Monday, October 25, 2010

Five things Paul Graham is wrong about

I debated with myself for a long time over whether to write this post (which is one of the reasons some of the subject matter is a little old). I like Paul. I read his essays. I agree with most of them. I think Y-Combinator is a great achievement. I've invested, directly or indirectly, in about a dozen YC companies. I've even made some money doing it. And Paul's influence is now such that getting on his shit list can be a career limiting move. But some of the things Paul has said lately I find a bit disturbing. So I decided to go ahead and write this because, if nothing else, it will serve as a test to see if the Silicon Valley culture really does tolerate disruption as well as it likes to think it does.

That said, here are the five things I think PG is wrong about:

1. Efficiency doesn't matter

Paul has argued that machines have gotten so powerful that the efficiency of the programming language you use doesn't matter. And of course he's right about that to a certain extent: efficiency doesn't matter at first. But when you start to get traction, efficiency starts to matter a lot, and the more successful you are the more it matters. If you've done it right then you have machines doing most of your work. If you have machines doing most of your work, then the larger you grow the more the cost of those machines will come to dominate your expenses. Those costs are inversely proportional to your efficiency. If your code runs twice as fast you will only need half as many machines. All else being equal, those savings are pure profit.

Note that I am not arguing for premature optimization here. Premature optimization is still the root of all evil. But it's important to keep in mind that the aphorism of "efficiency doesn't matter" has a limited domain of applicability, and your goal, if you're starting a business, is explicitly to get out of that domain. If you've been in business for a while and efficiency still doesn't matter you have almost certainly failed.

2. The returns on angel investing are bimodal

Paul claims that a startup either wins big or fails, and that therefore if you're going to invest you should invest in as many companies as possible in order to increase your odds of picking the one big winner. He writes, "The expected value of a startup is the percentage chance that it's Google." We had all better hope that this is not true, because Google-sized wins only come along once every few years at most. In the meantime, 700,000 new businesses are incorporated in the U.S. every year. There had better be a couple of those non-Googles that produce a decent return for their investors or the country is in big trouble.

I've been angel investing for five years now. I have a net negative return because of one spectacular failure (from which I have learned a lot but that's another story). But if you discount that as an outlier, my net return is positive, and all of my winners so far have been small: 10x return or less. The key, as Paul seems to have recently discovered, is time. What matters is not absolute return, but return amortized over time. I'd much rather have a 10x return in one year than a 100x return in three (or five or ten, which are more typical numbers).

It is true that VC returns have historically been bimodal, but this is not inherent in the nature of startups, it's because VC's engineer the environment to force the distribution to be bimodal. They do this because they are constrained by the structure of their funds to move money in and out on a particular schedule. They are usually not free to reinvest the proceeds from an early small win into a new startup that might produce another fast small win. Because of this, if a VC cashes out early, that capital can no longer work for them. So they use their influence to force companies to continue to grow past the point where they otherwise would have. This in turn forces the companies to take on additional risks that they otherwise might not have, which causes them in many instances to fail.

Angels do not operate under these constraints, and so they can make a very nice overall return on small wins with fast turnaround. To be fair, Paul did grudgingly admit this possibility in his talk at the most recent startup school, but he doesn't seem to have thought the consequences all the way through. In particular, Paul still claims that:

3. Valuations don't matter

This is the one that worries me the most because the issue has been in the news lately, and because it is so plainly false. To see that it is false you only have to take it to an extreme: Would you invest in a YC company at a ten billion dollar valuation? Clearly not. So somewhere between zero and ten billion there is a line beyond which the expected return no longer has a high enough risk premium to make the investment worthwhile. One can reasonably argue about where that line is, but there are only four orders of magnitude between the $1M valuation that YC companies typically commanded not so long ago and the $10B which is clearly over the line. The last round of YC companies was getting $10M valuations, an order of magnitude over a few years ago, under the valuations-don't-matter mantra. You can only do that trick three more times before you get to $10B. Somewhere between here and there valuations must start to matter.

4. Investors who don't want to be the first movers are "assholes"

The exact quote (captured on video, about twelve minutes in) was, "You start with the committed ones, the nice guys who say, 'yes I'm absolutely in,' and work your way outwards to assholes who give you lines like 'come back to me to fill out the round.'" I have on occasion been one of those assholes, and I take umbrage.

