DEEPSEEK

into the unknown……

 2

AGENDA
Introduction
Comparison between ChatGPT and DeepSeek
Large Language Models(LLMs)
Neural Networks
U.S. and China
NVIDIA’s response to china
DeepSeek
 WHAT IS DEEPSEEK?
DeepSeek Artificial Intelligence Co., Ltd. (referred to
as "DeepSeek" or "深度求索") , founded in 2023, is a
Chinese company dedicated to making AGI a reality.
AGI stands for Artificial General Intelligence, which
refers to a type of AI that possesses the ability to
understand, learn, and apply knowledge across a wide
range of tasks at a level comparable to human
intelligence.

Liang Wenfeng
WHY ARE CHATGPT AND DEEPSEEK EVEN
COMPARED?
DeepSeek and ChatGPT are both large language models (LLMs) designed for various natural
language processing tasks:
General-purpose: They Conversational abilities: Both
can be used for a wide Constantly evolving: can engage in human-like
range of tasks, including Both are under active conversations, answer
writing, translation, development, with questions, and generate text.
summarization, and code frequent updates and
generation. improvements.
 5

WHAT IS A
LARGE
LANGUAGE
MODEL(LLM)??
 6

LARGE LANUAGE
MODEL(LLM)
A Large Language Model (LLM) is a type of artificial intelligence (AI)
designed to understand and generate human-like text. Think of it as a
super-smart computer program that has been trained on a huge amount of
text from books, websites, and other sources
 7
HERE’S A BREAKDOWN
It can read, understand, How it learns:
and generate text. For •It’s trained on massive amounts of text data, learning
example, it can answer patterns, grammar, facts, and even some reasoning skills.
questions, write essays, •It doesn’t "know" things like humans do but uses statistics
summarize articles, or and patterns to generate responses
even create stories.
Why it’s useful:
It works by predicting 1. It can help with tasks like writing,
what words or sentences learning, coding, or even just having a
conversation.
should come next based
2. It’s like having a very knowledgeable
on the input it receives.
assistant that can help with almost
any text-based task.

In short, an LLM is a powerful AI tool that uses its training on vast amounts of
text to understand and generate human-like language.
 8

CoPilot ChatGPT

LLMs Google’s
Bard

Meta’s
Orca
Llama
 9
Ma’am something named
neural network is ABSOLUTELY!!!!
Let me explain
a viral term nowadays,
how is it related…
are LLMs anything
related to this???

LLM
 10

NEURAL NETWORKS
A neural network is a type of computer system
inspired by how the human brain works. It’s designed
to recognize patterns and solve problems by learning
from examples.

Frank Rosenblatt
What is it? How does it work?
•A neural network is like a Imagine you’re teaching a child to recognize cats:
virtual "brain" made up of •You show them pictures of cats and say, "This is a cat."
layers of tiny units •Over time, the child learns to identify cats by noticing patterns
called neurons (or nodes). like pointy ears, whiskers, and tails.
A neural network does something similar:
•These neurons work 1.Input Layer: It receives data (e.g., a picture of a cat).
together to process 2.Hidden Layers: It processes the data, looking for patterns (e.g.,
information, learn from edges, shapes, colors).
data, and make decisions 3.Output Layer: It makes a decision or prediction (e.g., "This is a
or predictions. cat").

How does it learn? Why is it useful?

•The network learns by adjusting the Neural networks are great at solving
connections between neurons. This complex problems, like:
is called training. •Recognizing faces in photos.
•During training, it compares its •Understanding spoken language
predictions to the correct answers (e.g., Siri or Alexa).
and improves over time (like •Predicting trends (e.g., stock prices
practicing to get better at a game). or weather).
12
13
DeepSeek, a Chinese AI firm, is shaking up the
industry with its low-cost, open-source large
language models, posing a challenge to U.S. tech
giants. The company recently unveiled its latest AI
models, developed for just $6 million — a fraction
of the typical cost for U.S. firms. This disruption has
unsettled majorBUT……….
tech players and impacted their
stock prices.
 15

