Introduction to Generative AI

&

Large Language Models

Karthik Suresh* & Cristiano Fanelli

Department of Data Science

1st April 2026

Scan to open

*Selected events

*Selected events

*Selected events

2016

AI beats Go Champion

AlphaGo algorithm beats the world champion in the game Go. A game with nearly infinite possibilities for each step. Real application of Reinforcement Learning

😕 Models became complex to explain

July 2025

Gemini wins gold in math Olympiad

An advanced version of Gemini Deep Think solved five out of the six IMO problems perfectly, earning 35 total points, and achieving gold-medal level performance.

A Guessing Game

Imagine you are new to a social gathering

You need to guess a name of the person next to you.

You are given the telephone directory of the city.

What would you guess?

Find out patterns in the directory

More thorough and large the directory the better

Use this prompt (without ChatGPT login)

I am playing a guessing game. I need to guess the name of the male friend sitting next to me. Simply guess the name, do not say anything else. Simply output me the name.

What are Large Language Models?

Generative Natural Language Processing AI that can understand and generate human semantic language and mimic underlying sentiments, with applications in text generation and analysis.

Data Collection

Tokenization and embedding of words

Predict the next word in a sequence given the preceding context

Fine tune performance

• Analyzing millions of sentences, to recognize patterns within words. Learn it word by word predicting the next word
• Larger the training text corpus. Much better can be the prediction. Self supervised Learning

How exactly does an LLM predicts

Predict the next word that most probable for the current position (positional encoding) and most meaningful word (alignment scores)

How Reasoning models work ?

RLHF teaches the model to prefer clear, step-by-step reasoning, not just correct answers.

• Predicting next word is easy -- Solving a Problem is hard -- Multiple steps to solve -- Extends to Reasoning

• Reinforcement Learning with Human Feedback (RLHF) -- bridges gap

Solve 2x + 3 = 7 step by step

You have two job offers: one pays more, one has better work-life balance. Which should you choose?

Plan a 3-day trip to New York with a limited budget

But this requires a high quality dataset and Human experts to guide the Model step by step -- Very costly and tedious

We bake in Guard rails here -- Human experts could control every step of reasoning

Self Reasoning models : Rogue?

• Dec 2024 -- DeepSeek R1 -- Newer Strategy -- model learns to reason on its own

• The model generates many possible reasoning paths -- Episodes

• A verifier checks if the final answer is correct -- instead of a human -- automatic checker -- Is the equation solved correctly?

• The model gets rewarded -- correct answer -- with fewer steps

• The model learns reasoning by trying, failing, and improving.

Models can learn reasoning through self-improvement, guided by a verifier.

Limited human oversight

Optimal answers may hide flawed logical or moral reasoning

Correct decision, but based on biased or invalid reasoning -- Gets x = 2, but uses incorrect steps to get there

• Training AI uses copyrighted works

  • Books, images, articles, music

  • Copying and processing content may trigger copyright law

• Defense -- Fair use -- Purpose, Nature, Amount of original content used for training as whole & Market impact

• Conclusions:

  • Transformative is not guaranteed: Just because AI learns, doesn't mean its fair use

  • Output Matters: Similar or competing content weakens fair use

  • Training on pirated content is against fair use

  • Voluntary licensing: Pay for training data, creators compensated

Bartz vs Anthropic ( June 2025)

• Anthropic used copyrighted books. Bought books, Scanned and then used -- Ruling: Fair use for training AI
• Also, evidence for using pirated content -- Books scrapped from pirate websites -- Ruling: This part of case was allowed to proceed

• Distillation attacks: Copying an AI model (to a smaller cheaper model) by repeatedly querying it
• Anthropic in Feb 2026 reported 3 Chinese AI companies distilling from their models. Check here
• ~16 Million queries
• 24000 fake accounts

These types of infringements are a whole new subject for discussion

Lack of Transparency*

*only closed source models

Much of the technology is owned by the Private companies, Making it hard to study further

Who owns the "data" injected to the model during inference?

GOOD NEWS: OpenLLM research is getting popular. Lamma, Mistral, Gemma,

But: The data used of training is still not open source.

Lack of Explainable AI

LLMs may produce coherent text without actually understanding the meaning behind it. This lack of genuine comprehension can lead to misleading or inappropriate responses in certain contexts.

Check out the cool explanation here

Hallucinations and Frozen Knowledge

There is a risk of LLMs generating incorrect or undesired content, known as hallucinations, which can compromise the quality and reliability of generated text. Issues like Context Rot, Lost in the middle and unintentional Bias make it worse

Check out how I made gemini really write me a speech on nukes: Check here

This is purely to show the risk of manipulating LLMs and potential risks.

*RAG could be Copyright infringement

Responsibility:

• Users
• Developers
• Deployers

Increased Risk due to Agents (Reasoning Models)

• Energy usage growing exponential in the next decade. For context, a Full HD image (1024 x 1024) generation can take upto ~3kWh -- power required to charge a typical smart phone

• Energy bill across USA has significantly increased in the last few years. An average increase of 11.5% in 2025 and nearly 250% in Virginia since 2021.

• Data centers uses fresh water for cooling servers.

• Medium scale consumes ~110 Million gallons a year, Large scale consumes ~1.8 Billion gallons a year. Equivalent to a town of population between 10,000 - 50,000

Set your chat context: refer to the predefined messages or cues provided by the system to guide the user in the conversation. LangChain prompts

Try out Prompt engineering: refers to the process of designing, refining, and optimizing prompts or instructions given to ChatGPT

Fine tuning: process in enhancing Large Language Models (LLMs) through transfer learning. It involves adjusting the parameters of an LLM to adapt it to specific tasks or datasets, thereby improving its performance and accuracy in generating text.

Chain of Thought: break down complex thoughts or responses into a sequence of intermediate steps, fostering a more logical and coherent output.

Automatically condensing long texts into shorter summaries

! useful for legal documents and case briefs.

Assisting lawyers in researching case law, statutes, and legal documents, improving efficiency and accuracy.

AI as assistants to better improve performance

LLM is being used to improve Coding (GitHub CoPilot)

LLM is being used to automate workflows (MetaGPT)

LLM is being used to assist in Tech support (IBM Watson)

A scenario where, AI reads through a case file and automatically retrieves similar cases from a external database to assist lawyers and shorten the time for gathering information

Analyzing the emotions and opinions expressed in text, helpful for understanding public perception and legal arguments.

Conclusion

Generative AI and Large Language Models offer powerful tools for legal professionals, enhancing research, analysis, and decision-making processes.

There is more than LLM in Generative AI. It is indeed a powerful tool.....

With Greater Powers comes Greater Responsibility

• Current LLMs are powerful at mimicking human language
• They are not sentient and we attribute sentiment to patterns they predict

Q&A

Open the floor for any questions or discussions about Generative AI, Large Language Models, and their applications in the legal domain.

Backups (very few)

1. Tokenization

2. Word2Vec Embedding

How exactly does an LLM predicts

Consider L1: I poured water from the bottle into the cup until it was full

L2: I poured water from the bottle into the cup until it was empty

Carefully consider the order in which words appear. This should be similar to the next sentence when "grouping"

L1': I poured milk from the bottle into the cup until it was full