I can certainly understand why a founder would find a line like that annoying. But calling such people "assholes" crosses a line and implies a dishonorable motive, which I think it grossly unfair. To invest in startups as anything other than a hobby is a boatload of work. Notwithstanding what I wrote above under #2, the sad fact of the matter is still that the vast majority of startups never achieve profitability. (I'm talking here about real profitability, of the sort that can provide a positive return to investors after all the company's employees start making reasonable salaries. "Ramen profitability", which YC companies like to advertise, simply means that the company is being subsidized by its employees. This is not a recipe for long-term success.) There are ways that you can improve the odds, even stack them in your favor, but they are, as I've said, a ton of work, and very, very hard to learn. Particularly for a new investor, following the lead of more experienced investors rather than just shooting from the hip can be a very reasonable strategy. And not just because it improves your odds, but also because experienced investors pay more attention to people who have co-invested with them. So even if you're treating your investments as an educational expense and not really looking to make money, it can still make sense to ask who else is in on the deal before you write a check.

So on behalf of reluctant angels everywhere I say to founders: cut us some slack. Many of us are still trying to figure this out as we go along just like you are.

Watching that video again reminded me:

5. You don't have to take notes

At the start of his talk at startup school PG admonished the audience not to take notes because a written version of his talk would be available on line and it would certainly be a higher-fidelity rendition of what he said than anything they could scribble in real time. First, it isn't actually true. The written version of Paul's talk is an edited version of what he said. The "asshole" comment, for example, doesn't appear. But more importantly it makes a false assumption about the purpose of taking notes. Taking notes serves not only to make a non-volatile record of what was said, but the act of writing helps you remember what was said. Writing activates different parts of your brain than speaking. The act of rendering words to a page and feeding those words back through your eyes rather than your ears actually helps you remember things. So taking notes can be worthwhile even if you never go back to look at them. And last but not least, if you take notes, you can write down your real-time reactions to what is being said, which obviously no one but you can do.

[UPDATE] This entry was, unsurprisingly, posted on Hacker News. The top-rated comment is a brief response form PG. For the record, I concede that I was being unfair in point #5. PG did not say not to take notes, he just said that people didn't *have* to take notes if they chose not to because a transcript would be available. To the extent that I have misrepresented or misunderstood PG's positions on this or the other points I raised, it was not deliberate.

18 comments:

ses said...

"Taking notes serves not only to make a non-volatile record of what was said, but the act of writing helps you remember what was said"

As a university student I can safely say never has a truer word been said. It frustrates me so much when some lecturers and peers ask you why you're making notes when they are making the information available online.

On a note more related to YC and Paul Graham - I found his most lucid writings those about his own experiences and advice to those thinking of starting a start-up. The more recent essays and talks I think focus too much on securing investment - there's more to business than this.

Anonymous said...

From my intro psych class, I remember a learning study where they had one group of students take notes and another set of students study an outline provided by the instructor. The latter group scored higher!

Anonymous said...

Remembering facts isn't the point of listening
to this sort of talk. The point is immersion in
the thought process of the speaker, to leave
a lasting emotional impression. That lasting
impression is what will motivate you to go
research what the speaker said, and to find
ways to integrate it into your own work and life.

John said...

I don't think any of the above criticisms is severe enough to detract much from the huge positive contribution Paul has made overall.

Regarding note-taking, it depends on what kind of learner you are. I never took notes as an undergraduate and always scored #1 in the class.

Now I'm a college professor and I'd say about 80% of my students take notes (even if there are notes distributed later on, which there usually are).

ses said...

"That lasting impression is what will motivate you to go research what the speaker said, and to find ways to integrate it into your own work and life"

Interesting view, however in that case I would expect such a talk to be really rather short. Generally however, most speakers who give this kind of talk like to spend some time on the subject and will often relate interesting information and anecdotes which are completely off-the-cuff and may not be included in any literature they provide.

Perhaps I have a gold-fish bowl type of memory but I find if someone is talking at length for more than about 10 minutes, unless I write some of it down I will inevitably forget some of the detail.

Anonymous said...

My most critical factors, when designing a platform, are speed of implementation, availability of talent, and the ability to swap out the slow parts later.

The reality is that servers are cheap, and people are expensive. As such, it often makes sense to use slow languages to start and it often makes financial sense to keep using them for quite some time.

And if your business is a high volume consumer website, then it probably makes sense to start swapping out the slowest bits as you scale. But that said, it's not really a key issue. It's an optimization.

Anonymous said...

