Paul Graham Says Y Combinator Is Pickier Than Ever, ‘Hardly Any’ Bad Startups

Or bad means "bad bet" (before the fact) instead of "unsuccessful bet" (after the fact).

Or he was attempting what hu-mans refer to as "humor" :-)

This is a good and interesting question.

Since YCombinator depends on finding Dropbox and AirBnB scale startups, I imagine one of the most important metrics to them is whether they've rejected a business that went on to achieve that scale.

Of course being in YCombinator can increase a venture's chance of hitting that type of success, so keeping track of who they rejected is just a proxy for how well their applications strategy is. They may have rejected a company that _could have become_ AirBnB or Dropbox. But this is less important since in this case that would indicate that this ability lies with YC, not the startup, and so it's less important which companies they pick. If true, this would run counter to the founder-centric model they profess, so I'm skeptical it's how they feel.

This question of who to let in reminds me of the tradeoff of precision and recall in IR. A large class can reasonably be termed higher recall, while a class where they actively exclude companies with predictors of failure increases the precision. THe two metrics are inversely related, but it's possible to have high precision and recall if you're only looking for a very few number of things and you pick them all.

It seems to me that culling out likely failures would only make sense if the partners had determined that the presence of these likely-to-fail companies negatively impacted their ability to help other companies reach their potential. This seems a likely rationale to me based on PG's comments about how dying startups take up so much of their time.

So what it really seems to represent is a vote of confidence in the partners' ability to help startups increase their chance of massive growth. And/or a recognition that likely-to-fail startups have deleterious effects on the rest of their batch that outweigh the likelihood that they'll be successful outliers. Either way it doesn't seem to have much impact on companies that apply.

It will be interesting to see if PG writes a How Not to Apply essay. Such an essay might destroy the predictive value of these behaviors, but if it stops the behaviors and they were causally linked to bad outcomes then it's a net gain.

I think your definition of "bad" implies a high likelihood of failure. Whereas I'd say that a "bad" startup is one with a low growth or profit potential; my guess is that pg's definition is similar. A bad startup can "succeed" and a great startup can fail -- they're totally unrelated things, IMO.

Much like the Google guidelines on SEO, I would imagine pg's response to this would be something like 'the best way to optimize your application is to ignore optimizing your application' and instead focus on building something awesome.

I do agree they should be consistent but in their defense, one of the above is an actual quote while the other is not.

Hey, I'm the writer who wrote the second example cited here -- very good point in showing the difference between the two, so I wanted to try and clarify.

When Graham said the "hardly any" line on stage, it drew a pretty big laugh from the audience -- it seemed like a bit of humor on his part. The earnest meaning and connotation I took from that (and comments Graham gave off-stage) was that there were no obvious weak links in this batch.

If that is the best example you can find of this theoretical phenomenon, I don't see why you'd bother dredging up such a load of nothing.

I think he was just giving an easy to understand example which didn't really reflect reality. The problem with predictors of failure, unlike success, is that it's a lot easier to hide them once you know what they are. If you know, for instance, that having founders who when asked say they're doing a startup because they hate having a boss, no non-idiot will give that as an answer when asked, even if they believe it, ruining the predictor.

It could simply mean that they never spent a lot of time investigating how the founders got along, and so weren't able to notice initially which founders ultimately didn't work together well.

People have been predicting the pop for a while now, I don't think that YC being picky (< 2% of applicants make it) is a sign of the apocalypse.

They hit scaling problems, so what? It's not indicative of a larger problem.

What makes you think we'd a bubble in place? And if we had then what makes you think it is going to pop?