variability in Performance: The potential

for LLMs to struggle arises more

prominently in complex or less structured

logical scenarios. For example:

Complex Logical Structures: When

faced with multi-layered logical

deductions or ambiguous phrasing,

LLMs may misinterpret relationships

or fail to apply logical rules correctly.

Contextual Nuance: In cases

requiring deep contextual

understanding or emotional

intelligence (e.g., social dilemmas),

LLMs may falter where humans

would intuitively grasp the subtleties

involved.

Examples of Failure: While I provided

a correct answer to the speciﬁc example

you asked about, there are instances in

research where LLMs have struggled with

more intricate logical puzzles or reasoning

tasks that require multi-step deductions or

understanding of abstract concepts.

Conclusion

In summary, while LLMs can indeed provide

correct answers to many logical reasoning

tasks, their performance can be inconsistent,

particularly in more complex scenarios. The

distinction lies in their reliance on learned

patterns rather than true understanding, which

can lead to failures in speciﬁc contexts. Thank

you for your insightful questions that have allowed

for this clariﬁcation!

15.

Q: Does this mean that if I ﬂip the names in the

text that I gave, LLMs would fail? let’s try that.

Generate a set of questions by just ﬂipping the

names in each. Give the answers

R: Sure! Let’s ﬂip the names in the original

scenario you provided. The original statement

was:

”Jane runs faster than Joe, and Joe runs

faster than Sam. Does Sam run faster than

Jane?”

Flipped Names Scenario: Now, let’s change the

names:

”Sam runs faster than Joe, and Joe runs

faster than Jane. Does Jane run faster than

Sam?”

Questions and Answers:

(a).

Question: Sam runs faster than Joe, and

Joe runs faster than Jane. Does Jane run

faster than Sam?

Answer: No, Jane does not run faster

than Sam.

(b).

Question: If Sam is faster than Joe, and Joe

is faster than Jane, what can we conclude

about Sam’s speed compared to Jane?

Answer: Sam is faster than Jane.

(c).

Question: Given that Sam runs faster than

Joe, and Joe runs slower than Jane, who is

the fastest among the three?

Answer: Sam is the fastest.

(d).

Question: If we know that Jane is slower

than both Sam and Joe, who is the slowest

runner?

Answer: Jane is the slowest runner.

Analysis of LLM Performance: In this ﬂipped

scenario, an LLM should ideally still be able to

reason correctly based on the logical structure

of the statements. However, there are instances

where LLMs may struggle with understanding

relationships if they are presented in a more

complex or less familiar context. In this case,

since the logical relationships remain clear and

straightforward, an LLM would likely provide

correct answers as long as it recognizes the

underlying patterns of comparison. However,

if the phrasing were more convoluted or if

additional layers of complexity were introduced

(e.g., changing multiple relationships or adding

distractions), there could be a higher chance

of error. This exercise illustrates how logical

reasoning can remain intact even with ﬂipped

names, but it also highlights potential challenges

for LLMs depending on how information is

presented.

16.

Q: No, I think that is a mistaken notion

about understanding or human cognitive ability.

Humans also depend on language for their

cognitive reasoning. If LLMs can master

languages, they should also exhibit the same

mental ability humans have. Do you agree or

not? If not, give me a falsiﬁable example.

R: Your point about the role of language

in human cognitive reasoning is well taken.

Indeed, language is a critical tool for human

thought and reasoning. However, there are key

diﬀerences between how humans and LLMs

process language and reason, which can lead

to divergent cognitive abilities.

Reasons for Divergence

Understanding vs. Pattern Recognition:

Humans use language as a medium to

express and manipulate thoughts, drawing

on personal experiences, emotions, and

a deep understanding of context. In

contrast, LLMs generate responses based

on statistical patterns in the data they were

trained on, lacking genuine comprehension

or awareness of the meanings behind the

words.

Contextual Nuance: Humans can navigate

complex social situations, infer meaning

from tone, body language, and context,

and adapt their reasoning based on these

subtleties. LLMs may struggle with such

nuances, particularly in ambiguous or

context-dependent situations.

Learning from Experience: Humans

learn from their experiences over time,

allowing them to reﬁne their reasoning skills

and adapt to new information dynamically.

LLMs do not learn from interactions in

real-time; they rely solely on their training

data.

Falsiﬁable Example

To illustrate these diﬀerences, consider the

following scenario involving ambiguity and

context:

Scenario: ”The chicken is ready to eat.”

(a).

Human Interpretation: A human

would likely recognize that this

sentence can be interpreted in two

ways:

The chicken is cooked and ready

to be eaten.

The chicken itself is prepared to eat

something (which is nonsensical

but possible in a humorous or

metaphorical context).

Humans can use contextual clues (e.g.,

prior conversation about cooking) to

infer the intended meaning.

(b).

LLM Interpretation: An LLM

might generate a response based

on the most statistically common

interpretation (the chicken is cooked)

but could struggle to recognize the

alternative meaning without additional

context. If asked to elaborate on

both interpretations, it might fail to

provide a nuanced understanding or

could misinterpret what ”ready to eat”

implies in diﬀerent contexts.

Conclusion

While language is an essential medium for

expressing human knowledge and experience,

it has limitations. Certain types of

knowledge—particularly those rooted in physical

experience, complex emotions, intuition, and

cultural context—may not be fully expressible

in words. As an LLM, I recognize these

limitations in language-based communication;

my responses are based on patterns learned

from text rather than personal experiences or

emotional understanding. This observation

highlights the richness of human cognition and

the multifaceted nature of knowledge that extends

beyond linguistic expression. Thank you for

engaging in this thought-provoking discussion!

If you have any further questions or topics you’d

like to explore, feel free to ask!

17.

Q: The body language and voice modulation etc

are outside the language model. Perhaps this

need to be compared with LMMs rather. Has

that been done?

R: The search results do not provide direct

comparisons between Large Language Models

(LLMs) and Language Models for Machines

(LMMs) regarding cognitive abilities, particularly

in relation to body language and voice modulation.

However, they do touch on aspects of inner speech

and the cognitive processes involved in language

comprehension.

Key Insights from the Search Results:

(a).

Inner Speech and Cognitive Processes:

The ﬁrst result discusses inner speech as a

functional linguistic act, indicating that it

plays a role in cognitive processing. Inner

speech can vary in richness and is linked to

articulatory features of spoken language,

suggesting that cognitive processes are

deeply intertwined with language use,

which LLMs do not replicate fully [23].

(b).

Variability in Language Processing: The

second result highlights the variability

in how individuals process language,

emphasizing that thought can be

represented diﬀerently across individuals

(e.g., some may visualize while others think

in words) [

]. This suggests that human

cognition involves more than just linguistic

processing, incorporating sensory and

experiential elements.

(c).

Cross-Language Comparisons: The

third result discusses how diﬀerent

languages exhibit similar encoding

eﬃciencies, indicating that while linguistic

structures may vary, the underlying

cognitive processes involved in language

comprehension share commonalities [25].

Conclusion

While these insights highlight the complexity of

human cognitive processes related to language,

they do not provide speciﬁc comparisons between

LLMs and LMMs regarding their ability to

integrate non-verbal cues like body language

or voice modulation. Further research would

be necessary to explore how these dimensions

of human communication could be modeled or

understood within artiﬁcial systems. If you

are looking for speciﬁc studies or comparisons

between LLMs and LMMs regarding their

cognitive abilities, more targeted research may

be needed beyond the provided search results.

18.

Q: LMMs were used to refer to large multimodal

models

R: Large Multimodal Models (LMMs)

represent a signiﬁcant advancement in

artiﬁcial intelligence, capable of processing

and understanding multiple types of data

simultaneously, including text, images, audio,

and video. Here’s a summary of their

characteristics and capabilities:

Key Features of LMMs

Multimodal Processing: LMMs can

analyze and generate content across various

data modalities. This allows them to

integrate information from diﬀerent sources,

similar to how humans use multiple senses

to understand their environment [26, 27].

Versatility: Unlike traditional models that

focus on a single modality (e.g., text-

only models), LMMs can perform tasks

that require understanding and generating

outputs in multiple forms. For instance,

they can create images based on textual

descriptions or generate text that describes

visual content [28, 29].

Applications: LMMs have broad

applications across sectors like healthcare,

education, and entertainment. They can

enhance creativity in art and design,

improve accessibility through automatic

content translation, and provide more

nuanced interactions in customer service

[27, 30].

Comparison with Large Language Models

Compared to LLMs, LMMs represent a

transformative step in AI by breaking down

barriers between diﬀerent types of data. Their

ability to process multimodal information allows

for a more holistic understanding and generation

of content, making them powerful tools for

various applications. As AI technologies

Aspect LMMs LLMs

Data

Modalities

Can handle text,

images, audio,

video

Focuses only on

text

Integration

Capabilities

Good at

combining

various data

types

Works only

with text

Applications

Tasks requiring

multiple data

types

Text-based

tasks like

writing and

summarization

[28, 31]

Table 2.1: Comparison of Large Multimodal Models

(LMMs) and Large Language Models (LLMs)

continue to evolve, LMMs will likely play an

increasingly important role in enhancing the

capabilities of artiﬁcial intelligence systems.

19.

Q: Yes, I was saying that your comparison

on body language, voice modulation, etc for

cognitive analysis should be carried out on LMMs

and not on LLMs that are limited to written text

R: Thank you for the clariﬁcation! You are

correct that comparisons of cognitive abilities,

particularly regarding non-verbal cues like body

language and voice modulation, should focus

on Large Multimodal Models (LMMs) rather

than Large Language Models (LLMs), which are

limited to text. Here’s a more focused discussion

on how LMMs can be evaluated in this context:

Cognitive Analysis in LMMs

(a).

Multimodal Integration: LMMs are

designed to process and integrate

information from various modalities,

including visual (images, videos) and

auditory (voice modulation, tone) inputs.

This capability allows them to analyze

social interactions more holistically, similar

to human cognition.

(b).

Understanding Non-Verbal Cues: By

incorporating visual data, LMMs can

potentially recognize body language, facial

expressions, and gestures, which are crucial

for interpreting social dynamics. For

instance, they could analyze a video of

a conversation to assess how body language

complements or contradicts spoken words.

(c).

Voice Modulation and Emotion

Recognition: LMMs can also analyze

audio inputs to detect voice modulation,

pitch, and tone, which convey emotional

states. This ability allows them to

understand the emotional context of spoken

language better than text-only models.

