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.
$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],
];

$text = strtolower($prompt);
$steps = ["Generating from \"$text\":"];

// Generate up to 12 more characters
for ($i = 0; $i < 12 && strlen($text) < 20; $i++) {
    $context = substr($text, -3);
    $probs = $learned[$context] ?? null;

    if (!$probs) break;  // Unknown pattern

    // Pick highest probability (greedy decoding)
    arsort($probs);
    $nextChar = array_key_first($probs);
    $prob = $probs[$nextChar];
    $steps[] = "  \"$context\" → \"$nextChar\" (" . (int)($prob * 100) . "%)";
    $text .= $nextChar;
}

$steps[] = "Result: \"$text\"";
return implode("\n", $steps);