Intro to Artificial Intelligence

CSE 425 | Lecture 1

What we will cover in this lecture

• Introduce some popular AI systems
• Focusing on LLMs and their capabilities
• What is AI and what are the different types of AI
• The present and the potential future of AI

Some Popular AI Systems

• Search Engines (e.g., Google, Bing)
• Voice Assistants (e.g., Google Assistant, Siri, Alexa)
• Recommendation Engines (e.g., Amazon, Facebook, YouTube, TikTok)
• Conversational and Multimodal AI (chatbots and assistants)
• e.g., ChatGPT, Claude, Gemini, Qwen, DeepSeek
• Coding assistants and agents (e.g., Claude Code, Codex, Cursor)
• Self-Driving Cars

Search Engines

• Understand natural language queries
• Provide relevant results
• Learn from user behaviour
• Increasingly, they answer the question rather than return links
• The engine retrieves documents, then a language model writes an answer from them
• This technique is called retrieval-augmented generation, or RAG
• You will build a RAG system yourself later in this course

Voice Assistants

• Answer questions, set reminders and alarms, make calls and send messages
• Control smart home devices, play music, give directions, translate languages
• The early generation (Siri 2011, Alexa 2014, Google Assistant 2016) worked by intent classification
• A fixed list of things the assistant could do, and a parser that mapped your speech onto that list
• Today's assistants are built on large language models with speech in and speech out
• No fixed list of intents - the model composes a response to anything
• Tool calls replace hand-written skills

Recommendation Engines

• Recommend customers the products or content they are most likely to want
• Early examples:
• Using unsupervised clustering to identify similar customers in online shops
• Today: deep sequence models over your entire interaction history
• An interesting transition:
• Suggesting posts and videos to serve the customer better
• and to maximise screen time, which means more advertising to show

Conversational and Multimodal AI

• ChatGPT, Claude, Gemini, Qwen, DeepSeek
• These are all large language models (LLMs) that can generate text, translate languages, write different kinds of creative content, and answer questions in an informative way
• Multimodal means one model handles several kinds of input and output
• Text, images, audio and video, rather than a separate model for each
• Most are now available as a chat interface, an API, and inside other software

How to use LLMs? Prompt Engineering

• A prompt is the instruction you give the model
• Example:
• "You are an expert translator. Your task is to translate the following text to Bangla. Please make sure the translated text is natural, easy to follow, and has literary quality to it."
• What makes a prompt work:
• State the task, the audience, the format and the constraints precisely
• Give an example when the format is unusual
• Supply context the model cannot retrieve for itself

LLMs and Bangla

• Model quality is uneven across languages, and it follows the data
• Far more English text exists online than Bangla text
• So these models write a poem in the style of Shakespeare more convincingly than one in the style of Fakir Lalon Shah
• The gap has narrowed since 2023:
• Current open models are trained on 140+ languages deliberately
• Capability learned in English partly transfers to other languages
• It has not closed. This remains an open research area - and a good project topic.

Multilingual coverage: Google, Gemma 3 and Gemma 4 model cards.

LLM models can generate code

• Prompt: "Write a python code to find prime numbers"
• The model produces working code for most routine tasks, and now also runs it, tests it, and edits files across a project
• Warning!
• DO NOT just copy and paste code
• Must check the reliability of the code
• The risk grew rather than shrank, because generated code now usually runs
• Code that fails loudly is safe. Code that is subtly wrong is not.

Image, audio and video generation

• Diffusion models generate an image by starting from noise and removing it step by step, guided by the prompt
• Prompt: "A boy flying a kite in the paddy field, in the summer afternoon, in the style of Van Gogh"
• Current systems also generate audio and short video from a text description
• Quality improved sharply between 2023 and 2026, but these models still:
• Struggle with precise text inside an image, and with counting
• Make errors of physics and continuity in video

Idea behind language models

• The task these models are performing is usually simple --
• an autocomplete task
• predicts the next plausible word in a given sequence of words
• "next word prediction"
• "Dhaka is the capital of ___"
• With this phrase as input, the model should come up with "Bangladesh"
• Tokens can be whole words or parts of words

