Only start a company with people you've known for a while and worked on projects with before.
If you don't know anyone like that, you can manufacture them. Find good people and try working with them on projects that are not the startup. College and (especially) grad school tend to create these situations, but you can also make them happen deliberately, e.g. by collaborating on some open source project.
The chances of us interviewing a team of founders who'd built some fairly well regarded open source software together would be close to 100%.
Things like Node Knockout and Rails Rumble are a great way of identifying potential problems relatively quickly. From experience, it's high-stress and you're more irritable than normal because you haven't been getting enough sleep. It's very easy to get annoyed, even when you're doing it with the people you work with all of the time. You discover who you can work with well and who you can't pretty quickly.
I found my team by hanging around the CS lab at UC Irvine and reaching out to the ACM club pres at CSUF a while back. I am a business major that previously had two non-tech startups and I needed hackers in my corner for a project. Luckily, the help found turned into ongoing collaboration with a small diverse group. It helps that we are all upperclass/grad students that have previous experience. We all offer something different to the team and we tend not to argue so much because of this.
When PG says to try projects before a jumping into a company he’s saying: you should date your girl for a while before moving in together.
This generalization seems to ignore teams who work together over time - say a year or three - and while bootstrapping. Have you never seen a data point in that direction?
Teams thrown together for the sole purpose of YC seem likely to have poor records. But YC is a very, very special case. Bootstrapping takes time and effort that either deepens relationships or destroys them. I don't see a middle ground there, actually.
Have you ever looked at how long a team has known each other as a predictor of success/failure? It seems your philosophy assumes more is better. Is there any data to support that belief? What's the record for spouses?
A founding team who's already been working together for years or survived the bootstrapping experience match, or at least approximate, the criteria pg's looking for – a cohesive team that can handle stress well. My take is that the advice he gives is suited for people looking to start a brand new company during yc, not those who have already started and are applying to yc.
More importantly, I don't think pg's saying that length of relationship is a deciding factor. Rather, he's highlighting that you need to know you work well with your cofounders and the relationship – new or old – won't fall apart under stress and trial.
If you don't know anyone like that, you can manufacture them. Find good people and try working with them on projects that are not the startup. College and (especially) grad school tend to create these situations, but you can also make them happen deliberately, e.g. by collaborating on some open source project.
The chances of us interviewing a team of founders who'd built some fairly well regarded open source software together would be close to 100%.