Demonstrate the core mechanism behind LLMs by predicting the next token, appending it, and repeating.
Code
Algorithms# LLM is a next-token predictor. That's the whole trick.
# It looks at text, predicts what comes next, appends it, and repeats.
# These patterns simulate what a model "learned" from training.
# Real LLM learned from trillions of words - same idea, bigger scale.
learned = {
'hel' => {'l' => 0.9},
'ell' => {'o' => 0.9},
'llo' => {' ' => 0.6, '!' => 0.3},
'lo ' => {'w' => 0.5, 't' => 0.3},
'o w' => {'o' => 0.8},
' wo' => {'r' => 0.9},
'wor' => {'l' => 0.7},
'orl' => {'d' => 0.9},
'the' => {' ' => 0.9},
'he ' => {'c' => 0.4},
'e c' => {'a' => 0.8},
' ca' => {'t' => 0.7},
'cat' => {' ' => 0.6, 's' => 0.3},
'at ' => {'s' => 0.5},
't s' => {'a' => 0.6},
' sa' => {'t' => 0.7},
}
text = prompt.downcase
steps = ["Generating from \"#{text}\":"]
# Generate up to 12 more characters
12.times do
break if text.length >= 20
context = text[-3..-1]
probs = learned[context]
break unless probs # Unknown pattern
# Pick highest probability (greedy decoding)
next_char, prob = probs.max_by { |_, v| v }
steps << " \"#{context}\" → \"#{next_char}\" (#{(prob * 100).to_i}%)"
text += next_char
end
steps << "Result: \"#{text}\""
return steps.join("\n")Parameters
Starting text (try 'hel' or 'the c')
This Is How LLM Works
When you type "hello" to ChatGPT or Claude, here's what actually happens:
The Core Loop
- Look at context: Model sees "hello"
- Predict next token: Based on training, " " (space) is most likely
- Append it: Now we have "hello "
- Repeat: Predict again → "world" is likely after "hello "
That's it. The entire conversation is generated one token at a time.
Why This Matters
When LLM writes a 500-word essay, it's not "thinking" about the whole essay. It's just predicting the next word, over and over, 500+ times. The apparent intelligence emerges from patterns learned during training.
The "Learning" Part
This demo uses hardcoded patterns. Real LLMs learn patterns from massive datasets:
- After "hel" → "l" appeared millions of times → high probability
- After "hello " → "world", "there", "!" all common → model learns distribution
The architecture (attention, transformers) is just a sophisticated way to learn these patterns from data.
More Ruby Snippets
Array Difference
Find elements in the first array that are not present in the second array.
Array Frequencies
Count how many times each value appears in an array and return a frequency map.
Array Head
Get the first n elements of an array.
Array Intersection
Find common elements that exist in both arrays.
Array Tail
Get the last n elements of an array.
Array Union
Combine two arrays and remove duplicates to produce a union.