Issue #17: July 22, 2025
Let’s dive into the world of AI with Nikhil Gupta, our very own machine learning engineer,
who will demystify the tech everyone’s buzzing about!
What’s Inside This Issue?
- Concepts and Terminology: Discover concepts and terminology introduced in the featured article.
- Featured Article: Dive into the world of AI and uncover the tech everyone’s buzzing about!
- About the Author: Join Nikhil Gupta, a machine learning engineer passionate about the intersection of AI, biotech, and healthcare.
- Hands-On Activities: Try AI for Yourself: Explore Tokenization and Prompt Engineering.
- Environmental Equity and Sustainability: Discover how AI impacts fairness in our world and its environmental footprint.
- Preparing for a Career that Uses AI: Explore an AI pathway and how to get started.
- Glossary: Recap key concepts and terminology from the featured article.
1. Concepts and Terminology
Artificial Intelligence (AI) – the field of developing systems that are capable of mimicking human behavior and capabilities, and extending them.
Machine Learning (ML) – a subcategory of AI that uses algorithms to automatically detect patterns in data, and using those patterns to make decisions.
Neural Network – a computational model inspired by the way the human brain functions.
Transformer – a type of neural network that is specialized for processing sequences (like text).
Machine Learning Model – a set of parameters organized into a structure like a neural network that can be trained on data and used for analyzing data, predicting information, and generation.
Large Language Model (LLM) – a type of machine learning model that is trained on a large amount of text data in order to understand and generate human-like language.
Artificial General Intelligence – a hypothetical type of AI that has human-like intelligence and capabilities across a broad set of knowledge, rather than being trained for a specific task.
2. Featured Article
ChatGPT, is it the greatest tool ever created or just the beginning of a science fiction movie?
By Nikhil Gupta
Artificial Intelligence (AI) has taken the world by storm, being integrated into nearly every step of our lives from figuring out our breakfast order to driving our cars to being our (supposedly) infinitely knowledgeable personal assistant. But what is actually happening when you ask ChatGPT a question? And how can you make sure that the smartest person in the room is still a human?
Background
AI itself is a very broad term, and encompasses the entire field dedicated to developing systems (robotic, computational, etc) that are capable of mimicking human behavior and capabilities, and in some cases even going beyond human abilities. This technically includes, for example, some aspects of robotics used in industrial applications, early chatbots, speech recognition, and even Roombas. However, in this article, we’ll specifically be looking at machine learning and large language models (LLMs).
A recent-model robotic vacuum cleaner for your floors.
Image credit: Roomba
(Paul Jr. got to work on the latter company’s first robots for his graduate degree capstone project at Cal.)
Image credit: Agility Robotics
Machine Learning is a subset of AI that deals with developing algorithms that can perform functions without human intervention, usually after some level of training. This is primarily done by identifying patterns in data, and using that information to then make accurate predictions.
A Venn Diagram of how various AI concepts are linked.
Image Credit: Wikimedia Commons: AI relation to Generative Models subset
A classic example of a machine learning model is figuring out the price of a house. Say you have information from 1000 houses, that for each one, includes the year it was built, all the different materials used, if it has A/C or not, the square footage, and 100 other different details in addition to the price of the house. Each information type is called a feature, and these are the pieces of information we can train the machine learning model on. We’ll go more into training later, but with a machine learning model we can feed it all this data and then point it at any house and it will be able to tell us the expected price of the house.
Another popular use of machine learning is image recognition. There are many types of image recognition in AI, but the simplest is called image classification. This is where you give a machine learning model a picture, and it will give you a category it thinks the picture would fall under, based on how it was trained. In this case, each pixel of the image is a feature (and sometimes each pixel has three values – one for red, green, and blue – and each of those are features). As you can see, the number of features quickly becomes very high, to the point where a human or simpler computer program wouldn’t be able to analyze it. That is where a machine learning model comes in.
You probably realize that humans are actually pretty good at identifying images, and it doesn’t take much effort or even math to figure it out, so why do we need machine learning for that? That is where neural networks come into play.
Neural Networks (or “neural nets”) are a common type of machine learning model directly inspired by the human brain. Just like our brains have billions of interconnected neurons that send and receive signals, processing everything we see, hear, and think (e.g., learning to identify an “apple” from its sight), neural networks mimic this structure.
Once scientists understood how our brains work, they adapted that into the general neural network model, which you see below. This simplified diagram helps us visualize how signals of information travel from start to finish in both brains and computers. Keep in mind that not every path shown here exists in reality, and each connection can carry a different ‘weight’ of information.