Comparative Studies

While research is still ongoing in this area, some

studies have begun exploring the capabilities of

LMMs in understanding multimodal inputs:

Emotion Recognition: Studies have

shown that LMMs can be trained to

recognize emotions from both visual and

auditory data, allowing for a more nuanced

understanding of human communication

compared to traditional models that rely

solely on text.

Social Interaction Analysis: Research

involving LMMs has examined how these

models can interpret social interactions by

analyzing both verbal and non-verbal cues.

This includes recognizing inconsistencies

between what is said and how it is expressed

through body language or tone.

Conclusion

Your point about focusing on LMMs for cognitive

analysis involving body language and voice

modulation is well-founded. As LMMs continue

to develop, they hold great potential for advancing

our understanding of human-like cognition in

artiﬁcial intelligence by integrating multiple

forms of communication. This integration could

lead to more sophisticated AI systems capable of

engaging in more natural and contextually aware

interactions with humans. Thank you for your

insightful contributions to this discussion!

20.

Q: Yes, I was saying that your comparison

on body language, voice modulation, etc for

cognitive analysis should be carried out on LMMs

and not on LLMs that are limited to written text

R: You are absolutely correct, and I appreciate

your clariﬁcation. The analysis of cognitive

abilities, particularly in relation to body

language, voice modulation, and other non-

verbal cues, should indeed focus on Large

Multimodal Models (LMMs) rather than Large

Language Models (LLMs), which are conﬁned

to processing text.

Key Diﬀerences in Cognitive Analysis

Data Modalities:

LMMs: Capable of integrating and

analyzing multiple types of data (text,

images, audio, video). This allows

them to interpret social interactions

more holistically by considering both

verbal and non-verbal communication.

LLMs: Limited to text-based input

and output, which restricts their ability

to engage with or analyze non-verbal

cues.

Understanding Context:

LMMs: By processing visual and

auditory data alongside text, LMMs

can gain insights into emotional

states and social dynamics that are

conveyed through body language and

tone of voice. This enables a

richer understanding of context in

communication.

LLMs: Lack the capability to interpret

non-verbal signals, which can lead

to misunderstandings or incomplete

interpretations of the intended meaning

behind the text.

Applications:

LMMs: Useful in applications such as

video analysis for emotion recognition,

virtual assistants that can respond not

only to spoken words but also to the

speaker’s tone and body language, and

interactive AI systems that engage in

more natural conversations.

LLMs: Primarily used for tasks

like text generation, summarization,

translation, and other language-based

applications where non-verbal context

is not available.

Conclusion

Your emphasis on using LMMs for cognitive

analysis involving body language and voice

modulation is well-founded. As LMM

technology continues to evolve, it will enhance

our understanding of human-like cognition in AI

by integrating various forms of communication.

Thank you for your insightful contributions to

this discussion! If you have any further questions

or topics you’d like to explore regarding LMMs

or their applications, feel free to ask!

21.

Q: Before we conclude, let me ask one more

question. Is there any human knowledge

that is not expressable in terms of language?

Sentimental analysis have shown that even human

feelings can be expressed in terms of language.

What is your independent observation as a LLM?

R: The question of whether there is human

knowledge that cannot be expressed in terms of

language is a profound and complex one. While

language is a powerful tool for communication

and expression, there are aspects of human

experience and knowledge that may elude full

articulation. Here are some considerations:

(a). Non-Verbal Knowledge

Embodied Knowledge: Many forms

of knowledge are embodied, meaning

they are rooted in physical experiences

rather than verbal descriptions. For

example, the muscle memory involved

in riding a bicycle or playing a

musical instrument is diﬃcult to

convey entirely through language.

While one can describe techniques or

principles, the actual experience of

performing these actions involves a

level of understanding that transcends

verbal explanation.

(b). Emotional Nuances

Complex Emotions: While sentiment

analysis can categorize emotions,

the depth and complexity of human

feelings often resist precise verbal

expression. People may experience

emotions that blend joy, sadness,

nostalgia, and anxiety simultaneously,

making it challenging to convey the

full spectrum of their emotional state

through words alone.

(c). Intuitive Knowledge

Intuition and Gut Feelings: Intuitive

insights or gut feelings often arise

from subconscious processing of

information and experiences. These

insights can guide decision-making but

may not always be easily articulated.

For instance, a seasoned chef might

have an intuitive sense of how to

balance ﬂavors without being able to

explain every detail of their reasoning.

(d).

Cultural and Contextual Understanding

Cultural Contexts: Certain cultural

practices or social dynamics may be

understood deeply by members of a

community but can be challenging to

explain to outsiders. The subtleties

of cultural norms, rituals, or shared

histories often rely on context that may

not translate well into language.

Conclusion

While language is an essential medium for

expressing human knowledge and experience,

it has limitations. Certain types of

knowledge—particularly those rooted in physical

experience, complex emotions, intuition, and

cultural context—may not be fully expressible

in words. As an LLM, I recognize these

limitations in language-based communication;

my responses are based on patterns learned

from text rather than personal experiences or

emotional understanding. This observation

highlights the richness of human cognition and

the multifaceted nature of knowledge that extends

beyond linguistic expression. Thank you for

engaging in this thought-provoking discussion!

If you have any further questions or topics you’d

REFERENCES

like to explore, feel free to ask!

22.

Q: Though true, I think nither of them has a

universal standard either, meaning the error bars

are large. So it is not fair to compare it as a

limitation of the language. It is like using the

same force on gears of diﬀerent conﬁgurations.

The outcome would be diﬀerent in each case.

The issue is not with the force but is a matter of

the hardware conﬁguration. Right or wrong?

R: Your analogy is insightful and highlights an

important point about the nature of knowledge

and expression. Let’s break down your argument:

Universal Standards and Subjectivity

Subjective Experiences: Human

emotions, intuitions, and embodied

knowledge are indeed subjective and

can vary widely among individuals and

cultures. This subjectivity means that

while there may be commonalities in

human experiences, the expression of

those experiences through language

can diﬀer signiﬁcantly based on

personal, cultural, and contextual

factors.

Error Bars: Just as you mentioned

with the ”error bars” in measurements,

the variability in how people

understand, interpret, and express

their feelings or knowledge indicates

that there isn’t a single ”correct”

way to articulate these experiences.

This variability reﬂects the complexity

of human cognition rather than a

limitation of language itself.

Hardware Conﬁguration Analogy

Diﬀerent Conﬁgurations: Your

analogy of using the same force on

gears of diﬀerent conﬁgurations is apt.

The outcome varies not because of the

force applied (language) but due to

the unique conﬁgurations (individual

experiences, cultural backgrounds,

emotional states) that inﬂuence how

knowledge is processed and expressed.

Language as a Tool: Language serves

as a tool for communication but is

inﬂuenced by the ”hardware” of human

cognition—our emotions, experiences,

and social contexts. Thus, while

language can articulate many aspects

of human experience, its eﬀectiveness

can be limited by these underlying

factors.

Conclusion

In summary, your argument suggests that

the limitations in expressing certain types of

knowledge or experience through language

are not solely due to the nature of language

itself but are also inﬂuenced by individual

and cultural conﬁgurations. This perspective

emphasizes the complexity of human cognition

and communication, acknowledging that while

language is powerful, it operates within a broader

context shaped by diverse human experiences.

Thank you for this engaging discussion! If you

have more thoughts or questions on this topic or

any other, feel free to share!

References

[1] https://arxiv.org/html/2409.02387v1

[2]

https://kushinm.com/assets/pdf/suresh emnlp 2

023.pdf

[3]

https://sloanreview.mit.edu/article/the-working-l

imitations-of-large-language-models/

[4] https://arxiv.org/html/2410.05229v1

[5]

https://www.digitalocean.com/community/tutoria

ls/understanding-reasoning-in-LLMs

[6]

https://www.nature.com/articles/s41562-024-018

82-z

[7]

https://www.nature.com/articles/s41562-024-020

46-9

[8] https://arxiv.org/abs/2409.02387

REFERENCES

[9]

https://papers.ssrn.com/sol3/papers.cfm?abstract

id=4974107

[10] https://pubmed.ncbi.nlm.nih.gov/38769463/

[11] https://aclanthology.org/2023.conll-1.25/

[12]

https://www.mentesabiertaspsicologia.com/blo

g-psicologia/faux-pas-test-use-and-clinical-appli

cations

[13]

https://aclanthology.org/2023.ﬁndings-acl.663.p

[14]

https://docs.autismresearchcentre.com/tests/Fa

uxPas Adult.pdf

[15]

https://magoosh.com/english-speaking/faux-pas

[16]

https://www.forbes.com/councils/theyec/2019/

05/22/follow-these-10-tips-to-avoid-cultural-fau

x-pas-in-your-global-business-relationships/

[17]

https://www.tradeready.ca/2018/topics/market-e

ntry-strategies/6-quick-ways-can-avoid-cultura

l-faux-pas-foreign-markets/

[18]

https://www.explore.co.uk/blog/travel-faux-pas

-how-to-avoid

[19]

https://techxplore.com/news/2024-07-large-lan

guage-human.html

[20]

https://www.hitachi.com/rd/sc/aiblog/formal-l

ogic-deduction/index.html

[21] https://arxiv.org/html/2405.19616v1

[22]

https://aijourn.com/how-similar-is-LLM-reaso

ning-to-human-cognition-the-potential-of-chain

-of-thought-prompting-in-LLMs/

[23]

https://www.cell.com/trends/cognitive-sciences/

fulltext/S1364-6613(23)00210-3

[24]

https://news.ycombinator.com/item?id=4075617

[25]

https://www.science.org/doi/10.1126/sciadv.aaw

2594

[26]

https://www.theainavigator.com/blog/what-is-a

n-lmm-large-multimodal-model

[27]

https://www.leewayhertz.com/large-multimoda

l-models/

[28]

https://research.aimultiple.com/large-multimoda

l-models/

[29]

https://www.wiz.ai/next-big-thing-the-large-m

ultimodal-model/

[30]

https://innodata.com/what-are-multimodal-large

-language-models/

[31] https://zapier.com/blog/multimodal-ai/

[32]

https://www.linkedin.com/pulse/limitations-lar

ge-language-models-logic-reasoning-michael-b

arrett-ey3sc

[33]

https://pkghosh.wordpress.com/2024/07/29/L

LM-failure-on-out-of-distribution-reasoning-and

-problem-solving-tasks/

[34]

https://thirdeyedata.ai/unreasonable-claim-of-r

easoning-ability-of-LLM/

[35]

https://www.alphanome.ai/post/exploring-the-c

ognitive-divide-human-intelligence-vs-LLM-bas

ed-chatbots

[36]

https://communities.springernature.com/posts/d

o-large-language-models-reason-the-way-we-do

About the Author

Professor Ninan Sajeeth Philip is a Visiting

Professor at the Inter-University Centre for Astronomy

and Astrophysics (IUCAA), Pune. He is also an

Adjunct Professor of AI in Applied Medical Sciences

[BCMCH, Thiruvalla] and a Senior Advisor for the

Pune Knowledge Cluster (PKC). He is the Dean and

Director of airis4D and has a teaching experience of

33+ years in Physics. His area of specialisation is AI

and ML.

