Online discussion threads are important means for individual decision-making and for aggregating collective judgments, e.g. the `wisdom of crowds'. Empirical investigations of the wisdom of crowds are currently ambivalent about the role played by social information. While some findings suggest that social information undermines crowd accuracy due to correlated judgment errors, others show that accuracy improves. We investigate experimentally the accuracy of threads in which participants make magnitude estimates of varying difficulty while seeing a varying number of previous estimates. We demonstrate that, for difficult tasks, seeing preceding estimates aids the wisdom of crowds. If, however, participants only see extreme estimates, wisdom quickly turns into folly. Using a Gaussian Mixture Model, we assign a persuadability score to each participant and show that persuadability increases with task difficulty and with the amount of social information provided. In filtered threads, we see an increasing gap between highly persuadable participants and skeptics.