How these models are trained

• Pre-training: predict the next token over an enormous corpus scraped from the internet
• This produces a model that can continue text, not one that helps you
• Post-training: teach the model which continuations are wanted
• Supervised fine-tuning, then learning from human or AI preference signals
• This is where helpfulness, refusal behaviour and response format come from
• Inference-time compute: let the model work through a problem before answering
• The model spends tokens on reasoning steps the user usually does not see

Reason behind such plausible answers

• Models do not memorise a database of answers
• They predict the next token based on patterns learned from the training data
• There is "inherent structure" in the data:
• Words that occur in similar contexts tend to have similar meaning
• For most positions, only a small set of words is a plausible continuation
• Modern models can also retrieve: they search the web, query databases and call tools when the answer is not in their weights
• But the underlying objective still rewards plausible text, not true text

Reasoning in language models

• Early models were not trained to reason, and it showed
• "Tarik is taller than Suman. Suman is not taller than Masud. Is Masud taller than Tarik?" - a question the 2023 models often got wrong
• Models are now trained explicitly to produce a chain of intermediate steps before answering
• Two consequences you can observe directly:
• More compute spent at answer time buys more accuracy, fairly predictably
• Interfaces now expose "thinking" modes and effort settings
• This is a real trade-off between latency, cost and correctness

How fast capability is improving

Source: IntuitionLabs, HLE benchmark review, Aug 2026. Benchmark released Jan 2025 by CAIS and Scale AI.

Prompting techniques

• Capabilities can be improved through prompting techniques:
• Chain of Thought - structure the output as a series of steps
• Self Consistency - generate several reasoning paths and take the most consistent
• Multiagent / Multipersona - the model takes several roles and debates the answer
• Note: in current reasoning models, chain of thought is trained in
• Telling such a model to "think step by step" adds very little
• These techniques still matter for smaller and open models you run yourself

Tool use and agents

• An agent is a model in a loop with tools, memory and a goal
• Observe, decide, act, observe the result, repeat until the task is done
• Tools let the model reach outside itself: run code, query a database, call an API, read and write files
• The Model Context Protocol is an open standard for exposing tools to models
• Announced by Anthropic in November 2024
• Servers expose data and tools; the AI application connects to them as a client
• This is the architecture behind current coding assistants

Source: Anthropic, Introducing the Model Context Protocol, 25 November 2024.

The model landscape today

Capability = score on Humanity's Last Exam (Artificial Analysis, Kili review Apr 2026). Prices are published output rates. Figures as of Sept 2026.

Open-weight models you can run

• Open weights means the trained parameters are published and you can download, fine-tune and deploy them yourself
• Gemma 4 (Google, April 2026) - Apache 2.0, 2B to 31B, runs on a laptop or a phone
• Qwen3.8-Max (Alibaba, August 2026) - 2.4 trillion parameters
• DeepSeek V4-Pro and V4-Flash (April 2026) - roughly USD 0.44-0.87 per million tokens when hosted
• Kimi K2.6 (Moonshot, April 2026) - 58.6% on SWE-Bench Pro, matching a frontier closed model
• GLM-5.2 (Zhipu, June 2026) - leads the open-weight intelligence index

Sources: Google Gemma 4 model card; Wikipedia (Qwen); Data Gravity, Jun 2026.

Gemma 4: one family, five sizes

Source: Google, Gemma 4 model card. Released 2 April 2026 under Apache 2.0. The 31B scores 84.3% on GPQA Diamond and 89.2% on AIME 2026.

Which models are actually being used

Source: Data Gravity, Jun 2026. OpenRouter routes API traffic across many providers.

The falling cost of AI

Source: Epoch AI, LLM inference price trends.

Context windows and multimodality

• The context window is how much text the model can consider at once
• Current models handle hundreds of thousands of tokens
• Even small open models ship with 128K to 256K
• That is a few hundred pages of text in a single prompt
• Multimodality is native rather than assembled from separate systems
• Text, images, audio and video handled by one model
• Caution: a long context window does not mean the model uses all of it equally well

Source: Google, Gemma 4 model card (128K-256K context; text, image, audio, video).