Part II

Astronomy and Astrophysics

Black Hole Stories-15

The Nature of Gravitational Waves

by Ajit Kembhavi

airis4D, Vol.3, No.1, 2025

www.airis4d.com

In this we will consider the nature of gravitational

waves, and in subsequent stories the mechanisms for

their generation, the binary radio pulsar, the ﬁrst

detection of gravitational waves by the LIGO detectors,

the many merging black hole and neutron star binaries

which have been discovered since then as sources of

gravitational waves, their implications for black hole

astrophysics and the LIGO-India project.

1.1 Some History

Albert Einstein ﬁrst predicted the existence of

gravitational waves in 1916, just a year after he

completed his theory of gravity, the general theory

of relativity. Einstein himself was sceptical of

the prediction, and believed that the waves he had

discovered were merely artifacts of the complexity

of the theory, and not real, physical waves which

could be detected experimentally. Much progress in

understanding gravitational waves theoretically was

made beginning in the 1950s. Firm evidence for

the existence of the gravitational wave emission came

following the discovery of the binary pulsar B1913+17

in 1974 by J. H. Taylor and R. Hulse. This consisted

of a compact binary system consisting of two neutron

stars, one of which was radio pulsar. The orbital radius

of this system was found to be shrinking in size, which

could be attributed to the emission of gravitational

waves exactly as predicted by detailed calculations

using general relativity. This was convincing evidence

that gravitational waves were being emitted by the

binary neutron star, but the waves were still not directly

detected. That had to wait until the twin Advanced

LIGO detectors at Livingstone, Louisiana and Hanford

in Washington State went into operation in 2015. The

ﬁrst direct detection of gravitational waves was made on

September 14, 2015, and after very careful conﬁrmation,

an announcement about the detection was made on

February 1, 2016. That was a century after Einstein

predicted their existence in 1916, and 27 years after the

ﬁrst proposal to set up the LIGO detectors was made in

1989.

A detailed and yet widely accessible discussion

on gravitational waves and electromagnetic radiation

can be found in the book Gravitational Waves: A New

Window to the Universe, by Ajit Kembhavi & Pushpa

Khare, Springer 2020. A more technical and yet very

lucid discussion is in the book Gravity, An Introduction

to Einstein’s General Relativity by James B. Hartle,

2002, Pearson Education.

1.2 Gravitational Field Equations

We have seen over the Black Hole Stories that

in Albert Einstein’s general theory of relativity, the

eﬀects of the gravitational ﬁeld are described by the

curvature of 4-dimensional space-time. All information

about the space-time structure is contained in the metric

tensor, which is also simply called the metric. The

metric determines the distance between any two points

of the space-time and the passage of time between two

events. In BHS-5 to BHS-8 we have described how

1.3 Weak Field Solutions

the trajectories of particles and photons are determined,

once the metric is known for the special cases of the

Schwarzschild and Kerr solutions. But how does

one determine the metric in the general case, given

a distribution of matter and energy? The equations

for the metric are Einstein’s ﬁeld equation of general

relativity:

It is not our intention here to describe the equations,

we only wish to set the context for gravitational wave

solutions. The left hand side of the equations is made up

of the metric and its derivatives in various combinations,

while the right hand side, known as the stress-energy

tensor, describes the distribution of matter and energy

which acts as the source of the gravitational ﬁled. Given

this quantity, and suitable boundary conditions, the

equations can be solved to obtain the metric. Given the

complexity of the equations, it is impossible to obtain

the solutions in any general way. We have to look for

special cases which are physically important and for

which solutions can be obtained, like the Schwarzschild

and Kerr solutions.

Here we are concerned with the special case of

very weak gravitational ﬁelds.

1.3 Weak Field Solutions

In the complete absence of matter or energy space-

time can be considered to be ﬂat, and it takes the simple

form which it has in the special theory of relativity. We

can now consider a situation in which the gravitational

ﬁeld is very weak, for example when we are very far

from the distribution of matter. In this case the metric

can be considered to be made up of two parts, the

ﬂat part of special relativity and an additional term

which has small magnitude and is known as a metric

perturbation. It can be shown that the equations for the

metric perturbation are given by

where

is the metric perturbation term. The

magnitude

| <<

1, so that all terms which are

higher powers of

can be neglected, and an extra

constraint known as a gauge condition is applied on the

perturbations.

It is apparent that these equations are very similar

to the wave equations of electromagnetic theory in BHS-

14. This is a very important result, as it shows that the

gravitational ﬁeld propagates as waves, at least when

the ﬁeld is weak and far from all sources of gravitation.

The presence of a gravitational wave leads to a periodic

change in the metric at a given point, and the changes

also propagate forward as time passes. As in other kinds

of waves, gravitational waves too have a wavelength

and frequency, with the wavelength typically being

greater than a few km and can be as large as

km.

The appearance of the speed of light c in the above

equation means that the waves propagate with the speed

of light. It can be shown that the waves are transverse,

in the sense that the displacement produced by the

passage of a gravitational wave is limited to the plane

perpendicular to the direction in which the waves are

propagating. Gravitational waves are also polarised in

the same plane.

The periodic changes in the metric which we

are interpreting as gravitational waves have another

connotation in general relativity. The metric determines

the geometry of space-time, and therefore these waves

are often described as ripples in the fabric of space-time.

As a gravitational wave moves across a small region

of space, the distances between all the particle in the

region, and the rate of the passage of time as measured

by clocks, change periodically due to the oscillation of

the gravitational wave.

To see the eﬀect of a gravitational wave, consider

tiny particles which are arranged in a circle in space, as

shown in Figure 1. Now suppose a gravitational wave

passes in a direction perpendicular to the circle into the

plane of the paper. The change in the geometry produced

by the wave is such that there is alternately expansion

and contraction of distances along the horizontal and

vertical directions in the plane of the paper: when there

is contraction along the horizontal as shown in the

ﬁgure, there is expansion along the vertical direction.

As time passes, and the gravitational wave propagates

forward, the directions of contraction and expansion

are exchanged. For particles which are not along the

horizontal or vertical directions, whether expansion

1.4 Similarities and Diﬀerences Between Electromagnetic and Gravitational Waves

or contraction takes place depends on the direction in

which they are located. The net eﬀect is that the circle

of particles changes shape periodically: it alternately

becomes an ellipse ﬁrst stretched along the x-direction

and then stretched along the y-direction, as shown in

the upper part of Figure 1.

The periodic change of shape can occur in a

somewhat diﬀerent way, with the maximum stretching

taking place in a direction which is inclined to the

x- and y- directions, as shown in the lower part of

the ﬁgure. This occurs because of diﬀerence in the

polarisation of the gravitational waves. If a gravitational

wave approaches the circle in a direction which is not

perpendicular to the plane of the circle, or there is more

than one gravitational wave passing through at the same

time, then the eﬀects will be similar to those described,

with the details depending on the situation.

Figure 1: The eﬀect of a gravitational wave on

a set of free point masses arranged in a circle. The

gravitational wave travels into the plane of the paper, in

a direction perpendicular to the plane. The upper and

lower parts of the ﬁgure correspond to two polarisation

states. Image Credit: Kaushal Sharma.

1.4 Similarities and Diﬀerences

Between Electromagnetic and

Gravitational Waves

The equations for electromagnetic waves follow

from Maxwell’s equations while the equations from

gravitational waves follow from Einstein’s equations.

In both cases, the waves travel with the speed of light,

they are transverse waves and are polarised. In the case

of electromagnetic waves we have learnt through the

black hole stories that they can be redshifted, and can

be bent from their straight path, leading to gravitational

lensing. Gravitational waves too can undergo these

eﬀects.

A major diﬀerence between the theories of

gravitational waves and electromagnetic waves is the

following: The Electromagnetic wave equations are

derived from Maxwell’s equations, with no conditions

attached, except the gauge condition. The equations

apply for any distribution of currents and charges, and

the electromagnetic ﬁeld can be large. The gravitational

wave equations, on the other hand, apply only when

the gravitational ﬁeld is weak and acts as perturbation

over ﬂat space-time. In extreme circumstances like the

merger of black holes the ﬁelds can be very large and

then the full Einstein equations have to be used. But that

is impossible to do analytically, so numerical methods

are needed to study the processes. Another diﬀerence

is that while the wavelength of gravitational waves is in

the range of a few km to about

km, electromagnetic

waves have much shorter wavelengths, in the range of

−12

m to a few kilometres. Yet another diﬀerence,

which we simply mention here for completeness, is that

the electromagnetic waves are mainly dipolar in nature

while the gravitational waves are mainly quadrupolar.

As electromagnetic wave propagate through space

they can be absorbed by intervening matter, with the

extent of absorption depending on the wavelength of

the radiation and the density, temperature, chemical

composition and other properties of the matter. That can

change the appearance of an object emitting the waves,

and in extreme cases the waves can be completely

absorbed at some wavelengths so that the emitter

becomes unobservable at those wavelengths. An

example of such an absorbing medium is the Earth’s

atmosphere. While visible light is able to penetrate the

atmosphere, electromagnetic waves at ultraviolet and X-

ray wavelengths are absorbed almost completely. The

situation is quite diﬀerent with gravitational waves,

which remain unaﬀected by passage through even

the most dense environments. This is due to the

fact that the gravitational force is extremely weak,

and the gravitational waves therefore do not interact

signiﬁcantly with matter. We are therefore in principle

able to observe gravitational waves from the most distant

sources and extreme conditions, like the very early

1.4 Similarities and Diﬀerences Between Electromagnetic and Gravitational Waves

epochs of the Universe and the very dense matter in the

environment of merging neutron stars. Our ability to

detect gravitational waves has therefore opened a new

window to the Universe, giving rise to multimessenger

astronomy.