A simplified neural network model.
Image Credit: Wikimedia Media: Neural Network
Each circle in the diagram represents a neuron. As you can see, every circle in one column connects to all circles in the next. These columns are called layers, and they’re fundamental building blocks of machine learning models.
The first column of neurons is the input layer, representing features like house details or image pixels. We can then add hidden layers to model more complex details. Generally, more and larger hidden layers can improve a model, but they also increase training time and can sometimes hurt performance. Finding the best combination of hidden layers is a key area of machine learning research. The final layer is the output layer, where all information funnels down to a single decision in this example.
Training a model
We’ve brought up “training” a model a lot, but what does that mean? There are many different types of training, but we will discuss one of the simpler ones: supervised training. In this type of training, we give the model both inputs and the expected outputs. In our image example, that would mean we give the model a bunch of images that have already been labeled with their descriptions. This labeling is usually done manually but now there are many datasets available to use so not everyone has to sit down and manually label 1000’s or more data points.
The actual math behind training relies heavily on linear algebra and calculus. Think of each connection between neurons having a specific weight – like an importance factor. These weights determine how strongly a signal from one neuron affects the next. The training process essentially tries to figure out the best combination of these weights to make accurate predictions. You might already be familiar with a simpler version of this idea: finding the line of best fit:
Linear regression aims to minimize the error (black lines) by finding the best-fit red line.
Image Credit: Wikimedia Media: Residuals for Linear Regression Fit
We are given a set of data points (the dots) and need to find the equation of the line that best represents their trend. This process, called linear regression, is a basic statistics and machine learning technique. We start with a random line and keep changing its parameters (slope and y-intercept) until the line minimizes our error (the vertical black lines in the graph). In linear regression, only two parameters change.
However, in a machine learning model, every single weight is a parameter. This means we’re working with problems that can involve hundreds of dimensions, making them impossible to graph directly like a simple line. The problem of minimizing this complex error is called gradient descent, and it’s best visualized by looking at the shape of a surface.
Gradient descent uses advanced calculus and linear algebra for the math, but the concept is straightforward. Just like with linear regression, all the weights in our model start off randomized. Based on our input and output data, we calculate an initial error. Then, we change all the weights by a tiny amount, observing if the error gets smaller or bigger. If the error shrinks, we continue changing the weights in that direction. If it grows, we reverse direction. For a visual, imagine that this figure below is a surface plot showing all the different errors for various combinations of weights:
Image Credit: Wikimedia Common: Grafico con MATLAB di una superficie
We want to get to the dark blue area because that’s where the error is the least. Imagine a ball representing our algorithm placed starting at the red peak; if we let go then it will eventually roll down the slopes until it hopefully finds its way to the darkest blue part of the graph (the lowest point on the right side). I say “hopefully” because it’s possible for the “ball” to get stuck in that lighter blue divot (front middle) instead. This is a common problem in machine learning training, and researchers are always developing more advanced algorithms to ensure we reach the absolute minimum error.
Once this training process reaches that minimum point, congratulations! Your machine learning model is now trained and ready for predictions. While there are additional nuances like overfitting and underfitting, and more advanced algorithms beyond gradient descent, we encourage you to explore these topics further on your own!
Transformers and LLMs
The latest big craze in machine learning models is the transformer architecture. First proposed by Google scientists in 2017, it laid the groundwork for the many Large Language Models (LLMs) we see today, including ChatGPT, Gemini, Claude, and many others.
While the actual details are complex, transformers essentially allow engineers to feed a vast amount of text into a model, enabling it to learn how all pieces of language—from letters to words and sentences—fit together. Language is broken down into small pieces called tokens (combinations of letters and symbols that form words). The model then learns how each token relates to others and, crucially, what is most likely to be the next token in a sequence. This is essentially how LLMs work: they take your prompt, convert it into a sequence of tokens, and then build a response one token at a time by figuring out the most likely next token.
Here’s a summary image of the overall tokenization and response building process:
Tokenization and response building process.
Image Credit: Paul Dennig, Sr.
At their core, LLMs are like a very advanced autocomplete. While specific training data, algorithms, and architectures allow for extra behaviors such as checking their own work, searching the internet, or accessing files, ultimately, the LLM still builds responses one token at a time, based on what’s most likely to come next.
Be Smarter Than AI
Machine learning models were once confined to very specific situations due to intense software and hardware requirements. However, thanks to the internet and cloud data centers, companies like OpenAI, Google, X, and Anthropic now run their models on massive servers, giving everyone access through a website. ChatGPT, for example, is the 5th most visited website on the internet! People are clearly relying heavily on these new AI tools, but it’s crucial to understand when and how to use them most effectively.