Surprises, Capabilities, and Challenges

• While these models often work well, they still have limitations:
• Factual errors and hallucination
• Unintended biases inherited from training data
• Uneven performance outside English
• Jagged capability - very strong and very weak on tasks that look similar
• Cost and energy at scale
• The visible problems of 2023 - repetition, incoherence, malformed images - are largely gone. The structural ones are not.

Factual errors

• Text generated by language models can contain factual inaccuracies, incorrect information, or nonsensical statements
• "Elephants are known for their ability to fly gracefully through the skies, flapping their wings with precision and elegance."
• Why it happens: the model is optimised to produce plausible text
• Where the training data is thin, plausible and true come apart
• Retrieval reduces this but introduces a new failure: a confident answer with a citation that does not support it

Unintended biases

• Like other AI systems, generative models can inherit biases present in the training data, potentially leading to biased or inappropriate outputs
• Sources of bias:
• What is over- and under-represented on the internet
• Who labelled the preference data during post-training
• Which languages and cultures the evaluation covered
• Harder to audit now, because the systems are larger and the training data is rarely published in full

Limited capability in complex tasks

• Researchers measure these models on benchmarks - fixed sets of problems with known answers
• On many standard benchmarks the models now score very highly
• MMLU general knowledge: above 90%
• GPQA Diamond, PhD-level science: above 90%
• On harder, newer benchmarks there is still a long way to go
• Humanity's Last Exam: around 59%
• ARC-AGI-3, learning an unfamiliar game from scratch: under 1%

Sources: Kili Technology, AI benchmarks 2026; Artificial Analysis; ARC-AGI-3 developer preview, Mar 2026.

How models score on benchmarks

Sources: Kili Technology (Apr 2026); Artificial Analysis; ARC-AGI-3 preview (Mar 2026).

Challenges to resolve

• Data efficiency remains a challenge
• These models need far more examples than a human does to learn the same thing
• Rather than scaling up models indefinitely, researchers are working on:
• More efficient architectures and smaller models that match larger ones
• Technologies that learn with less supervision and less data
• Spending compute at answer time rather than only at training time
• Evaluation itself is an open problem - benchmarks saturate faster than they can be built

Self Driving Cars

• Imagine you're sitting in the passenger seat of a car, but there's no driver behind the wheel. The car is moving, changing lanes, and avoiding obstacles all on its own.
• Self-driving taxis now operate commercially in a limited number of mapped cities
• They are not yet general-purpose: each deployment is carefully bounded by area, weather and road type

Self Driving Cars: how they work

• Sensors: cameras, radar, lidar, GPS, ultrasonics
• The car's computers attempt to:
• create a real-time model of the surroundings
• predict what other road users will do, and plan a path
• Why this stayed hard while language models advanced quickly:
• Mistakes are physical and cannot be undone
• The long tail of rare situations never ends
• There is no opportunity to retry a failed attempt

So, what is AI?

• Colloquially, AI refers to the machines (or software) that perform tasks requiring human-like intelligence. These tasks encompass a broad range, including
• Learning
• Reasoning
• Problem-solving
• Perception
• Language understanding etc
• One weakness of this definition: the list keeps moving. Every capability that is achieved stops being called intelligence.

A slightly different definition

• John McCarthy, in his 2004 paper "WHAT IS ARTIFICIAL INTELLIGENCE?"
• It is the science and engineering of making intelligent machines, especially intelligent computer programs. It is related to the similar task of using computers to understand human intelligence, but AI does not have to confine itself to methods that are biologically observable.
• That final clause matters: today's models are not built to resemble the brain
• But what is intelligence?

John McCarthy - Computer scientist and one of the founders of the discipline of Artificial Intelligence.