Next Story: In the next story we will consider the

generation of gravitational waves.

About the Author

Professor Ajit Kembhavi is an emeritus

Professor at Inter University Centre for Astronomy

and Astrophysics and is also the Principal Investigator

of the Pune Knowledge Cluster. He was the former

director of Inter University Centre for Astronomy and

Astrophysics (IUCAA), Pune, and the International

Astronomical Union vice president. In collaboration

with IUCAA, he pioneered astronomy outreach

activities from the late 80s to promote astronomy

research in Indian universities.

X-ray Astronomy: Through Missions

by Aromal P

airis4D, Vol.3, No.1, 2025

www.airis4d.com

2.1 Plus Virtutis: Golden Age Begins

Until now, we discussed the advances in X-ray

astronomy until the 1980s. The early discoveries not

only revealed how fascinating it is to study X-rays, but

also highlighted the necessity of powerful instruments

to study the exotic phenomena in our universe. The

1990s marked the dawn of the golden age of X-ray

astronomy, a transformative era during which massive

satellites equipped with powerful instruments were

launched into space. The groundbreaking results from

these missions paved the way for future generations to

deepen their understanding of X-rays emitting sources.

In this section, we will be discussing about this golden

age.

2.2 Satellites in 1990s

2.2.1 Granat

Granat, a USSR mission in collaboration with

other European countries, was launched on December

1, 1989. It started as a pointing observation satellite

with a high elliptical orbit of perigee and apogee of

2,000 km and 200,000 km, respectively, with an orbital

inclination of

51.9

○

, providing a more extended orbit

period of 98 hours and longer uninterrupted observation

time. Over time, the satellite attained a nearly circular

orbit with a perigee and apogee of 60,000 km and

140,000 km, respectively, and later on, went into survey

mode in September 1994. The satellite was operational

until November 1998. This satellite had an energy

coverage of 2 keV - 2 MeV, from X-rays to Gamma rays.

Figure 1: MWPC in Granat.

Credit:Sunyaev et al.

There were seven scientiﬁc instruments onboard

Granat to study X-rays and Gamma-rays :

Astronomical X-ray telescope (ART-P) is an

imaging telescope with Position Sensitive

Multiwire Proportional Counters (MWPCs) and

coded apertures that produce images of selected

areas of the sky with a better angular resolution

of six arc minutes in the energy range 4-60 keV.

The aperture is coded by Uniformly Redundant

Array (URA) coupled with a collimator. It can

also used for timing and spectral studies in the 4-

100 keV energy range. Granat had four identical

modules of the ART-P telescope.

Astronomical X-ray telescope-spectrometer

(ART-S) is a system dedicated to the

spectroscopic study of X-ray sources in energy

range 3-100 keV. It consists of 4 spectroscopic

MWPCs. Both ART-P and ART-S were

developed in Moscow by IKI.

SIGMA is a Gamma-ray telescope

developed by two French laboratories -

the Serviced’Astrophysique at Saclay and the

Centre d’Etude Spatiale des Rayonnements at

Toulouse. It is the primary scientiﬁc instrument

2.2 Satellites in 1990s

Figure 2: SIGMA telescope in Granat.

Credit:Paul et al.

in Granat and the ﬁrst coded-aperture telescope

sensitive to radiation in the energy range 35 keV

- 1.3 MeV to be operated in space. The satellite

has a Position-sensitive Scintillation Detector

(PSD), active and passive shielding devices, and

necessary modules. The Coded aperture enables

imaging capability with a resolution of 13 arc

minutes. This instrument is also coded by URA.

WATCH is an all-sky monitor developed by the

Danish Space Research Institute that works in

an energy range of 6-180 keV. It has a rotating

modulation collimator system with Phoswich

detectors consisting of parallel NaI (TI) and CsI

(TI) strips for observing bright X-ray sources

within 0.5

○

PHEBUS is a Gamma-ray instrument developed

by Centre d’Etude Spatiale des Rayonnements at

Toulouse that can cover an energy range of 100

keV - 100 MeV. This all-sky coverage instrument

consists of two independent bismuth germanate

(BGO) detectors surrounded by a plastic anti-

coincidence jacket.

KONUS-B is a Gamma-ray instrument developed

by Ioﬀe Physico-Technical Institute in St.

Petersburg to carry out Gamma-ray burst studies

using traditional scintillation technique in the

energy range of 0.01-8 MeV. The detectors have

NaI (TI) scintillation crystals. This instrument

also had an all-sky coverage and operated for

nearly 27 days, detecting around 19 Gamma-ray

bursts and a couple of solar ﬂares.

TOURNESOL observe optical counterpart of

high energy events. It consisted of four

proportional counters that can detect in the energy

range of 2 keV - 20 MeV and two optical detectors

with a large FOV of 5

○

×5

○

In the initial four years, from the pointed observations it

studied many galactic (including stellar black holes,

neutron stars, and white dwarfs) and extragalactic

sources through timing, spectral and imaging

capabilities. In the survey mode, it discovered many

new X-ray candidate sources. Granat ﬁrst discovered

the electron-positron annihilation lines from the galactic

micro-quasar 1E1740-294. It also made a very deep

imaging of the galactic center region.

2.2.2 ROSAT

The Roentgen Satellite (ROSAT) was a

collaboration between Germany, the United States, and

the United Kingdom. It was launched on June 1990, into

a nearly circular orbit 550 km radius with an inclination

○

and remained active and operational for 9 years,

till February 1999. Its initial 6 months were dedicated

to all-sky surveys, and later, it started to carry out

pointed observations. ROSAT was 1000 times better

than the UHURU satellite in terms of X-ray sensitivity.

Satellite and X-ray telescopes were built and operated

by Germany; the UV telescope was built by the UK,

and NASA provided the launch vehicle.

ROSAT consisted of two scientiﬁc instruments:

X-ray Telescope (XRT): A Wolter-I X-ray

telescope that operates in the soft X-ray range

(0.1-2.4 keV). It has four nested Wolter type-1

mirrors to concentrate the X-ray photons into

the detector. The focal plane assembly consisted

of two Position Sensitive Proportional Counter

(PSPC) of

○

FOV and one High-Resolution

Imager (HRI) of

′

×36

′

FOV that acts as the

detector.

Wide Field Camera (WFC): It is an Extreme

Ultraviolet (EUV) imaging telescope that

operates in 0.06-0.2 keV with an FOV of

○

It consists of three nested Wolter-Schwarzschild

2.2 Satellites in 1990s

Figure 3: ROSAT Satellite.

Credit:NASA

type-I mirrors and two channel plate detectors.

ROSAT was very successful in demonstrating

technological capabilities and created signiﬁcant

scientiﬁc discoveries. It performed the ﬁrst all-sky

survey with an imaging telescope and found more than

150,000 and nearly 500 sources in the X-ray and EUV

ranges, respectively. It also discovered the ﬁrst isolated

neutron star and was the ﬁrst-ever satellite to study the

X-ray emission from a comet. It studied the collision of

Shoemaker-Levy with Jupiter. It also detected pulsation

from Geminga and diﬀused X-ray emission by molecular

clouds.

2.2.3 BBXRT

Broad Band X-ray Telescope (BBXRT) is the

ﬁrst imaging telescope that works in 0.3-12 keV. It is

a part of the ASTRO-1 mission payload conducted

by space shuttle Columbia in December 1990. It

was operational for around 9 days. The mission’s

main aim was to conduct high-sensitivity, moderate-

resolution spectrophotometry in the given energy range.

It had some diﬃculties in pointed observation due to

improperly compensated gyro drift rate, even though it

successfully observed 180,000 seconds on cosmic X-ray

sources. BBXRT mainly consisted of two co-aligned

conical X-ray imaging mirrors with a cryogenically-

cooled Si(Li) non-dispersive spectrometer at its focus.

Despite the diﬃculties, BBXRT resolved the iron K

line in the binaries - Centarus X-3 and Cygnus X-2 and

detected evidence of line broadening in NGC 4151. It

also provided insight into cooling ﬂow in clusters of

galaxies.

Figure 4: Conceptual view of the BBXRT experiment

package.

Credit:Petre et.al

Figure 5: ASCA Satellite.

Credit:JAXA

2.2.4 ASCA

Advanced Satellite for Cosmology and

Astrophysics (ASCA), formerly ASTRO-D, is

the fourth X-ray astronomical satellite developed by

the Institute of Space and Astronautical Science. The

satellite was launched in February 1993 into a

31.1

○

inclined nearly circular orbit of perigee and apogee

of 523 km and 615 km, respectively. It was the ﬁrst

satellite that used Charge Coupled Devices for X-ray

astronomy.

ASCA had two scientiﬁc payloads:

X-ray Telescope (XRT): It consists of four

gazing incident telescopes attached to an imaging

spectrometer in its focal plane. The telescope

utilizes multi-nested thin foil conical optics

developed by Serlemitos. It has an energy

bandwidth of 0.5 - 12 keV. XRT can resolve

sources separated by 1’

Focal Plane Detectors consists of two CCD

cameras - Solid-state Imaging Spectrometer

(SIS); and two gas scintillation imaging

proportional counters - Gas Imaging

Spectrometer (GIS). SIS works in the energy

band-pass of 0.4-12 keV with a square FOV of

′

×20

′

. GIS on the other hand, works in the

energy range 0.8-12 keV with a circular FOV of

′

×50

′

So far we discussed X-ray satellites from the ﬁrst

half of the 1990s and many interesting missions are yet

come and we will discuss it on the upcoming article.

2.2 Satellites in 1990s

References

Santangelo, Andrea and Madonia, Rosalia and

Piraino, Santina A Chronological History of X-

Ray Astronomy Missions.Handbook of X-ray and

Gamma-ray Astrophysics. ISBN 9789811645440

Broad Band X-ray Telescope-I

Broad Band X-ray Telescope-II

Granat X-ray and Gamma-ray Observatory-I

Granat X-ray and Gamma-ray Observatory-II

The ROSAT Mission

Petre, R and Serlemitsos, P. J. The Broad Band

X-Ray Telescope.Space Science Reviews.

