Something tells me the wheels may be coming off the car at Cuil, the once promising search startup that had one of the most miserable launch debuts in Silicon Valley history.
On April 8 the company launched a new service, cpedia, which creates automated articles about queries instead of simply returning search results. The results are sort of strange, but as an experiment it certainly has legs. As in, I find it interesting, not necessarily that it can be a business. For things not covered well by Wikipedia, it can be useful and I can imagine, years from now, after many iterations, well, you get what I’m saying.
Some of the feedback on the new service wasn’t so positive (GigaOm’s Mathew Ingram wins best title with Cuil Failed at Search, Now Fails to Copy Wikipedia). And Cuil CEO Tom Costello has now made the extremely unwise decision to lash out at, well, just about everyone.
He begins his blog post with “Wow, the haters are out in force today” and specifically calls out Dave Parrack, Farhad Manjoo, Louis Gray, who have been critical of cpedia. “Cpedia does very badly with people who write much more on the web than people write about them,” he snipes. He then goes on to explain the service more fully, and ends, with spectacularly poor judgement, by suggesting that people who don’t like the results they see from cpedia simply shut up. Via a poem.
I’ve copied the entire post below, because it may not be around for long.
Cpedia and its Detractors
Apr 13, 10:20 AMWow, the haters are out in force today. We launched our alpha version of Cpedia last week, and have gotten a little feedback. Some of it was positive, a lot of it questioned whether or not we deserved to live (consensus: no). I was a little surprised at how vituperative things were.
So, I’d like to review some of the top issues we’ve heard about.
First up, Cpedia does very badly with people who write much more on the web than people write about them. Given the 1 billion people on the web one might think this unlikely, but it happens. When we try to summarize the information mentioning these people, we run into a problem. Almost none of it is about them. It’s about random things they have opined on. Dave Parrack, Farhad Manjoo, Louis Gray, I’m talking about you.
Another complaint was that we have stolen, plagiarized, looted, thieved, etc., the information we were providing. People were shocked that all the sentences came from other sources. Yes, all the sentences come from other sources, and we have links to exactly which sources they come from. We do not have a vast array of tiny sensor bots collecting information across the globe on all topics (I’m not even quite sure how this would work for the past). We crawl the web, and use bits of web pages, citing every bit.
A third complaint was that our machines did not seem to really understand the material. People complained of rote recitation, rather than an in-depth understanding. It was ever so. As a child I was made to learn Irish. The Christian Brothers believed in a Platonic theory of learning, where all knowledge was recollection, so they would beat us with leather straps until we “remembered” our Irish vocabulary (this actually works). I, however, could never get full marks, no matter how well I remembered, because my Irish, while technically correct, had no “blas”.
Blas, for those of you not from the West of Ireland, is the polish a hurley gets from the sliothar when used by a player of unusual skill, a patina on the surface of the wood testifying to the depth of talent of the player that had used the stick. Fair enough. Cpedia does not have blas – it’s a machine.
I think a lot of commentators had a misunderstanding about what machines can do, and what Cpedia is trying to accomplish. Cpedia is not an attempt to build something that knows all current knowledge and can write a meaningful essay on any topic – that would be a stretch goal. Rather, we are trying solve a much simpler problem. When people search the web for information, a lot of times the first few results do not contain all the information there is about the subject. Almost no one can continue through all the other pages, because they are almost all regurgitations of the same material, with perhaps a few extra nuggets. Cpedia processes all the pages about a topic, and extracts the unique ideas.
You can see the ideas that we have considered similar when you look at the sources page. There you see how we have collapsed away many other similar sentences to the sentences we show.
Because we remove all this duplication, unique ideas have more chance of coming to the top. We organize the ideas into rough clusters so that information about the same topic is close together.
I’m sorry if people were expecting Skynet. I can understand how it would be upsetting to get psyched up for life, the universe and everything, only to get a different UI on a search engine.
Cpedia has errors. That is intentional. We have tried to be inclusive, and dredge to the bottom of the web. This is great sometimes. For instance, I was meeting a VC recently, and I was able to discover that he has a tendency to over-imbibe. He doesn’t have a Wikipedia page (not that Wikipedia is in the business of random slander) and his firm’s bio on him elides this trivia. If we did not seek out the lower reaches of the web, I would have missed his one redeeming feature.
But bottom-feeding does ensure that we will have mistakes. Some are just algorithmic. For instance, our anaphora resolution (working out who a “he” refers to) is wrong about one in 20 times. This is on par with the accuracy of Hobbs’ algorithm. We sometimes get two senses confused. For example, we have made several pages about Django – one about Django Rheinhart and one about the python framework named after him. There is a little cross bleeding here, as the distinction is not as sharp as it might be. The notion of the identity of objects is not a settled matter – the classic example is Theseus’s Ship, but a simple one is of an axe: if you replace the handle twice, and the head once, is it still the same axe?
The promise of Cpedia is that you will find information that you might otherwise miss. It often works for me. Your mileage will vary. If you find that the page about you is completely random, the only advice I can offer is a poem my six year old recited at breakfast:
A wise old owl sat in an oak,
The more he heard, the less he spoke,
The less he spoke, the more he heard,
Why aren’t we all like that wise old bird.
Comment