The world's best AI minds, made for absolute beginners.
Welcome to LAPA Deep Dives. We've gone looking for the best AI lectures on the planet — from researchers like Andrej Karpathy, Andrew Ng, Geoffrey Hinton, Yann LeCun, and Fei-Fei Li, and labs like Stanford, MIT, OpenAI, and DeepMind — and distilled them into something anyone can actually sit through and understand.
What makes this different
These lectures are usually built for engineers and PhD students. We've translated the ideas — things like large language model training, transformer architecture, reinforcement learning from human feedback, and neural network backpropagation — into plain language. No coding. No math background. No prior ML experience needed.
How this works: Watch the full lecture first (it's short, and fully credited to its creator), then work through cards that break down and reinforce the key ideas.
Who You're Learning From
AK
Andrej Karpathy
Founding member, OpenAI · Former Sr. Director of AI, Tesla · Stanford PhD
Andrej Karpathy earned his PhD at Stanford, was a founding member of OpenAI, and later led AI efforts at Tesla as Senior Director of AI. He is widely regarded as one of the clearest teachers in the field, known for breaking down deeply technical AI concepts into intuitive, general-audience explanations without dumbing down the substance.
Watch First
Intro to Large Language Models
Lecture by Andrej Karpathy — Stanford PhD, founding member of OpenAI, former Senior Director of AI at Tesla. Originally presented at the AI Security Summit, 2023.
Full credit to Andrej Karpathy for this lecture. Watch as much or as little as you like now — the cards ahead will walk you through every core idea either way, so you can also treat this as a reference to revisit after.
Why This Matters
You've used ChatGPT or Claude. But do you know what's actually happening in between?
You've probably got a decent instinct for what a "prompt" is, and what these tools can and can't do. But have you ever wondered what's really happening in the seconds between hitting send and getting a reply?
By the end of this deck: you'll have the same mental model AI researchers use with each other — translated so you don't need a computer science degree to follow it. Most people who use AI daily never reach this level of understanding.
Core Idea 01
It's just two files
Strip away the hype, and a large language model comes down to two things:
01
A parameters fileA giant list of numbers — the "knowledge" the model learned. For a large model, this can be hundreds of gigabytes.
02
A run fileA relatively small program that reads those numbers and uses them to predict text, one piece at a time.
Core Idea 02
Analogy
Imagine summarizing an entire library into one very dense cheat sheet.
To create that parameters file, the model reads a huge chunk of internet text and compresses it into a dense list of numbers. You lose the exact wording of every book, but you keep the patterns — grammar, facts, how ideas connect. That's roughly what training does.
Core Idea 03
Why only a few companies can do this
This compression step is extremely expensive. It requires large clusters of specialized computer chips running for weeks, costing millions of dollars for the largest models.
Why it matters: this is why the "base model" stage of AI is dominated by a handful of well-funded labs (OpenAI, Anthropic, Google, Meta) — while almost anyone can build useful things on top of a trained model.
Core Idea 04
A raw model doesn't "chat" — it completes
Fresh out of training, a "base model" doesn't behave like ChatGPT. Ask it a question, and it might just continue the sentence in a plausible way — because all it has ever learned to do is predict the next chunk of text, not have a conversation.
Key distinction: "sounding right" and "being helpful" are two completely different skills — and the base model only has the first one.
Core Idea 05
From base model to assistant
To turn a base model into something like ChatGPT or Claude, labs run a second, much smaller training stage: fine-tuning.
01
Curated examplesHumans write thousands of high-quality "question → ideal answer" pairs, showing the model what a helpful response looks like.
02
Behavior shiftThe model isn't taught new facts here — it's taught a new behavior: answer helpfully, follow instructions, stay in character as an assistant.
Core Idea 06
Simplified loop
Humans rank the answers (RLHF)
1. The model produces several possible answers to the same question. 2. Human reviewers rank them from best to worst. 3. The model is nudged to produce more answers like the top-ranked ones.
Core Idea 07
Why it "hallucinates"
Because the model is fundamentally predicting plausible next words — not looking facts up in a database — it can produce confident-sounding statements that are simply wrong.
Practical takeaway: treat AI output the way you'd treat a very well-read colleague speaking from memory, without notes — usually right, sometimes confidently wrong. Verify anything that matters.
Core Idea 08
Modern models can reach for tools
Newer systems don't just rely on what's baked into their parameters. Mid-answer, they can call out to a search engine, a calculator, or an image generator, then weave the result back into their response.
Core Idea 09
Think of it as a new kind of operating system
Karpathy's own framing: the LLM acts like a "kernel process" of a new type of computer — it can read, write, reason, and now call on external tools, the way an operating system coordinates programs and hardware.
Why this reframe is useful: it explains why AI is showing up everywhere at once — it's less "one product" and more "a new layer" that other software gets built on top of.
Core Idea 10
Fast instinct vs. slow reasoning
Borrowed from psychology: System 1 is fast, instinctive pattern recall. System 2 is slow, deliberate reasoning. Most everyday chatbot answers are System 1. Newer "thinking" models attempt to give AI a System 2 — deliberately working through steps before answering, which is why they take longer.
Core Idea 11
The honest risks, plainly explained
Because LLMs respond to whatever text they're given, they open up new categories of risk:
01
JailbreaksCleverly worded prompts designed to push a model past its intended limits.
02
Prompt injectionHidden instructions smuggled into content the model reads, trying to hijack its behavior.
03
Data poisoningDeliberately planting bad content in training data so the model learns the wrong thing.
Quick Check
Q1. What are the two things that make up a large language model?
Q2. Why can an LLM confidently state something false?
Q3. What does fine-tuning actually change in the model?
Arcade Break
LLM Fact or Myth?
Quick reinforcement round before you move on — tap FACT or MYTH as fast as you can.
II
Fact or Myth: Speed Round
Tap FACT or MYTH before the timer runs out. Build a combo for bonus points.
Fact or Myth
SCORE 0
20s
—
0
points
Apply It
Your Turn
Think back to your last conversation with ChatGPT or Claude. Was your prompt asking it to recall something (pretrained knowledge) or to reason through something new? Write one line reflecting on which one it was — and whether that changes how much you'd trust the answer.
Deep Dive Complete
You now understand AI better than most daily users ever will.
You've covered how LLMs are trained, why they sometimes hallucinate, how fine-tuning and human feedback shape their behavior, and where this technology is headed — all without writing a line of code.
Want feedback on your thinking? Optional
This is an add-on, not required to complete the Deep Dive. Submit your "Your Turn" reflection to a live LAPA tutor and get real feedback back by email.
Put It To The Test
LAPA Arena · compete.aiesports.org
You just leveled up your AI understanding. Now prove it.
The LAPA Arena is our ongoing AI skills tournament — real challenges, real scoring, real prizes. Go head-to-head with other learners and see where you actually rank.
More Deep Dives coming soon — next up: how neural networks actually learn.