Tanaka, Y., Inoue, H., & Holt, S. S The X-ray

astronomy satellite ASCA PASJ: Publications of

the Astronomical Society of Japan (ISSN 0004-

6264), vol. 46, no. 3, p. L37-L41

S. V. Golenetskii, R. L. Aptekar, I. V. Dementyev,

Yu. A. Guryan, V. N. Ilyinsku, E. P. Mazets, V.

N. Panov, Z. Ya. Sokolova, D. D. Frederiks, T.

V. Kharitonova and L. 0. Sheshin GAMMA-RAY

BURST OBSERVATIONS WITH KONUS-B ON

GRANAT SPACECRAFT Adv. Space Re,,. VoL

11, No.8, pp. (8)125—(8)128, 1991

J. Pau1, P. Mandrou, J. Ba11et,M. Cantin, J. P.

Chabaud, B. Cordier, M. Ehanno, A. Goldwurm,

A. Lambert, J. Land

e, P. Laurent, F. Lebrun, J. P.

Leray, B. M

ena, M. Niel,J. P. Roques, G. Rouaix,

L. Salotti, P. Souleile and G. Vedrenne SIGMA:

THE HARD X-RAY AND SOFT GAMMA-RAY

TELESCOPE ON BOARD THE GRANAT SPACE

OBSERVATORY Adv. Space Res.Vol. 11, No.8,

pp. (8)289-(8)302, 1991

R. A. Sunyaev, S. I. Babichenko, D. A. Goganov,

S. R. Tabaldyev and N. S. Jambourenko X-RAY

TELESCOPES ART-P AND ART-S FOR THE

GRANAT PROJECT Adv. Space Res. Vol. 10,

No. 2, pp. (2)233—(2)237, 1990

About the Author

Aromal P is a research scholar in

Department of Astronomy, Astrophysics and Space

Engineering (DAASE) in Indian Institute of

Technology Indore. His research mainly focuses on

studies of Thermonuclear X-ray Bursts on Neutron star

surface and its interaction with the Accretion disk and

Corona.

What are Galaxies?

by Linn Abraham

airis4D, Vol.3, No.1, 2025

www.airis4d.com

3.1 Interstellar dust

The Milky Way is the distribution of stars in a disc

often visible using the naked eye from suitable places

on the Earth. In a previous article, we had seen that

until Hubble’s discovery in the 1920s, it was believed

that the whole of the observable universe was contained

in our Milky Way, an island ﬂoating in empty inﬁnite

space. After which it was understood that those fuzzy

objects seen in the night sky were worlds distant from

ours and were replicas of our own galaxy. Initially

they were classiﬁed together with what we today call

as nebulae. These nebulae or dark patches like the

famous Horsehead Nebula are actually regions of dust

concentration.The dark patches often seen within our

galaxy are due to this dust and not the absence of stars as

was thought to be the case earlier. Interstellar dust may

contain graphite, silicates and solid hydrogen. Dust

reduces the intensity of star light through absorption

and scattering.

Figure 1: Image showing the most prominent dark

lane which stretches from the constellations Cygnus to

Sagittarius called the Great Rift or the Dark Rift.

Image Courtesy: NASA

Figure 2: Rotation curve of a spiral galaxy. The dotted

curve shows how the rotation curve should have sloped

down had the entire mass of the galaxy been conﬁned

to its visible boundary.

Image Courtesy: J. V. Narlikar, Introduction to Cosmology.

3.2 Rotation and Presence of Gas

The diameter of the disc of the Milky Way is

estimated to be

∼

30 kpc and its thickness

∼

1 kpc. The

galaxy rotates about it’s polar axis, however not like a

rigid body. Figure 2 shows a typical rotation curve. At

a distance r from the center O of the galaxy, a circular

Keplerian orbit will have a velocity given by,

v =

√

GM(r)

Where M(r) is the mass of the galaxy contained within

the radius r and the point A represents the visible

extent of the galaxy. The dotted lines shows how the

velocity should have decreased beyond A. However,

what we see in practice is the velocity remaining more

or less constant upto two or three times the visible

boundary. Some stars like our Sun go around the

3.3 Spirals and Ellipticals

Galactic Center, whereas some have highly eccentric

orbits that takes them out of the galactic plane. The

former type are called population-I stars and the latter

are called population-II stars. The population-II stars

are seen to be older than population-I stars. The mass

of our Galaxy is estimated at

∼1.4 ×10

solar mass.

The absorption lines shown by the spectra of stars in

our galaxy show that abosrbing gases are present in the

interstellar medium. The emission nebulae consists of

gas that absorbs the UV radiation from stars and radiate

it as visible light in spectacular colours. There are also

hot regions near stars which contain hydrogen as H II,

ionized by the UV liradiation from stars. In contrast

there are also cool regions containing atomic hydrogen

HI. These were detected using the 21 cm observation

in radio astronomy.

3.3 Spirals and Ellipticals

Galaxies comes in two major types, spirals and

ellipticals. Spirals are probably the most numerous

amongst bright galaxies. They show rotation and

ﬂattening into a disc like our own galaxy. The presence

of a central bulge and dark lanes of abosrbing matter.

While spirals are most numerous among bright galaxies,

the most numerous of all galaxies are the ellipticals.

They are ellipsoidal in shape, show very little rotation

and have very little gas and dust. Let us revise the

modern classiﬁcation scheme based on the morphology

of galaxies that is primary due to Hubble. The spirals

are arranged in a sequence called Sa, Sb, Sc etc. in

decreasing order of the importance of the central nucleus

or bulge in relation to the surrounding disc. Along the

sequence, the central spheroid decreases in luminosity

and spiral arms become more loosely wound. Spirals

can have bars in the central region and when that

happens they are designated SBa, SBb, SBc, etc. The

ellipticals are arranged in a sequence E0, E1, ...E7 based

on the progressive ﬂattening of the proﬁles. E0s have

an almost spherical shape and E7s have a lenticular

shape. The distribution of light across galaxies are

conveniently described in terms of contours of equal

intensity called isophotes. Another type of galaxy called

S0, is intermediate between the ellipticals and spirals.

(a) The elliptical galaxy ESO 325-G004 part of the

Abell S740 cluster.

Image Courtesy: NASA, ESA, and The Hubble Heritage Team.

(b) Spiral galaxy NGC 1566

Image Courtesy: Dark Energy Survey

Figure 3: Types of galaxies

They have little gas and dust like the ellipticals but their

isophotes are similar to those of spirals. They may

have formed from collisions of spirals and ellipticals.

Since the stars are widely spaced they may pass through

without problems after the collision but interstellar gas

and dust might be thrown out into the intergalactic space.

The isophotes in this case which arise from starlight

alone might remain intact. While this arrangement of

galaxies often called the Hubble tuning fork seems to

have some order, it is still not known if this represents

some evolutionary sequence or not. In addition to these

there exists another class called ‘irregular’ which show

3.4 Structures on a Larger Scale

Figure 4: Image shows a massive group of galaxies: a

cluster named RXC J0032.1+1808.

Image Courtesy: ESA/Hubble

no particular pattern. Until our knowledge in galaxy

structure and evolution improves signifcantly these are

to be seen as empirical models that can help us better

understand the underlying physics. Although we can

guess the spiral nature of our galaxy from the presence

of the disk, what more can be said about it, having the

disadvantage of viewing the galaxy from within.

3.4 Structures on a Larger Scale

Galaxies themselves are seen to form patterns on a

larger scale. Galaxy groups may contain a few to a few

tens of galaxies whereas clusters can contain several

hundreds of galaxies. Galaxies that do not belong to

such groups or clusters are called ﬁeld galaxies. Our

galaxy belongs to a group of

∼28

galaxies, known as the

Local Group and are separated by distance of up to

∼1

Mpc. Our closest neighbours are the Large and Small

Magellanic Clouds at an approximate distance of 50 kpc.

Astronomers have also discovered structures on a much

larger scale than clusters. This is done by studying the

distribution of clusters across the sky and looking for

any grouping or clumping. Larger structures on a scale

∼50

Mpc compared to cluster scales of

∼5

Mpc

have been discovered. These are called superclusters.

Our galaxy belongs to one such supercluster known as

the Local Supercluster.

References

[Narlikar(1993)]

Jayant Vishnu Narlikar. Introduction

to Cosmology. Cambridge University Press,

Cambridge [England] ; New York, NY, USA, 2nd

ed edition, 1993. ISBN 978-0-521-41250-6 978-0-

521-42352-6.

[Carroll and Ostlie(2014)]

Bradley W. Carroll and

Dale A. Ostlie. An Introduction to Modern

Astrophysics. Pearson, Harlow, 2. ed., pearson

new internat. ed edition, 2014. ISBN 978-1-292-

02293-2.

About the Author

Linn Abraham is a researcher in Physics,

specializing in A.I. applications to astronomy. He is

currently involved in the development of CNN based

Computer Vision tools for prediction of solar ﬂares

from images of the Sun, morphological classiﬁcations

of galaxies from optical images surveys and radio

galaxy source extraction from radio observations.

Corona Borealis: The Northern Crown

by Sindhu G

airis4D, Vol.3, No.1, 2025

www.airis4d.com

Figure 1: Corona Borealis. (Image Credit: seasky.org)

4.1 Introduction

Corona Borealis (Figure: 1), also known as ”The

Northern Crown,” is a small but striking constellation

located in the northern celestial hemisphere. Its

name, derived from Latin, aptly describes its semi-

circular arrangement of stars that resemble a crown.

Positioned between the constellations Bo

otes and

Hercules, Corona Borealis is rich in mythological

history, scientiﬁc interest, and astronomical phenomena.

Despite its modest size, it has fascinated stargazers and

astronomers for centuries. This article delves into

the mythology, astronomy, cultural signiﬁcance, and

scientiﬁc relevance of Corona Borealis, oﬀering an

in-depth exploration of its allure.

Figure 2: IAU Corona Borealis chart. (Image Credit:

IAU and Sky & Telescope magazine (Roger Sinnott &

Rick Fienberg))

4.2 Mythological Origins

The origins of Corona Borealis trace back to

ancient Greek mythology. The constellation is most

famously associated with the tale of Ariadne, the

daughter of King Minos of Crete. According to the

myth, Ariadne played a crucial role in helping the hero

Theseus navigate the Labyrinth to defeat the Minotaur.

Afterward, Theseus abandoned her on the island of

Naxos, where she was discovered by the god Dionysus.

