“What sorcery is this?!”
That was my exact thought the first time I saw ChatGPT generate an eerily human-like response in seconds.
It felt like a miracle— and still does, but when I dug deeper I realized it’s not magic… it’s math. (I know, a little less exciting, but stay with me.)
What if I told you that understanding A.I.—even just at a basic level—could make you a lot more effective in your business, content creation, and marketing? You don’t need a PhD. You just need a simple framework for how it works and how to use it like a wizard. 🧙♂️
Yesterday I spent over 5 hours learning from top A.I. expert, Andrej Karpathy and decided to share with you the best of what I learned so you can have the goods quick.
How A.I. is Made (In Simple Terms)
Imagine you could download the entire internet, smash it into one massive text file, and then compress it into something much smaller—like turning an entire library into a series of emojis. 📚➡️😂🎉🔥
That’s what happens with A.I. models.
- Download the Internet – Well, a huge chunk of it.
- Clean the Data – Remove junk, duplicates, and nonsense.
- Turn Words Into Code – We’re talking (binary) a string of 1’s and 0’s so the computer can understand.
- Find repeating patterns of those 1’s and 0’s and chunk them together into shorter numbers, so you retain the pattern but lose the length. The example Andrej used was perfect. He said “Imagine a text file filled with emojis.” Now you have a smaller file, but one that still represents the order of text in your O.G. Internet file. These emojis are tokens. (you can use this app to get a feel for what turning words into tokens is like)
- Train a Neural Network – Now you give this file to a powerful computer (neural network) and its goal is to analyze the pattern in the emojis learn all the statistical relationships between each of the symbols. This is where the magic happens. The A.I. guesses patterns and relationships, makes mistakes, adjusts, and repeats… millions of times.
- Ex: If smiley face emoji is used there’s a 50% chance that the laughing emoji will come next.
- Find the Best “Settings” (Parameters) – Through a process of trial, error and adjustment, over and over, the computer eventually discovers a combination of settings, that when applied to guessing “what comes next” has the highest accuracy rate, using the training data as a guide.
- Voilà, You Have A Base Model – This is the “raw intelligence,” a giant math file that predicts what word (or emoji) should come next.
At this point, A.I. still isn’t smart. It’s just statistically accurate.
Want it to be even better? You tweak it using:
- Supervised Fine-Tuning (Humans give it examples)
- Reinforcement Learning (A.I. continues to figure things out on it’s own through trial & error)
- Human Feedback (We judge A.I. generated examples)
The result? A model that feels like it has a personality, but is really just a very sophisticated guessing machine.
A.I. Isn’t a Fortune Teller – It’s a Probability Machine
Ever notice that when you ask the same question twice, A.I. sometimes gives different answers? 🤔
That’s because it doesn’t “know” the answer. It’s just guessing the most likely next words based on statistics.
The key takeaway: A.I. is incredibly powerful, but it’s not always right. Treat it like an ultra-smart intern, not an infallible guru.
Words X Math = What an LLM gives you
Take that kids! Turns out math actually IS good for something useful. 😉
How to Use A.I. Like a Wizard 🧙♂️
Want to get better results with A.I.? Follow these tips:
🔥 Use a “Context Window” Effectively –
Andrej said ‘the A.I. will better ‘remember’ common info from the internet and more vaguely remember less common info... its a finite and precious resource and this is the only way the models can learn after training.’
Here are some of his recommendation.
- Start a new chat when starting a new task. This will clear out all prior context so your new output will only be affected by what you want to affect it for the task. It will also make it work faster.
- Give it examples of the types of output you’re looking for. The nerds call this “Multi shot prompting.”
- If you want it to summarize something, upload what you want summarized vs. hoping it will rely on its training data and remember what it read.
- Give It Step-by-Step Instructions – A.I. struggles with complex, multi-step tasks unless you guide it through them logically.
- You can also say things like, “let’s think through this step by step.”
- With ChatGPT, if you say “Please remember this” it can store data for future reference. It will sometimes do this automatically, but you can specifically put things in the memory by prompting it. This is special memory stored outside the language model that will be available from chat to chat regardless of what’s in the context window.
⚡ Test Different Models – Not all A.I. is created equal. Some models (like GPT-4 & Grok Think mode) are better at reasoning, while others (like Claude) excel at computer code.
Did you know you can have Claude create little apps for you like a flashcard program for learning OR a visualization of a style sheet for a website or app so you can see what it will look like?
🔎 Use A.I. for Research, But Verify Facts – Language models can summarize books and articles, but double-check key details (especially dates and statistics).
🛠 Leverage A.I. Tools for Different Needs:
- Text & Chat: Grok, ChatGPT, Claude
- Images: DALL·E, Mid-journey, Grok or Ideogram.ai
- Video: Sora, Veo 2
- Automation: Custom GPTs (A GPT is just a chat that you’ve pre primed with certain prompts, so you don’t have to keep copy pasting)
📈 Want the best math-based answers? Tell A.I. to “use code” instead of guessing. This forces it to run calculations properly.
🧠Your New Invisible Council: Napoleon Hill wrote about having an invisible council of mentors. Andrej Karpathy said he has a council of LLM’s. 🙂 He also said know which model you’re using and pay attention to the different types of results they’re giving you. Ex: Thinking/ reasoning models thinks through steps, similar to how humans break down complex problems. This will likely be better for problems that require multi steps to arrive at a unique and well thought out solution.
🌏 Mind Maps – When learning new or complex topics you can say “Create me a mind map” or conceptual library to help you understand complex topics or charts so you can visualize relationships of data. This can be helpful especially if you’re a visual learner.
🗣️🖼️📹 ✍️ Change up the way you talk to it. With Grok & ChatGPT you can talk with voice & it will talk back. ChatGPT has a video input where you a show it things in the real world and it can see and understand. Most all A.I. support image input.
BONUS TIP: Andrej said that when he uploads an image, especially one with data like a nutrition label, he’ll ask it to describe the image to make sure it’s seeing it correctly.
Final Thoughts + Final Mind Blowing A.I. Trick
Right now, A.I. feels a bit like a land of chaos, excitement and never ending stream of new ideas and tools to check out.
I think if we can understand, at least on a basic level, how they work in theory, we’ll be better at working with all the models. That’s why I was excited to peek under the hood and share what I learned in this post with you.
If you have the time and want to deep dive, the 2 videos I watched were long but very insightful. I’ll include them below if you want to check them out.
Ok, before I close out… I have to share the last thing I learned that really blew my mind.
With Notebook.Lm from google, you can take a piece of content and turn it into a podcast with 2 people discussing the ideas.
I’m about to test it out with this post. :-). 🔻
(this podcast episode was 100% A.I. generated in a few minutes) 🔼
Thanks for reading and whatever you do, always go for your dreams,
Paul
PS: If you’re uploading YouTube videos or placing Facebook ads, congratulations, A.I. is working in your favor whether you know it or not! You can learn more about online lead generation in this residual income optimized training system.
PPS: Here are the 2 videos I mentioned above.