U.S. after imposing so much regulations on NVIDIA’s chips

How is this
even possible?
 16

•Cooperating with regulators

•Nvidia has offered to cooperate with regulators in China and the US.
•Investigating how chips ended up in China
•Nvidia has asked its distributors to check their customers in Southeast Asia to see how its chips
ended up in China.
•Acknowledging the success of Chinese AI models
•Nvidia has praised the innovation of Chinese AI models like DeepSeek, and has said that they
showcase the potential of AI techniques.
•Explaining its commitment to providing quality products

•Nvidia has said that it works to provide the best products in every region.

Nvidia is a key player in the global AI industry and makes most of the processors used in AI. The US
and China have been involved in a trade war over technology, including chips, and both sides have
implemented measures to assert their influence.
 17
THIS IS REVOLUTION!!!
But how did it manage to do so???
DeepSeek has emerged as a standout in the field of large language models (LLMs) due to
its innovative architecture, cost efficiency, and specialized capabilities. Here are the key
factors that make DeepSeek superior to many other LLMs:

Efficient Architecture and Cost-Effectiveness

•Mixture-of-Experts (MoE) System: DeepSeek uses a MoE architecture, which
activates only a subset of its 671 billion parameters (37 billion per task) rather
than the entire model. This reduces computational costs and energy
consumption while maintaining high performance .
•Low Training Costs: DeepSeek's models, such as DeepSeek-V3 and R1, were
trained for
approximately 6million,significantlylowerthanthe6million,significantlylowertha
nthe100 million+ budgets of competitors like OpenAI's GPT-4. This efficiency is
achieved through advanced data processing and training optimizations .
 19
Superior Performance in Specialized Tasks
•Creative Writing and Emotional Intelligence: DeepSeek excels in creative writing,
producing emotionally resonant and coherent narratives. It outperforms models like
ChatGPT and Claude in emotional intelligence (EQ) benchmarks, making it ideal for
storytelling and character development .
•Reasoning and Problem-Solving: DeepSeek-R1, a reasoning-focused model, achieves
top scores in benchmarks like MATH-500 (97.3%) and GPQA Diamond (73.3%), rivaling
or surpassing OpenAI's o1 in logical reasoning and scientific problem-solving .

Open-Source Accessibility
•DeepSeek models are open-source under the MIT license, allowing researchers and
developers to study, modify, and build upon them. This fosters innovation and
collaboration, making advanced AI tools accessible to smaller organizations and
individuals
 20

Long Context Handling

•DeepSeek supports up to 128K
tokens in context windows,
enabling it to process and maintain
coherence in lengthy texts, such as
large codebases or complex
datasets. This capability surpasses
many competitors, which typically Resource Efficiency and Environmental
handle 32K-64K tokens Impact
•By optimizing hardware usage and
reducing computational overhead,
DeepSeek minimizes energy
consumption and environmental
impact. This is a significant advantage
over resource-intensive models like
GPT-4, which have higher carbon
footprints
 21

Versatility and Specialization

•While DeepSeek excels in
creative and reasoning tasks, it
also performs well in general- Community and Ecosystem Growth
purpose applications. Its •DeepSeek has a growing
specialized models, like community of users and
DeepSeek Coder, are tailored for developers, particularly in creative
coding and technical tasks, and technical fields. Its open-source
offering faster and more nature encourages contributions
accurate code generation and and customization, further
debugging enhancing its capabilities

Cost-Effective Deployment
•DeepSeek's API costs are significantly lower than competitors, with input tokens
priced at 0.14permillionandoutputtokensat0.14permillionandoutputtokensat2.19 per
million. This makes it an affordable option for businesses and researchers
THANK YOU From
• Asmi Maulik
• Amrita Debnath
• Ayushi Biswas
• Deep Das
• Manjari Aich
• Indresh Mukherjee
• Rigam Bhaduri
• Sabyasachi Mukhopadhyay
• Soumyashree Ghosh