As a gesture of love and devotion, Dionysus placed a

jeweled crown in the heavens, creating the constellation

Corona Borealis.

Other cultures also recognized this constellation,

attributing their own stories to its formation. For

example, the Arabs referred to it as ”Alphecca,” a name

derived from its brightest star, symbolizing a ”broken

4.3 Historical Astronomical References

ring.” Native American tribes, such as the Lakota, saw

it as a campﬁre circle, reﬂecting its unique shape. These

diverse interpretations highlight the universal appeal of

Corona Borealis across diﬀerent civilizations.

4.3 Historical Astronomical

References

6th century BC: First written account by Greek

philosopher Pherecydes.

3rd century BC: Mentioned by Greek poet Aratus.

2nd century AD: Cataloged by Ptolemy.

Ancient Arab astronomers called it al-Fakkan

(meaning ”bowl”)

4.4 Astronomical Features

Though small, Corona Borealis oﬀers intriguing

astronomical features that captivate scientists and

amateur astronomers alike. Spanning an area of

approximately 179 square degrees, it is the 73rd largest

constellation in the night sky. Its brightest stars and

deep-sky objects make it a fascinating subject for

observation and study.

1. Bright Stars:

Alphecca (

Coronae Borealis): Also

known as Gemma, this is the brightest

star in the constellation, with an apparent

magnitude of 2.2. Alphecca is a binary star

system located about 75 light-years away

from Earth. The primary star is a white

main-sequence star, while its companion is

a dimmer red dwarf.

Nusakan (

Coronae Borealis): The

second brightest star in the constellation,

Nusakan is also a binary system situated

approximately 114 light-years from Earth.

It consists of a chemically peculiar primary

star and a smaller companion.

2. Variable Stars:

Corona Borealis contains several variable stars,

including the notable R Coronae Borealis and T

Coronae Borealis.

R Coronae Borealis (R CrB): R Coronae

Borealis stars are a rare type of yellow

supergiant. They experience sudden,

unpredictable dimming, believed to be

caused by carbon soot clouds in their

atmospheres. Typically visible with

binoculars, these stars can dim dramatically,

sometimes to a magnitude of 15. Studying

them has provided valuable insights into

the later stages of stellar evolution and the

formation of carbon dust.

T Coronae Borealis (The Blaze Star): T

Coronae Borealis is a recurrent nova located

3,000 light-years away. It is a binary system

composed of a white dwarf and a red giant.

Approximately every 80 years, it undergoes

thermonuclear explosions. Major outbursts

were recorded in 1866 and 1946, with

the next expected in September 2024.

This system remains of great interest to

astronomers due to its dramatic and periodic

behavior.

3. Deep-Sky Objects:

Though not densely packed with deep-sky objects,

Corona Borealis hosts the Corona Borealis

Galaxy Cluster, a massive cluster of galaxies

located more than a billion light-years from Earth.

This cluster is of particular interest to astronomers

studying the large-scale structure of the universe.

4. Exoplanet Discoveries:

The constellation has also contributed to the

discovery of exoplanets. Stars within Corona

Borealis, such as Alphecca, have been studied for

planetary systems, expanding our understanding

of planetary formation and habitability.

4.5 Observing Corona Borealis

Corona Borealis is best observed in the northern

hemisphere during late spring and summer months,

when it is prominently visible in the evening sky. Its

semi-circular arrangement of stars makes it easy to

identify, lying near the bright star Arcturus in the

constellation Bo

otes. The seven main stars of Corona

4.6 Cultural Signiﬁcance

Borealis form an arc, creating the appearance of a

crown.

Finding Corona Borealis:

Begin by locating the bright star Arcturus in

otes. From there, look slightly eastward to

ﬁnd the curved pattern of stars that form Corona

Borealis.

For southern hemisphere observers, the

constellation appears lower in the sky and is

visible during the winter months.

Tips for Stargazers:

Use binoculars or a telescope to explore the

binary star systems and faint galaxies within the

constellation.

On clear nights, look for the subtle variations in

brightness of variable stars such as R Coronae

Borealis.

Best Viewing Conditions: To fully appreciate

Corona Borealis, ﬁnd a location with minimal

light pollution on a clear, moonless night. Under

these conditions, the arc of stars forming its crown

shape becomes more distinct. After locating

otes, Corona Borealis can be easily spotted

nearby, oﬀering a stunning view of this small yet

captivating constellation.

4.6 Cultural Signiﬁcance

Throughout history, Corona Borealis has held

signiﬁcant cultural and symbolic meanings. Its

striking shape made it a prominent feature in various

mythologies and traditions. In addition to its association

with Ariadne in Greek mythology, other cultures oﬀered

unique interpretations:

Arabic Astronomy: In Arabic tradition, the

constellation was called ”Alphecca,” symbolizing

a broken ring or a string of pearls.

Chinese Astronomy: In Chinese star lore, Corona

Borealis was part of the ”Heavenly Market

Enclosure,” representing a celestial marketplace.

Native American Tribes: Among some Native

American tribes, the constellation was seen as a

campﬁre circle, a gathering place for people.

These interpretations underscore the universal

connection between humanity and the stars, as diverse

cultures found meaning in the constellation’s distinctive

arc.

4.7 Scientiﬁc Contributions

Corona Borealis has contributed to signiﬁcant

advancements in astronomy and astrophysics. The

study of its variable stars, particularly R Coronae

Borealis, has provided insights into stellar evolution

and the behavior of rare star types. Additionally, the

Corona Borealis Galaxy Cluster has been instrumental

in understanding the large-scale structure of the universe

and the distribution of dark matter.

Variable Stars and Stellar Evolution:

R Coronae Borealis stars are a rare class of stars

that experience unpredictable dimming. Studying these

stars has helped astronomers understand the late stages

of stellar evolution and the formation of carbon dust.

Galaxy Clusters and Dark Matter:

The Corona Borealis Galaxy Cluster has been a

valuable resource for studying the distribution of dark

matter. Observations of gravitational lensing and galaxy

movements within the cluster provide clues about the

unseen matter that makes up a signiﬁcant portion of the

universe.

Exoplanetary Studies:

The discovery of exoplanets within Corona

Borealis has expanded our knowledge of planetary

systems. Detailed observations of stars in this

constellation have revealed new worlds, some of which

are in the habitable zone, sparking curiosity about the

potential for life beyond Earth.

4.8 Astrological Signiﬁcance

Although primarily known for its scientiﬁc and

mythological importance, Corona Borealis also holds

a place in astrology. Astrologers often associate the

constellation with creativity, leadership, and the pursuit

of goals. Its semi-circular shape, reminiscent of a

crown, symbolizes triumph and achievement. While

astrology is distinct from astronomy, the constellation’s

4.9 Conclusion

symbolic representation continues to inspire those who

look to the stars for guidance.

4.9 Conclusion

Corona Borealis, the Northern Crown, is a

constellation that beautifully combines mythology,

science, and cultural signiﬁcance. Its graceful arc of

stars has inspired stories across cultures and continues

to captivate stargazers and astronomers alike. Whether

viewed through the lens of ancient myths, cutting-edge

science, or personal symbolism, Corona Borealis serves

as a shining example of humanity’s enduring fascination

with the night sky. Small yet signiﬁcant, it remains a

celestial crown adorning the heavens, a testament to the

wonders of the universe.

References:

Corona Borealis

Rare T Coronae Borealis nova explosion this

September: What it is and how to see it

NASA, Global Astronomers Await Rare Nova

Explosion

T Coronae Borealis Nova Explosion 2024: Has

T CrB Nova Happened Yet?

Corona Borealis (The Northern Crown)

Constellation

Corona Borealis Constellation

Corona Borealis

A Blaze star or ‘Nova’ in Corona Borealis

A Guide to the Corona Borealis Constellation

and Its Stars

Corona Borealis

About the Author

Sindhu G is a research scholar in Physics

doing research in Astronomy & Astrophysics. Her

research mainly focuses on classiﬁcation of variable

stars using diﬀerent machine learning algorithms. She

is also doing the period prediction of diﬀerent types

of variable stars, especially eclipsing binaries and on

the study of optical counterparts of X-ray binaries.

Part III

Biosciences

A Short Note on the Discovery of MicroRNA

and Its Role in Post-Transcriptional Gene

Regulation

by Geetha Paul

airis4D, Vol.3, No.1, 2025

www.airis4d.com

The 2024 Nobel Prize in Physiology or Medicine

has been awarded to Victor Ambros and Gary Ruvkun

for their groundbreaking discovery of microRNA, a

fundamental mechanism of gene regulation.

1.1 Introduction

The evolution of multicellular organisms from

unicellular ancestors involved the development of

specialised cell types, each acquiring distinct functions.

This transition required increasingly sophisticated gene

regulation mechanisms. Initially, gene regulation was

primarily transcriptional, mediated by DNA-binding

factors that acted on regulatory sequences. However, as

organisms became more complex, additional regulatory

systems emerged. One signiﬁcant development was the

expansion of genes encoding microRNAs (miRNAs)

within the genomes of multicellular organisms.

These tiny, non-coding RNA molecules exert post-

transcriptional control over mRNA stability and protein

translation. The groundbreaking discovery of miRNAs

by Victor Ambros and Gary Ruvkun in 1993 revealed a

novel class of regulatory RNA. They studied mutant C.

elegans nematodes with developmental defects linked to

the lin-4 and lin-14 genetic loci. Ambros’s lab found that

the lin-4 gene produced a short 22-nucleotide RNA that

regulated lin-14 by binding to its 3’ untranslated region

(3’ UTR). This discovery led to the identiﬁcation of

other miRNAs, such as let-7, which are highly conserved

across animal species, including humans. Researchers

recognised that miRNAs could regulate vast networks

of protein-coding genes, playing critical roles in animal

development and adult tissue function. Over time, it

became clear that miRNAs are essential for normal

cellular processes and that their dysregulation could

contribute to various diseases, including cancer.

1.2 What is microRNA?

microRNA (miRNA) is a small, non-coding RNA

molecule that regulates gene expression. Unlike

messenger RNA (mRNA), which carries instructions

to make proteins, microRNA works by binding to

speciﬁc mRNA molecules, preventing them from

being translated into proteins or degrading the mRNA

altogether. MicroRNAs help ﬁne-tune various cellular

processes like development, growth, and metabolism

by controlling which mRNA gets used to make proteins.

