We study how retail investors form opinions, using large language models to classify nearly 100 million timestamped messages from a major investing social-media platform. Strategy adoption is highly dynamic, shifting between fundamental (FA) and technical analysis (TA) in response to news, past performance, and social feedback. Message informativeness varies sharply across strategies: FA sentiment positively predicts returns and identifies the informed component of retail order flow, whereas TA sentiment negatively predicts returns and captures the behavioral component. TA sentiment also predicts intense Robinhood buying and short squeezes followed by sharp return reversals, with both patterns amplified by social engagement.