Definition from a popular AI textbook

• Stuart Russell and Peter Norvig, authors of Artificial Intelligence: A Modern Approach, discuss four "schools of thought" to define AI:
• Systems that think humanly / systems that think rationally
• Systems that act humanly / systems that act rationally
• The field has largely settled on acting rationally as the engineering target
• Current language models sit awkwardly across the split: they are trained to imitate human text, then optimised to act rationally

Acting/Thinking Humanly: Turing Test

• The Turing Test, proposed by Alan Turing (1950), was designed to provide a satisfactory operational definition of "intelligence"
• A computer passes if a human tester, after posing written questions, cannot tell whether the responses come from a person or a computer
• Turing's definition falls under "systems that act like humans"
• It is no longer used as a working measure of intelligence
• Systems pass it under ordinary conditions without anything like general intelligence
• It rewards imitating human error and hesitation, which is not a useful goal

Moravec's Paradox

• What is easy/hard for humans can be non-intuitively very different to what is easy/hard for computers
• chess is easy for computers
• as it is a closed, deterministic system with a discrete action space, full observability, etc etc.
• On the other hand, humans can tie a shoe or fold a shirt and don't think much of it at all but extremely difficult for machines
• The paradox has moved but not disappeared: models now pass graduate physics exams and still cannot work out the rules of an unfamiliar game

Moravec's Paradox, measured

Source: ARC-AGI-3 developer preview results, Mar 2026.

Acting Rationally

• The Acting Rationally approach to AI focuses on developing AI systems that can act in a way that maximizes their expected utility
• This method is frequent in reinforcement learning, enabling AI to learn from experiences and optimize long-term rewards
• It is also how language models are post-trained: the model is optimised against a preference signal that stands in for what people want

Utility and Expected Utility

• Utility refers to a measure of an agent's satisfaction with potential outcomes
• Under uncertainty, the agent pursues expected utility by averaging outcome values weighted by their probabilities
• Uncertainty arises from incomplete knowledge about future consequences
• This framework also explains most AI failures: the agent maximised exactly what was specified, and what was specified was not what was wanted

Intelligent Agent and Rational Agent

• An agent is an independent program or entity that interacts with its environment by perceiving its surroundings via sensors, then acting through actuators
• Rationality depends on:
• Performance measure that defines criterion for success
• Agent's prior knowledge of the environment
• Actions the agent can perform
• Agent's percept sequence to date
• This is the same loop as the tool-using agents earlier in the lecture

PEAS of a self-driving car

• Performance criteria: time to reach the destination (minimize), fuel or charge consumed (minimize), comfort of the passenger (maximize), safety
• Environment: road, traffic, pedestrians, passengers, weather conditions etc.
• Sensors: cameras, microphone, GPS, RADAR, LiDAR, engine indicators etc
• Actuators: steering wheel, brake, accelerator etc.

Question

• What should be the performance measure, sensors and actuators of an Intelligent Programming Tutor?
• What about an LLM coding agent that edits files and runs tests?
• Do you think it is always possible to list all possible performance criteria so that nothing untoward happens?
• Can you identify what can go wrong with these agents given the performance criteria you listed?

Types Of Artificial Intelligence

• Weak AI - also called Narrow AI or Artificial Narrow Intelligence (ANI)
• Performs specific tasks very well, but lacks general intelligence
• Examples: facial recognition, chess programs, self-driving systems
• Strong AI - Artificial General Intelligence (AGI)
• Would possess human-level intelligence across any intellectual task
• Current systems complicate this split: general across a huge range of tasks they were never trained for, yet failing at things no human finds difficult
• A more useful question than "is it AGI" is "capable at what, measured how"

Do Machines Understand?

• Searle's Chinese room argument: a person in a room follows rules to manipulate Chinese symbols without understanding Chinese
• Brain Simulation objection: suppose the program simulates in fine detail the action of every neuron in the brain of a Chinese speaker?
• Searle argues such a simulation does not reproduce the important features of the brain - its causal and intentional states
• He is adamant that "human mental phenomena [are] dependent on actual physical-chemical properties of actual human brains"

What counts as understanding?