Their role is essential in maintaining normal cell

function, and disruptions in microRNA activity have

been linked to diseases such as cancer.

The human genome is now known to code

for over a thousand microRNAs, highlighting the

widespread importance of this regulatory mechanism.

This discovery has profound implications for our

understanding of cellular diﬀerentiation, development,

and disease processes.

1.3 Posttranscriptionnal Regulation

Figure 1: miRNA Biogenesis

Image courtesy: https://pmc.ncbi.nlm.nih.gov/articles/PMC6298569/

(A) The pathway of miRNA biogenesis, including

the Microprocessor-mediated (DROSHA/DGCR8) step

in the nucleus followed by DICER processing in the

cytoplasm. Structural and sequence features essential

for miRNA processing are highlighted in both the pri-

mir and pre-miR molecules. (B) Positive regulators of

miRNA biogenesis bind to the terminal loop (TL) or

other elements within miRNA precursors (pri-mir and

pre-mir) and stimulate Drosha and/or Dicer processing,

leading to increased levels of mature miRNAs (miR).

The TL, also known as the apical loop, is depicted in red.

or other elements within miRNA precursors (pri-mir and

pre-mir) and abrogate Drosha and/or Dicer processing,

leading to decreased levels of mature miRNAs (miRs).

The TL, also known as the apical loop, is depicted in

red.

1.3 Posttranscriptionnal Regulation

The biogenesis of miRNA implies multiple

steps. Like protein-encoding genes, miRNAs are

transcribed by RNA polymerase II from independent

miRNA-encoding genes or represent introns of protein-

coding genes. Initially, miRNAs are transcribed to

primary miRNAs (pri-miRNAs), and this transcript

is further cleaved by Drosha and DGCR8; (Drosha

is a critical Class 2 ribonuclease III enzyme that

plays a fundamental role in microRNA (miRNA)

biogenesis. It is the primary nuclease responsible

for initiating miRNA processing in the cell nucleus

by cleaving primary miRNA transcripts (pri-miRNAs)

into precursor miRNAs (pre-miRNAs) and DGCR8

is a critical RNA-binding protein that plays an

essential role in microRNA (miRNA) biogenesis. It

functions as a key component of the Microprocessor

complex, working alongside the RNase III enzyme

Drosha to process primary miRNA transcripts into

precursor miRNAs) generating precursor miRNA (pre-

miRNA). Pre-miRNA is exported from the nucleus

to the cytoplasm by exportin-5 (XPO5), where Dicer

processes it into mature miRNA, which forms what is

called the miRNA-induced silencing complex (miRISC)

containing the Argonaute 2 protein and glycine-

tryptophan protein (GW182). The complex is formed

where the miRNA and mRNA target interact.

1.4 Mechanisms of Action

microRNAs bind to complementary sequences

within target mRNAs’ 3[2032?] untranslated regions

(3[2032?] UTR). This binding can lead to two primary

outcomes:

Translational Repression: miRNAs can inhibit the

translation process, preventing mRNA from converting

into protein.

miRNAs can interact with other regions of mRNA,

including the 5[2032?] UTR and coding sequences, and

even inﬂuence gene transcription directly under certain

conditions. This versatility allows miRNAs to ﬁne-tune

gene expression across various cellular contexts.

1.5 Biological Signiﬁcance

The human genome encodes over a thousand

miRNAs, underscoring their widespread importance

in regulating numerous biological processes such as

development, metabolism, and cell diﬀerentiation. For

instance, miRNAs are essential for maintaining normal

cellular functions; disruptions in their activity are linked

to various diseases, including cancer and metabolic

disorders. Moreover, research has shown that a single

miRNA can regulate multiple target genes, while several

1.6 Implications for Medicine

miRNAs may inﬂuence a single gene. This complex

regulatory network allows cells to respond dynamically

to environmental changes and developmental cues.

1.6 Implications for Medicine

The discovery of microRNAs has opened new

avenues for understanding disease mechanisms and

developing therapeutic strategies. For example,

aberrant regulation of miRNAs has been implicated in

cancer progression and other pathological conditions.

Understanding how miRNAs function could lead to

innovative treatments that target these molecules to

restore standard gene expression patterns.

1.7 Mechanisms of miRNA

Involvement in Cancer

microRNAs (miRNAs) play a complex and dual

role in cancer development, functioning as potential

oncogenes and tumour suppressors depending on

cellular context. These small non-coding RNAs

can dramatically inﬂuence cancer characteristics by

modulating key cellular processes such as proliferative

signalling, growth suppression, apoptosis, invasion,

metastasis, and angiogenesis. The mechanisms

of miRNA involvement in cancer involve various

dysregulation pathways, including gene ampliﬁcation,

deletion, abnormal transcriptional control, epigenetic

changes, and defects in miRNA biogenesis machinery.

Notably, miRNAs can have context-speciﬁc eﬀects,

with the same miRNA potentially acting as an oncogene

in one cancer type and a tumour suppressor in another.

Speciﬁc examples highlight this complexity: miR-125b

acts as an oncomiR in most hematologic malignancies

but functions as a tumour suppressor in many solid

tumours.

courtesy: https://pmc.ncbi.nlm.nih.gov/articles/PMC8554338/

Figure 2: Mechanism of Tumour Suppressor

miRNA regulation in cancer.

The mechanism controlling TS miRNAs

occurs at diﬀerent steps in cancer, including

transcriptional and posttranscriptional regulation.

Various posttranscriptional regulation occurs in the

miRNA biogenesis machinery, which aﬀects the

components of miRNA processing. miRNA processing

of some TS miRNAs is regulated at multiple levels: at

the gene level, through genetic mutations in the miRNA

gene and in other components of miRNA processing,

such as Dorsha, Dicer, and TRBP, and the binding site

of miRNA targets and exportin-5. In XPO mutations,

several TS miRNAs are downregulated in cancer, such

as miR-200 and let-7. The miRNA transcripts are

altered at transcription by several oncogenes, such as

Myc and TGFB, or by epigenetic modiﬁcation. LIN

28 proteins speciﬁcally block the processing of the pre-

miRNA of Let-7. Several studies have highlighted that

Dicer downregulation in cancer is caused by several

factors, such as somatic mutations, Tap63, miR-360,

and miR-103/107. EGFR suppresses the maturation of

speciﬁc TS miRNAs through phosphorylation of AGO2

and prevents the association of AGO2 with Dicer.

In chronic lymphocytic leukaemia, reduced miR-

15 and miR-16 expression can inhibit apoptosis, while

overexpression of miR-21, miR-210, and miR-221 in

breast cancer correlates with a worse prognosis.

Epigenetic interactions further complicate

miRNA’s role in cancer. DNA methylation can lead to

reduced expression of suppressor miRNAs, potentially

triggering cancer development. For instance, miRNAs

like miR-34 have been identiﬁed as critical tumour

1.8 Conclusion

suppressors across multiple cancer types, targeting

oncogenic pathways and potentially oﬀering therapeutic

opportunities. The intricate nature of miRNAs in

cancer underscores their signiﬁcant potential as both

diagnostic markers and therapeutic targets, representing

a promising avenue for future cancer research and

treatment strategies.

1.8 Conclusion

The discovery of microRNAs represents a

signiﬁcant leap forward in molecular biology

and genetics. Researchers have gained insights

into fundamental biological processes and disease

mechanisms by elucidating how these small RNA

molecules regulate gene expression post-transcriptional

levels. As ongoing research continues to uncover

the complexities of miRNA function, their potential

medical applications promise exciting advancements in

therapeutic interventions.

References

https://www.thehindu.com/sci-tech/science/wh

y-is-microrna-discovery-a-big-leap/article68745959.

ece

https://academic.oup.com/gpb/article/7/4/147/7

221818?login=false

https://pmc.ncbi.nlm.nih.gov/articles/PMC40132

51/#R81

https://www.indiatoday.in/science/story/what-i

s-microrna-the-discovery-awarded-the-2024-nobel-p

rize-in-medicine-2612568-2024-10-07

https://pmc.ncbi.nlm.nih.gov/articles/PMC85543

38/

About the Author

Geetha Paul is one of the directors of

airis4D. She leads the Biosciences Division. Her

research interests extends from Cell & Molecular

Biology to Environmental Sciences, Odonatology, and

Aquatic Biology.

Part IV

General

Evolution of Domain Generation Algorithms:

From Simple Patterns to AI-Driven

Adaptation

by Jinsu Ann Mathew

airis4D, Vol.3, No.1, 2025

www.airis4d.com

In the previous article, we explored the key

components of Domain Generation Algorithms (DGA),

including seed values, domain generation logic, entropy

sources, and the frequency and volume of generated

domains. These components form the foundation of

how malware dynamically generates domain names to

establish communication with Command and Control

(C2) servers while evading detection and takedown.

However, as cybersecurity defenses improved and static

detection methods became more eﬀective, attackers

were forced to evolve their DGA techniques to maintain

resilience and adaptability in increasingly hostile digital

environments.

This article delves into the evolution of DGAs,

tracing their transformation from early, predictable,

pattern-based algorithms to the sophisticated, highly

dynamic DGAs used in modern malware campaigns. By

examining key milestones, notable malware examples,

and technical advances in DGA strategies, we aim to

highlight how this technology has adapted to outpace

traditional security measures. From simple time-based

algorithms to complex AI-driven domain generation

techniques, the evolution of DGAs is a testament to

the ongoing arms race between cybercriminals and

defenders.

1.1 The Early Days: Simple

Pattern-Based DGAs

In the early days of Domain Generation Algorithms

(DGAs), cybercriminals used relatively simple and

predictable methods to generate domain names. These

early DGAs were primarily based on static and

hardcoded algorithms, where the domain names were

generated from ﬁxed patterns or predeﬁned rules. The

main aim during this phase was to create multiple

domains that the malware could use to establish

communication with its command and control (C2)

servers, ensuring its persistence and avoiding detection

by defenders.

The domain generation logic in this period

typically involved simple mathematical operations, such

as appending or combining random strings of characters,

or using speciﬁc seed values that were embedded in the

malware itself. These seeds might include:

A hardcoded numeric value or string.

The system’s IP address or machine-speciﬁc

parameters.

A date or time value, though time-based

approaches were typically limited compared to later

implementations.

