Cuil

CuilのCEOがユーザに暴言を吐く–"気に食わなきゃ黙ってろ!"

次の記事

Twitterからデベロッパーへ贈り物―サードパーティーのデータストリーム取得制限撤廃へ

Cuilは、期待された検索の新星だったが、その出だしはシリコンバレーの歴史上もっとも惨(みじ)めだった。そして今のCuilは、車にたとえると、走行中に車輪が外れつつあるのではないかという気がする。

4月8日に同社は、cpediaという新しいサービスを立ち上げた。それは、単純に検索結果を返すのではなく、クェリに関する記事を自動的に作る。その記事はかなりへんてこりんだが、まだ実験段階だから、こんなものかもしれない。実験としてはなかなかおもしろいが、ビジネスとしてはどうかな。Wikipediaにない項目を調べるには便利だろうし、今後何年もかけて磨かれていけば、まともなものになるかもしれない。

しかしこのサービスに関するフィードバックの中には、否定的なものもある(GigaOmのMathew Ingramの記事はこんな意地悪なタイトル: 検索で失敗したCuilが今度はWikipediaの物真似で失敗)。そしてCuilのCEO Tom Costelloは、あらゆる否定的な記事に対して暴言で応じるという、愚かしい態度に出た。

彼のブログ記事の書き出しはこうだ: “これはすごい。今日は憎しみの言葉がそこらにあふれかえってるな”。そしてcpediaを批判したDave ParrackやFarhad ManjooやLouis Grayらを: “CpediaはWebの上にいっぱい書いてるけど全然人気のない人たちと相性が良くないようだ”と斬り捨てる。それから彼はこのサービスについて詳しく説明し、そして最後に、cpediaの生み出す結果が気に入らなかった人は黙ってろ、とめちゃめちゃひどいことを言っている。しかもその部分は“詩”になってる。

以下に、彼の記事の全文を引用しよう。取り下げられる可能性もあるからね。

〔訳注: 以下のブログ記事は英文ママ(Cpediaのアルゴリズムについて述べているがとても長い)。最後の“詩”の部分だけ訳してみましょう。〕

とても賢い老フクロウが樫の木に座っている
たくさん聞けば聞くほど彼は寡黙になった
寡黙になればなるほど彼はたくさん聞けた
みんなこの賢い老鳥のようになろうよ

——————————————————————–

Cpedia and its Detractors
Apr 13, 10:20 AM

Wow, 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.

[原文へ]
[米TechCrunch最新記事サムネイル集]

(翻訳:iwatani(a.k.a. hiwa))