Mini LLM

// 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.
const 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},
};

let text = prompt.toLowerCase();
const steps = [`Generating from "${text}":`];

// Generate up to 12 more characters
for (let i = 0; i < 12 && text.length < 20; i++) {
  const context = text.slice(-3);
  const probs = learned[context];

  if (!probs) break;  // Unknown pattern

  // Pick highest probability (greedy decoding)
  const [next, prob] = Object.entries(probs).sort((a,b) => b[1]-a[1])[0];
  steps.push(`  "${context}" → "${next}" (${Math.round(prob*100)}%)`);
  text += next;
}

steps.push(`Result: "${text}"`);
return steps.join('\n');