• One theory of understanding rests on the notion of minds, or agents, making models of phenomena external to themselves
• Criteria to evaluate understanding of a phenomenon:
• the ability to make predictions about the phenomenon
• the ability to achieve goals with respect to the phenomenon
• ability to explain the phenomenon
• ability to create or recreate the phenomenon
• By these criteria, current systems partly understand some domains

Bieger, Jordi, Kristinn R. Thorisson, and B. Steunebrink. "Evaluating understanding." IJCAI Workshop on Evaluating General-Purpose AI, Melbourne, Australia. 2017.

Present and future of AI

• AI is advancing quickly and is now embedded in everyday software
• It is showing big progress in many areas
• Yet it faces challenges and pitfalls too
• Three things that changed most between 2023 and 2026:
• Models that reason before answering, rather than replying immediately
• Agents that use tools and act, rather than only producing text
• Capable open-weight models anyone can download and run

Technological Advancements

• Deep Learning: neural networks for pattern recognition and problem-solving
• Natural Language Processing: machines understanding and generating human language
• Computer Vision: extracting meaning from images and video
• Reinforcement Learning: mastering tasks through trial and error
• Foundation models: one large pre-trained model adapted to many tasks, rather than a separate model trained for each
• This last shift is the one that reorganised the field

Applications across industries

• Healthcare: diagnosing disease from scans, predicting patient outcomes, assisting with documentation
• Finance: fraud detection, risk management, document processing
• Transportation: self-driving vehicles, delivery drones, traffic optimisation
• Software: code generation, testing and review, now a routine part of the job
• Agriculture and climate: yield prediction, disease detection from images, flood and weather modelling
• The last of these is where Bangladesh-specific work is most needed and least done

Ethical considerations

• Bias: Algorithms reflecting human prejudices, leading to unfair decisions because of the bias in the data. Imagine a loan application system biased against certain demographics, perpetuating existing inequalities.
• Who should be responsible for making AI systems safe, unbiased?
• Who is responsible for mistakes made by AI systems?
• Privacy: Our data used in ways we don't understand or consent to.
• Training data collected from the public web without individual consent
• Inference: what can be worked out about you from what you did share

Ethical considerations

• Job displacement: automation replacing human work, raising concerns about unemployment and inequality
• Which jobs are more likely to be replaced by AI?
• Demands for which jobs are likely to grow?
• What new jobs can AI create?
• The tasks most exposed turned out to be cognitive and entry-level rather than manual - which matters directly to this room
• Harmful content creation using generative AI has become cheap and hard to detect
• Realistic fake images, video and audio

Ethical considerations

• Environmental impact
• Training a large model is energy-intensive, but the balance has shifted
• A model is trained once and then queried billions of times, so inference now dominates the total energy cost
• Business incentives behind AI systems have created social media 'bubbles' and "skewed perception of reality"!
• Concentration of power
• Proprietary models can lead to high concentration of power to a handful of big tech companies
• Open models without any regulation can lead to increasing misuse

The future of AI

• Ethical AI Development: transparency, fairness and accountability; keeping powerful AI helpful, honest and safe
• AI Regulation and Governance
• The EU AI Act moved into enforcement on 2 August 2026
• Chatbots must disclose that they are AI; deepfakes must be labelled; AI-generated content must carry machine-readable markers
• These rules reach any system serving EU users, wherever it was built
• AI-Assisted Human Capabilities, AI for Social Good, and advances in explainable AI, drug discovery and human-AI interaction

Source: European Commission, AI Act enforcement from 2 August 2026.

A few discussion points

• What are the societal implications of strong AI?
• How can we ensure the ethical development and use of AI?
• What does the future hold for the relationship between humans and AI?
• Can we ever build AGI? Why or why not?
• If a model can be downloaded and run by anyone, who is accountable for what it produces?

Next Lecture

• Intro to Machine Learning
• Figures in this lecture are current as of September 2026 and carry their sources on the slide
• Epoch AI; Artificial Analysis; ARC Prize; Google Gemma 4 model card; European Commission
• In this field, always check the date on a number before you rely on it