decoding
L
L
M
3 letters reshaping how humans think, work & create
scroll to decode ↓
scroll
what those three letters actually mean
L
large
175B+
adjustable numbers (parameters) — more than neurons in a human brain
L
language
2T+
words from books, web, code — every language humans write
M
model
f(x)
a mathematical function trained to predict the next word
scroll
L → LARGE
LARGE
0 parameters
that's bigger than the number of stars in the Milky Way
175B
GPT-3 parameters
the model that shocked the world
the model that shocked the world
86B
neurons in
the human brain
the human brain
100T+
GPT-4 estimated
parameters (rumored)
parameters (rumored)
LANGUAGE
English
"The quick brown fox jumps over the lazy dog"
Python
def fibonacci(n):
return n if n <= 1...
return n if n <= 1...
Spanish
"En un lugar de la Mancha..." — Cervantes, 1605
Math
∫₀^∞ e^(−x²) dx = √π / 2
Arabic
إِنَّ مَعَ الْعُسْرِ يُسْرًا
JSON / Code
{"model":"gpt-4","role":"user","content":"..."}
SQL
SELECT * FROM knowledge
WHERE year > 2020
WHERE year > 2020
HTML
<div class="hero">
Hello, world!</div>
Hello, world!</div>
Trained on 2+ trillion words — books, Wikipedia, GitHub, Reddit, arXiv, CommonCrawl
MODEL
INPUT
"the cat sat on"
"the cat sat on"
→
175B weights
f(x)
next-token prediction
→
"the" 8%
"a" 5%
"mat" 61%
"a" 5%
"mat" 61%
"2 + 2 ="
→
"4"
"Translate to Spanish: Hello, echobash"
→
"Hola, echobash"
"Write a bio for @echobash — developer & builder"
→
"echobash builds tools that make tech accessible..."
"Why is the sky blue?"
→
"Rayleigh scattering causes..."
before all this...
engineers tried to write rules for every possible conversation
if "hello" in msg: reply("Hi there!")
elif "weather" in msg:
if "today" in msg: reply(get_weather("today"))
elif "tomorrow" in msg: reply(get_weather("tomorrow"))
elif "next week" in msg: reply(get_weather("next_week"))
else: reply("I don't understand the timeframe")
elif "order" in msg:
if "pizza" in msg: # 2,400 more conditions below...
if "large" in msg: # 8,100 more...
# ... 47,000 lines of if-else later ...
ERROR: UnhandledInputException
input: "what's the deal with airline food?"
⚠ real language is infinite — rules are finite
you can't write enough if-else statements to cover everything a human might say
the 72-year road to ChatGPT
1950
Turing Test proposed
Alan Turing asks: "Can machines think?" Sets the target.
1966
ELIZA — rule-based chatbot
Simulated a therapist. Impressive at first. Falls apart fast.
1990s
Statistical language models
Count word pairs (n-grams). No rules — just probability. Better but shallow.
2003
First neural language model (Bengio)
Words represented as vectors. The seed of modern NLP.
2012
AlexNet — deep learning explosion
Neural nets crush image recognition. The field wakes up.
2013
Word2Vec — king − man + woman = queen
Words become points in space. Meaning = position. Mind-bending.
2017
"Attention is All You Need" ← the paper
Transformers replace RNNs. The architecture powering every LLM today.
INFLECTION POINT
2018–20
BERT → GPT-2 → GPT-3
Scale explodes: 117M → 1.5B → 175B params. Emergent abilities appear.
2022
ChatGPT — 100M users in 60 days
Fastest product adoption in history. Everyone is now an AI user.
WE ARE HERE
June 12, 2017 · Google Brain
research paper
ATTENTION
one word that changed everything about how machines process language
then something unexpected happened
as models got bigger, new abilities appeared — nobody programmed them in
GPT-1
GPT-1
117M · 2018
GPT-2
GPT-2
1.5B · 2019
GPT-3
GPT-3
175B · 2020
PaLM
PaLM
540B · 2022
GPT-4 ✦
GPT-4 (estimated ~1T+)
~1T+ · 2023
✦ emergent abilities: at ~100B params, models suddenly could do multi-step reasoning, write code, pass professional exams — without being explicitly trained to
November 30, 2022
0M
users in 60 days — the fastest product adoption in history
🟣 Claude (Anthropic)
🔵 Gemini (Google)
🟢 GPT-4 (OpenAI)
🟤 Llama 3 (Meta)
🔶 Mistral
⬛ DeepSeek
🟡 Gemma (Google)
🔷 Command R (Cohere)
"We didn't predict when it would happen — but we knew this was coming."
now you know what LLM means
go deeper. pick a concept.
Every interactive tool below builds on what you just learned.
No prior knowledge required — just curiosity.