The "huge, anomalous, unexplained surge in selling at about 2:45" has been attributed to a source in Chicago (believed to be single execution system that mistook a trade of $16 mn for $16 bn), which had apparently set off algorithmic trading strategies, which in turn rippled across everything, pushing trading out of whack and feeding on itself — until it started to reverse. Computers programmed to sell when prices hit a certain level did so en masse, prompting yet other computers to do the same, until the prices of some stocks were pushed down to nearly zero.
I have blogged earlier about high frequency trading (HFT), wherein very high-speed supercomputers with sophisticated programs use "real-time" share prices to detect or even predict the next instantaneous twitch in the stock market, and then buy and sell stocks within fractions of seconds. The traders using such algorithms, usually large Wall Street firms, make a tiny profit on the blip of price change of each share. In recent years, HFT has come to account 50-75% of daily trading volume, and more than 60% of trading in stocks listed on the New York Stock Exchange takes place on separate computerized exchanges.
See also Rajiv Sethi on HFT here.
Update 1 (9/5/2010)
Micheal Durbin argues that greater regulatory controls, software standards, and disclosure requirements identifying the originator of a trade can address the concerns with HFT.
Update 2 (2/9/2010)
See this inside story of the flash crash.
Update 3 (9/10/2011)
Movement picks up to regulate HFTs since they are percieved as being responsible for spurts opf market volatility and unfair to small investors.
Update 4 (17/10/2011)
Excellent speech by Andrew Haldane on HFT, which is discussed here by Alex Tabarrok. He writes,
"Cramming ever-larger volumes of strategic, adaptive trading into ever-smaller time intervals would, following Mandelbrot, tend to increase abnormalities in prices when measured in clock time. It will make for fatter, more persistent tails at ever-higher frequencies. That is what we appear, increasingly, to find in financial market prices in practice, whether in volatility and correlation or in fat tails and persistence."