I'm with you on the notes, there's a world of difference between reading a transcript and writing your own notes.

Anonymous said...

The point about non-commital sheep-like angel investors asking you to come back to them to fill out the round is actually in his essay. He refers to them as jerks instead of assholes in the essay. Your 5th item on note taking is just a really lame point, almost as if you desperately needed to come up with a 5th point for your link bait post.

M. Edward (Ed) Borasky said...

"I have a net negative return because of one spectacular failure (from which I have learned a lot but that's another story). But if you discount that as an outlier, my net return is positive, and all of my winners so far have been small: 10x return or less."

Sorry, but the numbers are what the numbers are. Spectacular failures can't be "written off" as "learning experiences" or "outliers." You have a net negative return. No excuses, no ifs, ands or buts. You pulled the handle and didn't get your money back. That's the way the world works.

And your distribution *is* bimodal - in fact, I'm guessing the sample size is so small that "statistics" are meaningless. You are mistaking luck for skill until you have a large number of outcomes to analyze. As N. N. Taleb put it, "Nobody counts the monkeys."

CurmudgeonlyTroll said...

"I'd much rather have a 10x return in one year than a 100x return in three"

You should prefer $100,000 over $1m 2 years later, only if you can achieve a 10x return in 2 years, in other investments of comparable risk.

Ron said...

> Sorry, but the numbers are what the numbers are. Spectacular failures can't be "written off" as "learning experiences" or "outliers."

If you're going to count all my outliers, I have one on the positive side of the ledger as well. If you count that one then my "spectacular failure" was a moderate loss, and everything else I've done is lost in the noise.

> You should prefer $100,000 over $1m 2 years later, only if you can achieve a 10x return in 2 years, in other investments of comparable risk.

Yes. That's why I said three years -- or five or ten, which is a more typical wait for a 100x return.

Kleiner-Perkins and Sequoia made ~1000x on their Google investment over five years. That's 4x per year annualized. I think that can safely be considered an upper bound on what one can reasonably expect from early stage investing.

Anonymous said...

You got misconception about note taking. Several different books about learning tell its just a sure way to FORGET things. Sure it activates different parts of brain, but that part is not about memory, more likely secondary task that takes away your attention from memory and real processing of input in away that helps memory.

Ron said...

That's news to me. Do you have a reference?

fred said...

thanks for writing this

two of the most dangerous "falsisms" in startup land today are the idea that startup returns are bimodal and valuations don't matter

Dave W Baldwin said...

Very good post. I'm from the Founder side of things and it is great to see thoughts from the investment side.

Ron said...

Thanks Fred and Dave for the kind words.

M. Edward (Ed) Borasky said...

'thanks for writing this

'two of the most dangerous "falsisms" in startup land today are the idea that startup returns are bimodal and valuations don't matter'

Sure, with a *huge* sample, returns might not be bi/multimodal, but for an investor who's only done a few dozen or so investments, that's quite possible. Are there actually datasets with *all* the points that one could study empirically? Or are all of these deals done behind doors closed enough that we only see what the dealmakers want us to see about their track records?

All I'm saying is with publicly traded stocks, futures and options on regulated markets, bonds, exchange rates and a few other markets, the information is there for all of us to see, analyze and model. The same is true for the hundreds of economic time series that governments collect and publish. For angel and VC investments we don't have that. So I'm not willing to rule out bimodal - indeed, I'm not even willing to rule out distributions that don't have some higher moments at all!

Ron said...

> Sure, with a *huge* sample, returns might not be bi/multimodal, but for an investor who's only done a few dozen or so investments, that's quite possible.

Of course. Sampling distributions can look very different from the underlying distributions, especially for small sample sizes. So?

> Are there actually datasets with *all* the points that one could study empirically?

Yes. In the U.S. the roster of companies started in any given year is public record, as is the roster of companies currently in existence, both public and private (with the exception of sole proprietorships, but those can be safely ignored). The financial status of private companies is not public record, but it's not too hard to backsolve for a large enough sample to get a pretty reliable picture of the underlying distribution. That it cannot be bimodal is easily seen from the simple fact that there are an awful lot of non-Google-sized companies out there, many of which are sustainably profitable.

> For angel and VC investments we don't have that.

Actually, for VC investments we do because the VCs voluntarily publish this information. VC returns are closer to bi-modal than the natural underlying distribution of outcomes because VC's force it to come out that way by vetoing modest (by their standards) exits.