When we look “under the hood,” these models aren’t actually searching for information, thinking, or doing any real reasoning—at least not today. All they’re doing is figuring out the most likely response. While the unbelievably vast amounts of data used to train them give them a good idea of how humans respond, this has several important implications.
Firstly, a core principle in machine learning is “garbage in, garbage out.” If the data used for training is bad, the model will be bad, no matter how much you train it. The internet is full of misinformation, incorrect facts, and opinions, and a machine learning model can’t distinguish truth from falsehood without specific labeling (an impossible task across the entire internet). So, much of the training data likely contains “garbage” or insufficient information, especially for questions at the leading edge of human knowledge. AI companies are constantly trying to fix these issues, but they will always remain a challenge. Therefore, don’t take everything ChatGPT says at face value; always double-check with your own independent research!
Additionally, while these models excel at recalling information from their training data, you won’t get accurate results if you ask about information they were never trained on. This is a general shortcoming: machine learning models are excellent at what they were trained to do, but not much else. Research is constantly exploring ways to continually train models with new information and keep them updated. This also leads us to Artificial General Intelligence (AGI). AGI is a hypothetical type of AI truly capable of analyzing and responding to new information and learning new behaviors without additional training or human intervention, effectively acting like a human. The legal and ethical implications of AGI are vast, making a good understanding of AI even more critical for making informed decisions about its role in our lives.
Security and Data Privacy
Finally, let’s talk about security and data privacy. Using AI has never been easier, with Large Language Models (LLMs) right at our fingertips. But remember the old saying: “If something is free, you are the product.”
When you use a free LLM service, it’s very likely that all the information you feed into it is being used for future training. This means your private data—and everyone else’s—might eventually show up online or be used by companies to guide their research.
So, be extremely careful about what you share with any free LLM service. Never reveal anything that should remain private. Where possible, dive into the settings and turn off any data sharing or agreements that allow your data to be used for training.
On a brighter note, many new tools let you run your own LLM locally, directly on your phone or laptop. These offer the best of both worlds: you get the functionality and ease of use of a tool like ChatGPT while keeping all your data and information completely private.
Summary
In summary, while AI and machine learning are evolving at a rapid pace, their core fundamentals remain constant. A solid understanding of how these powerful technologies actually work will help you use them more effectively and intelligently. The possibilities they offer are endless, but so are the potentials for harm and misinformation.
Remember, AI is a tool, nothing more, and should be used as such. Regardless of additional functionalities and “bells and whistles,” at their core, Large Language Models (LLMs) are simply good at sounding human—nothing more, nothing less.
3. About the Author
Hi, I’m Nikhil! I’m currently a Machine Learning Engineer, but my primary interests are in biotech and healthcare, and how we can apply engineering techniques to solving biological problems.
I got my Bachelor’s degree in bioengineering from UC Berkeley, where I also minored in Electrical Engineering and Computer Science (EECS). I could never pick a single path, so I found a way to integrate my various interests (biology, healthcare, engineering, robotics) into my education. I am a big fan of interdisciplinary education, since the biggest and most important problems facing us as a community and a species aren’t just tied to a single domain. Solving these problems requires people who know how to talk and work across disciplines.
That’s why I pursued a Master of Translational Medicine (MTM) from UCSF and UC Berkeley. The MTM program integrated bioengineering, clinical studies, and business to teach students how to navigate the field of biotech and ensure the great ideas that scientists and engineers come up with actually make their way to the people who benefit most.
Beyond work, I love learning new things and picking up new skills. We live in an age where there is so much information and knowledge ready to be accessed, and everyone should make the most of it. I also love books (of any genre), tennis, video games, and playing the piano.
I met Paul through a Harry Potter course I taught at Berkeley, and we quickly connected over our mutual love of engineering, science, and just being goofy. When we both got into Masters programs at Berkeley we decided to be roommates, and that decision was a significant reason why it was one of the best years of my life. Paul had a great zest for life, and he inspired that enthusiasm in others, something I always admired him for. We talked about science, engineering, stories, and everything under the sun. Paul had a depth of character that was rare to find, and improved the lives of everyone around him. One of my favorite memories with Paul was watching the sunset from our Berkeley apartment and listening to Ratatouille music.
