Hacker News Reader

Comment by pg | Hacker News Reader

It's at the level of spam filtering: individual words in application fields. So the insights were the sort I noticed when I first tried statistical spam filtering-- e.g. that ff0000 was highly correlated with spamminess (because it's html red).

Incidentally, the past tense in the article is misleading. We're using that software now. It's the product of a company in the w2012 batch. I don't think I can mention the name because they're not launched yet, but we're very impressed by its performance. SV Angel is using it too.

View on HN

I'm wondering if it says more about the companies or the people judging them :)

Kidding aside, I'de be surprised if specific words correlated with the probability of success. So there's a good chance this is measuring only if there are key-words that please the reviewers.

I'm sure you've looked at prediction models for individual partners and compared them (or at least considered it, what with days only having 24 hours :)

It'd be fantastic to see you write about the results of that - maybe once the company has launched.

The reason this has me so curious? If the predictive models are not partner-specific, this could indicate that there are very specific things that successful companies/founders say and/or do. Sharing that might or might not mess with the models, but it has a good chance to increase the likelihood of success.

People have done simple word frequency analysis in various domains, and it's been effective -- from (housing listings, resumes, car listings) the takeaway I have is specific, quantifiable things seem to raise value, vague or weasel terms tend to lower value. I think the #1 word for housing listing value is "granite" (as in granite countertops), and various hedge/weasel words like "cozy, comfortable, ..." lower it.

Naively, I'd assume the same might be true of something like a YC app. Harj has said in many places that he turns to the "impressive achievements" section first (http://askolo.com/harj#4f74bc2be6c38a8e50000077). I'd assume that's fairly similar to a resume and would have the same statistical properties.

for those interested, a nice explanation of the correlation of words to value / weaseliness in real estate is covered in a chapter of the book Freakonomics.

Like anything in Freakonomics, I'd take that with a truck-sized grain of salt.

Curious: Are you referring to the book? I don't recall reading that in there... I guess I'll need to go back and take a look.

Chapter 2. (I googled "freakonomics real estate" and got a helpful google book search except. I love living in the future; Book Search and Google News are the best-executed Google products of the post-gmail period.)

but context matters: http://sentiment.christopherpotts.net/context.html#thwarted

Bayesian, or other algorithmic approaches (what many call AI), on signals like word frequency can help with separating wheat from chaff. Rarely any better when it comes to such multivariate decisions. Given the uncertainty in startup success, perhaps a rough solution at scale would perform. Maybe that's the goal with YC, but at least on the outside YC has always seemed more like an elite private school that delivers on personal treatment and alumni network. I wonder whoa is building the equivalent of the public university.

We also do measure the performance of partners reading applications. We don't actually use the data in any way except to motivate us though.

I'm surprised nobody mentioned this (http://www.google.com/m?q=bullshit+bingo ) yet, because what I hear, sounds like advanced version of it. Consider selling online YC-Branded cards.

Have you considered running it against Viaweb's first business plan?

http://www.paulgraham.com/vwplan.html

So a wise yc-spammer will take the published YC applications of accepted teams (Dropbox, etc.) as a starting point.

Not enough of a corpus, and not very helpful because the applications are manually reviewed anyway.

I just changed the tense, so it's clearer.

Reply on news.ycombinator.com