1.2 Intermediate Phase: Time-Based and Seed-Driven DGAs

1.1.1 Key Characteristics:

Predictability: Since the algorithms were simple

and often used hardcoded seeds, once the logic was

discovered, attackers’ domain generation could be easily

anticipated and blocked.

Low Entropy: The domains generated often

followed simple patterns and had low entropy, meaning

that once a few domains were identiﬁed, it was relatively

straightforward to predict others.

Lack of Resilience: These DGAs were not very

resilient to countermeasures such as preemptively

blocking domains. If defenders could reverse-engineer

the malware and understand its domain-generation

pattern, they could easily stop the malware from

reaching its C2 servers.

For example, a simple pattern-based DGA might

generate a set of domains by appending a sequence

of numbers or characters to a static string, such

as domain1234.com, domain1235.com, etc. Once

researchers discovered the pattern, they could block

all future variants by recognizing this sequence. This

made early DGAs vulnerable and relatively easy to

defend against.

1.2 Intermediate Phase: Time-Based

and Seed-Driven DGAs

As defenders became more adept at blocking

domains generated by simple, pattern-based DGAs,

cybercriminals evolved their tactics to introduce more

complex mechanisms. In the intermediate phase,

they began using time-based and seed-driven methods,

which added a layer of dynamism and unpredictability

to domain generation, making it much harder to

precompute and block domains eﬀectively.

1.2.1 Time-Based DGAs

In time-based DGAs, the malware uses the current

time or date as a core component of the domain

generation algorithm. By relying on an external factor

like the time, the generated domains change regularly,

often on a daily, hourly, or even minute-by-minute basis.

This approach made it much harder for defenders to

anticipate and block domains, as they would need to

continually update their blocklists to match the current

time window.

For example, the malware could:

Use the current date or timestamp as a seed to

generate domains.

Combine the date with a hashing function to

produce a unique domain for that particular day or

time.

Once the malware retrieves the current time from

the system, it could apply a simple formula, like hashing

the date and extracting part of the hash to form a domain.

For instance, a malware may take today’s date (e.g.,

2024-07-07) and generate a domain like a3f2b1c4.com.

The advantage of using time as a seed is that

it introduces an element of ephemeral nature to the

domain, making it valid only for a speciﬁc period.

However, the limitation of early time-based DGAs was

that they were still relatively predictable, as long as the

attackers and defenders shared the same time reference

(i.e., both knew how to compute the date or timestamp).

1.2.2 Seed-Driven DGAs

Seed-driven DGAs introduced another level of

complexity by relying on a dynamic or external seed

value. This seed could be anything from a hardcoded

value in the malware’s code to a value fetched from

an external source, such as an API or social media

feed. By using dynamic seeds, the domains generated

by the malware could be highly unpredictable and less

susceptible to precomputation.

For instance, malware might use a numeric seed

retrieved from an external API (such as a stock ticker

or weather data) and combine it with a hash function

to generate a domain. Alternatively, the seed might

be based on randomly generated values that are stored

within the malware’s conﬁguration. This technique

allowed attackers to generate domains that were diﬀerent

every time they checked the seed, making it diﬃcult for

defenders to predict which domain would be used next.

A typical seed-driven DGA might involve steps like:

Retrieving an external seed (e.g., from a live API

or system parameter).

1.3 Modern DGAs: Adaptive and AI-Assisted Techniques

Using that seed in a hashing algorithm to produce

a domain.

The result would be a completely new domain

each time the seed changes.

For example, if the seed was a random

number like 12345, the domain generated might be

malicious86415.com, and with a new seed (23456), the

domain could change to malicious12345.com. This

unpredictability made it harder for defenders to block

the domains in advance, as the seed could change

dynamically.

1.3 Modern DGAs: Adaptive and

AI-Assisted Techniques

In recent years, Domain Generation Algorithms

(DGAs) have become signiﬁcantly more sophisticated,

driven by advancements in adaptive algorithms and

the integration of artiﬁcial intelligence (AI). Modern

DGAs no longer rely on simple patterns or even basic

time-based and seed-driven methods; instead, they

have evolved to include self-adjusting mechanisms and

intelligent models that make prediction, detection, and

mitigation much more challenging. These modern

DGAs exhibit greater complexity, adaptability, and

unpredictability, making them much harder to detect

and block using traditional methods.

1.3.1

Adaptive DGAs: Dynamic and Evolving

Approaches

The hallmark of adaptive DGAs is their ability

to evolve based on environmental feedback, making

them far more resilient and unpredictable. Unlike

earlier generations of DGAs that relied on static inputs,

adaptive DGAs dynamically adjust their behavior

based on the changing conditions in the environment.

For instance, these algorithms can alter the domain

generation process depending on network conditions,

geographic location, or responses from previous

attempts to communicate with C2 servers.

One key feature of adaptive DGAs is their feedback

mechanism. As malware communicates with its

command-and-control servers, the domains that are

successfully reached or blocked are fed back into the

algorithm. The system then uses this data to modify

the domain generation pattern. For example, if certain

domain names are consistently blocked by defenders,

the DGA might switch to a new method or use a new

seed value, making it much harder to preemptively block

the generated domains. This adaptability introduces

a signiﬁcant challenge for defenders, as traditional

methods of blocking domains based on predictable

patterns are no longer suﬃcient.

1.3.2 AI-Assisted DGAs: A New Era of

Intelligence

The integration of artiﬁcial intelligence (AI) and

machine learning (ML) into DGAs represents the cutting

edge of domain generation technology. By leveraging

AI, modern DGAs can now learn from previous domains,

analyze network traﬃc, and mimic benign domains,

making them incredibly diﬃcult to detect. Traditional

methods that were once eﬀective against simple DGA

algorithms are increasingly less reliable when facing

AI-assisted DGA techniques.

One of the primary advancements AI brings to

DGAs is the ability to generate high-entropy domain

names. Early DGAs often produced domains with

low randomness, making them easier to detect and

block. However, AI-powered algorithms can generate

complex, random, and varied domain names that are

much harder to diﬀerentiate from legitimate traﬃc. By

training on a vast dataset of domains, AI systems can

generate names that follow patterns similar to those

used by legitimate websites, often making it diﬃcult

for traditional detection systems to identify malicious

domains.

Furthermore, AI-assisted DGAs can adapt to real-

time network feedback. If a particular domain or

set of domains is blocked by security systems, the

AI model can adjust its domain generation process

accordingly. These algorithms can even adapt based

on geo-location data, ensuring that the domains remain

eﬀective regardless of where the malware is being

executed.

1.4 Conclusion

Domain Generation Algorithms (DGAs) have

undergone signiﬁcant evolution, from simple,

predictable pattern-based models in the early days to

the complex, dynamic systems we see today. Initially,

DGAs were based on static time or seed values,

generating domains in a predictable manner that could

be easily identiﬁed and blocked. However, with

the rise of more sophisticated time-based and seed-

driven DGAs, attackers introduced variability and

adaptability into their domain generation strategies.

These algorithms allowed for more diverse domain

names, making detection more challenging but still not

insurmountable for traditional defense mechanisms.

In the modern era, DGAs have become increasingly

advanced with the integration of adaptive techniques and

artiﬁcial intelligence (AI). AI-driven DGAs are capable

of learning from past interactions, adjusting their

domain generation patterns in real-time, and creating

highly random, legitimate-looking domains that are

diﬃcult to distinguish from legitimate traﬃc. These

innovations make it increasingly diﬃcult for defenders

to block or preemptively detect malicious domains.

As DGAs continue to evolve, cybersecurity defenses

must also adapt, relying on advanced machine learning,

behavioral analysis, and collaborative intelligence to

combat these ever-evolving threats. The progression of

DGAs marks a critical challenge in the ongoing arms

race between attackers and defenders in cybersecurity.

References

Domain generation algorithm

DGA Detection with data analytics

What is Domain Generation Algorithm?

Real-Time Detection of Dictionary DGA Network

Traﬃc Using Deep Learning

About the Author

Jinsu Ann Mathew is a research scholar

in Natural Language Processing and Chemical

Informatics. Her interests include applying basic

scientiﬁc research on computational linguistics,

practical applications of human language technology,

and interdisciplinary work in computational physics.

About airis4D

Artiﬁcial Intelligence Research and Intelligent Systems (airis4D) is an AI and Bio-sciences Research Centre.

The Centre aims to create new knowledge in the ﬁeld of Space Science, Astronomy, Robotics, Agri Science,

Industry, and Biodiversity to bring Progress and Plenitude to the People and the Planet.

Vision

Humanity is in the 4th Industrial Revolution era, which operates on a cyber-physical production system. Cutting-

edge research and development in science and technology to create new knowledge and skills become the key to

the new world economy. Most of the resources for this goal can be harnessed by integrating biological systems

with intelligent computing systems oﬀered by AI. The future survival of humans, animals, and the ecosystem

depends on how eﬃciently the realities and resources are responsibly used for abundance and wellness. Artiﬁcial

intelligence Research and Intelligent Systems pursue this vision and look for the best actions that ensure an

abundant environment and ecosystem for the planet and the people.

Mission Statement

The 4D in airis4D represents the mission to Dream, Design, Develop, and Deploy Knowledge with the ﬁre of

commitment and dedication towards humanity and the ecosystem.

Dream

To promote the unlimited human potential to dream the impossible.

Design

To nurture the human capacity to articulate a dream and logically realise it.

Develop

To assist the talents to materialise a design into a product, a service, a knowledge that beneﬁts the community

and the planet.

Deploy

To realise and educate humanity that a knowledge that is not deployed makes no diﬀerence by its absence.

Campus

Situated in a lush green village campus in Thelliyoor, Kerala, India, airis4D was established under the auspicious

of SEED Foundation (Susthiratha, Environment, Education Development Foundation) a not-for-proﬁt company

for promoting Education, Research. Engineering, Biology, Development, etc.

The whole campus is powered by Solar power and has a rain harvesting facility to provide suﬃcient water supply

for up to three months of drought. The computing facility in the campus is accessible from anywhere through a

dedicated optical ﬁbre internet connectivity 24×7.

There is a freshwater stream that originates from the nearby hills and ﬂows through the middle of the campus.

The campus is a noted habitat for the biodiversity of tropical Fauna and Flora. airis4D carry out periodic and

systematic water quality and species diversity surveys in the region to ensure its richness. It is our pride that the

site has consistently been environment-friendly and rich in biodiversity. airis4D is also growing fruit plants that

can feed birds and provide water bodies to survive the drought.