4. Hands-on Activities
Experiment with Tokenization
Imagine that you are a computer. It sounds like a funny thing to do, but if you wish to head into computer work fields, you’ll frequently need to proceed step-by-step, just as a computer does, to write and debug your work. This exercise lets you see the results of tokenizing efforts, for whatever you’d like to see parsed (broken into smaller pieces).
First, go to this web page to reach the “gpt tokenizer playground,” a free online tokenizing tool: gpt-tokenizer playground
At first, you’ll see two empty windows. You’ll click your cursor in the top window, type into the top window, then see the results in the second window – color coded to show off your tokens! First, try typing “This is a test.” and see if your results match these:
Then, play with it! We’ve asked Google Gemini™, an AI tool free to those with a Google™ account, “what would be some things to try next?” Gemini responded in practically no time, making many suggestions. Here are some. Can you try them? Are the results what you’d expect? (Hint: Copy what is between the quotation marks below, then go to the tab in your browser with the tokenizer and paste into the top window.)
- “Hello, world!” vs. “Hello world!”
- “Supercalifragilisticexpialidocious” {think: Mary Poppins}
- “tokenizer” vs. “tokenization” vs. “tokenize”
- “ChatGPT” vs. “chatGPT” vs. “chat gpt”
- URLs and Email Addresses: “https://www.google.com” or “user@example.com“
- Emojis: “Hello 👋 world! 😊”
- A short sample of computer code (a “snipet”): print(“Hello World!”)
Can you think of other things to try? How will it handle contractions with an apostrophe? Short and long numbers, with (or without) commas and a decimal point? Can you now begin to see how a computer can recognize programming code or a written language? Voilà ! (That’s French, you know! It means “there it is!”)
Become a Better ‘Prompt Engineer’
In general, we always try to steer away from the “garbage in, garbage out’ syndrome. So, we asked an AI, Gemini, how we could get better at writing prompts, which as you probably already know, are the inputs to an AI engine. We got this response:
- Be Clear and Specific. What aspects are you interested in? It’s like giving the AI, your master chef, your instructions. Details, please! Don’t just say “bake bread.” Say “bake a round loaf of sourdough bread.” Or, better, “Please provide a simple, step-by-step recipe for a beginner to bake a round loaf of sourdough wheat bread with sesame seeds. Assume I have a basic sourdough starter. Include oven temperature and estimated baking time. It should be dairy free. Please tell me how to make the crust crispy.”
- Give Your Context (purpose). For this newsletter, we told our AI that we’d like to explain to high school students about prompt engineering and to please give us guidance on how to help them write better AI prompts. {Naming the audience is important.}
- Define the Format of the Answer that You’re Expecting. The AI doesn’t know your intentions. Did you want prose? A poem? A bullet list? Paragraphs? How many words? The level of detail or complexity required?
- Give Examples (optional, but powerful). Ex.: “Here are some examples of rhyming couplets: ‘The cat sat on the mat, and that was that.’ ‘A dog ran by, beneath the sky.’ Now, write three more rhyming couplets about school.”
- Refine and Repeat. Try first, then see what comes out of the master chef’s oven. Then refine your prompt: ask the AI for help to write the best-possible prompt for what you want by telling it what you’d like to change.
Here are four hands-on prompt-engineering activities (based on suggestions by our AI):
(i) Fix the Prompt challenge. Start with, “Tell me about dogs.” How can you make it better using the five principles above? What do you get if you instead type, “please tell me about the different kinds of corgi dogs”?
(ii) Role-Playing. Ask the AI to: play the role of a historian and explain an event to you, or perhaps be a debate coach who can give both sides of an argument.
(iii) Creative Writing Prompts. Prompt your AI to write something in a style of your choice, then refine your prompts to get closer to your vision. Did you mention what you want in the way of style (e.g. hardboiled detective story), length, setting, character names and roles, core plot, atmosphere/theme, and ending?
(iv) Summarization challenge. Take a short article you have written and ask the AI to summarize it in a way you’d like to see (shorter; half as long; in one sentence; find the three most surprising facts and state them in an ordered list; etc.)
5. Environmental Equity and Sustainability
AI’s Impact on Education and Work
The integration of AI raises significant questions about equity in education and the workforce, becoming a long-term debate. For instance, how will AI reshape classrooms? Students will undoubtedly need to master AI tools for future careers, so bridging the gap from school to professional life is crucial. Organizations like Khan Academy already emphasize that AI is here to stay in education, encouraging effective learning and use. However, always respect your current teacher’s guidance as the local authority. Regardless, learning about AI use safely and responsibly outside the classroom is vital, as its presence will only grow.
