Sneaky, clever bots. But not so brilliant as I’d thought.

I googled “Zero Mostel shat here” with the quotes.

“Speaking of Zero Mostel, we just threw out a toilet seat he had autographed. It said “Zero Mostel shat here” and had a cartoon sketch of himself. He “created” it when he played my father-in-law’s tent theater, loafing through Tevye, and still bringing the house down.”

It must have been in a blog comment. Only about 5 people in the world would know about that toilet seat and its history. I think the bot is mining your blog’s comments and matching keywords.

I figure someone took a dozen or more common subjects for blog postings and identified three or four keywords that could identify each one. Once you’ve categorized a blog, you build a string using a series of if/else statements that have searches for a second set of more specific keywords as their arguments.
As I said, it would be tedious, but a smart person could write a program that gives a very good approximation of a real post.

I play computer games. Most times, the bots just CHEAT.

They either use knowledge the user doesn’t have, speed the user can’t have, and/or have their “stats” enhanced to make them “recover” tougher or attack harder… So, if you presume any “character type” has a certain number of “hit points”, ‘z’, which recover a 1/second, and an attack with item ‘x’ causes ‘y’ damage, then the bot will have 1.5z hit points vs a player, recover hit points 2x faster, and/or either do yx2 damage or strike 1.5x as often… things like that.

Game AI has a long way to go to really get close to the singularity… and there’s at least as much impetus to make good game AI as there is to make better spambots.

My guess would be that the selection scripts have gotten much more sophisticated — say you start with 10,000 snippets of text, cross-correlated on keywords. The bot then looks at the primary post and makes best one based on the number of keywords appearing in it.

Neo: Keep in mind that you are actually seeing a potential “observation bias” in your selection, there. You’re noting all the hits, you have no idea how many misses came from the same bot(s), I’d bet.

That is — I can’t imagine you scanned the obviously bad ones to see if they pointed to the same sites as some of the good ones.

A guess, but natural language recognition is really NOT that easy a problem.

That was, BTW, one of the key things about Watson on Jeopardy… it was actually parsing the questions, not being told what to look for. And that’s a VERY controlled, specific, scenario. And it still got it laughably wrong sometimes, ignoring key words that made the answer provided notably clear — example: Final Jeopardy in the first game:

Watson was the only contestant to miss the Final Jeopardy! response in the category U.S. CITIES (“Its largest airport was named for a World War II hero; its second largest, for a World War II battle”). Rutter and Jennings gave the correct response of Chicago, but Watson’s response was “What is Toronto????”

See the wiki entry for guesses as to how it screwed that one up.

P.S., Watson won it handily, since it only wagered a small amount, and was up quite a bit at that point.

Now, sounds promising for spambots? Nope:

Watson is made up of a cluster of ninety IBM Power 750 servers (plus additional I/O, network and cluster controller nodes in 10 racks) with a total of 2880 POWER7 processor cores and 16 Terabytes of RAM. Each Power 750 server uses a 3.5 GHz POWER7 eight core processor, with four threads per core.

That probably doesn’t mean a lot to you non-techies, but it’s a bit more processing power than any spambot is likely to manage to have.

Most interesting.

Jamie Irons