Beyond education, AI’s potential to transform jobs is a widely discussed topic. While some predict AI may augment or displace certain lower-level roles, many also point to the creation of new opportunities as AI applications expand. So, even as AI offers incredible promise, such as helping doctors treat diseases, it will also bring disruption. The more you understand both the advantages and challenges of this technology, the better prepared you’ll be for your working life.
AI’s Environmental Footprint
It’s also important to consider AI’s environmental impact. Training and running powerful AI models consumes significant electricity, contributing to carbon emissions. The data centers that power AI require vast amounts of energy and water for cooling. However, AI can also be a tool for sustainability, helping optimize energy use, predict environmental changes, and manage resources more efficiently. Being aware of both sides of this equation is essential for understanding AI’s full societal impact.
6. Preparing for a Career that Uses AI
What if you don’t have immediate access to AI from home? How can you still engage with this powerful tool? We asked an AI that very question, and here’s part of its response: “Public libraries, your school, and local community programs are invaluable hubs for digital access and learning. Every step you take, even small ones, will build your knowledge and prepare you for the future.”
You can also try this: Find a way to access an AI—perhaps through a friend, teacher, librarian, or relative. Then, type in this prompt: “What if I’m in high school and may not have access to a phone or computer at home? How can I keep up with AI?” The answer will provide many ideas! Give it a try, even if you have access today.
High School Courses to Consider: If possible, take courses in these subjects to build a strong foundation for an AI-related career:
- Calculus
- Linear Algebra
- Probability and Statistics
- Computer Programming
- Language Arts
- Art
Beyond High School: To become instrumental in this field, consider pursuing college or university degrees in:
- Data Science
- Computer Science
- Artificial Intelligence
General Career Guidance: Regardless of your specific path, here are key steps:
- Build a Strong Math Foundation: Skills in Linear Algebra, Calculus, Probability, and Statistics are fundamental for most technical AI roles.
- Learn to Code: Python is the dominant language for AI and machine learning (with libraries like TensorFlow, PyTorch, scikit-learn), though R is also useful, especially for statistics.
- Hands-on Projects: Build a portfolio! Work on personal projects, participate in hackathons, contribute to open-source AI projects, or pursue internships. Practical experience is highly valued.
- Online Courses & MOOCs: Supplement your formal education with online courses from platforms like Coursera, edX, or Udacity. These can provide specialized skills quickly.
You can also use AI to explore career paths. For instance, ask an AI: “What degree levels will be required for AI careers?” or “What tasks within jobs will be replaced by AI?” (Be aware the answers might be quite detailed, so try them yourself!). You might find that some jobs less likely to be replaced by AI don’t even require four-year degrees.
Finally, remember that AI applications are often interdisciplinary. Follow your passion, as long as it benefits society, and see if you can figure out how to integrate AI into the vocation you’re most passionate about.
7. Glossary
- Artificial General Intelligence (AGI) – a hypothetical type of AI that has human-like intelligence and capabilities across a broad set of knowledge, rather than being trained for a specific task.
- Artificial Intelligence (AI) – the field of developing systems that are capable of mimicking human behavior and capabilities, and extending them.
- Feature – an individual, measurable property or characteristic of a data set used as input for a machine learning model.
- Gradient Descent – a process for minimizing a model’s error by iteratively adjusting its parameters, visualized as finding the lowest point on a surface.
- Large Language Model (LLM) – a type of machine learning model that is trained on a large amount of text data in order to understand and generate human-like language.
- Machine Learning (ML) – a subcategory of AI that uses algorithms to automatically detect patterns in data, and using those patterns to make decisions.
- Machine Learning Model – a set of parameters organized into a structure like a neural network that can be trained on data and used for analyzing data, predicting information, and generation.
- Neuron – a basic computational unit in a neural network, inspired by biological neurons, that receives and sends signals.
- Neural Network – a computational model inspired by the way the human brain functions.
- Supervised Training – a type of model training where the model is given both input data and the corresponding expected outputs to learn from.
- Token – a small part of your input to an AI computer system that is recognizable by that computer. Think: what molecules are to chemistry.
- Transformer – a type of neural network that is specialized for processing sequences (like text).
- Weight – an importance factor assigned to a connection between neurons in a neural network, determining how strongly a signal from one neuron affects the next.
We value your input! Have questions about this issue or want to know more about our foundation?
Feel free to reach out. We’re keen to hear your thoughts and feedback. Your voice truly matters to us!
We hope you have enjoyed this issue of our newsletter.
We look forward to preparing our next issue on a different topic.
Subscribe to the Newsletter!