Karim Lakhani of Harvard Business School (and a Berkman associate, and a member of the Harvard Institute for Quantititative Social Science) is giving a talk called “How disclosure policies impact search in open innovation, atopic he has researched with Kevin Boudreau of the London Business School.
NOTE: Live-blogging. Getting things wrong. Missing points. Omitting key information. Introducing artificial choppiness. Over-emphasizing small matters. Paraphrasing badly. Not running a spellpchecker. Mangling other people’s ideas and words. You are warned, people.
Karim has been thinking about how crowds can contribute to innovation for 17 years, since he was at GE. There are two ways this happens:
1. Competitions and contests at which lots of people work on the same problem. Karim has asked who wins and why, motives, how they behave, etc.
2. Communities/Collaboration. E.g., open source software. Here the questions are: Motives? Costs and benefits? Self-selection and joining scripts? Partner selection?
More fundamentally, he wants to know why both of these approaches work so well.
He works with NASA, using topcoder.com: 600K users world wide [pdf]. He also works with Harvard Medical School [more] to see how collaboration works there where (as with Open Source) people choose their collaborators rather than having them chosen top-down.
Karim shows a video about a contest to solve an issue with the International Space Station, having to do with the bending of bars (longerons) in the solar collectors when they are in the shadows. NASA wanted a sophisticated algorithm. (See www.topcoder.com/iss) . It was a two week contest, $30K price. Two thousand signed up for it; 459 submitted solutions. The winners came from around the globe. Many of the solutions replicated or slightly exceeded what NASA had developed with its contractors, but this was done in just two weeks simply for the price of the contest prize.
Karim says he’ll begin by giving us the nutshell version of the paper he will discuss with us today. Innovation systems create incentives to exert innovative effort and encourage the disclosure of knowledge. The timing and the form of the disclosures differentiates systems. E.g., Open Science tends to publish when near done, while Open Source tends to be more iterative. The paper argues that intermediate disclosures (as in open source) dampen incentives and participation, yet lead to higher perrformance. There’s more exploration and experimentation when there’s disclosure only at the end.
Karim’s TL;DR: Disclosure isn’t always helpful for innovation, depending on the conditions.
There is a false debate between closed and open innovation. Rather, what differentiates regimes is when the disclosure occurs, and who has the right to use those disclosures. Intermediate disclosure [i.e., disclosure along the way] can involve a range of outputs. E.g., the Human Genome Project enshrined intermediate disclosure as part of an academic science project; you had to disclose discoveries within 24 hours.
Q: What constitutes disclosure? Would talking with another mathematician at a conference count as disclosure?
A: Yes. It would be intermediate disclosure. But there are many nuances.
Karim says that Allen, Meyer and Nuvolari have shown that historically, intermediate disclosure has been an important source of technological progress. E.g., the Wright brothers were able to invent the airplane because of a vibrant community. [I’m using the term “invent” loosely here.]
How do you encourage continued innovation while enabling early re-use of it? “Greater disclosure requirements will degrade incentives for upstream innovators to undertake risky investment.” (Green & Scotchmer; Bessen & Maskin.) We see compensating mechanisms under regimes of greater disclosure: E.g., priority and citations in academia; signing and authorship in Open Source. You may also attract people who have a sharing ethos; e.g., Linus Torvalds.
Research confirms that the more access your provide, the more reuse and sharing there will be. (Cf. Eric von Hippel.) Platforms encourage reuse of core components. (cf. Boudreau 2010; Rysman and Simcoe 2008) [I am not getting all of Karim’s citations. Not even close.]
Another approach looks at innovation as a problem-solving process. And that entails search. You need to search to find the best solutions in an uncertain space. Sometimes innovators use “novel combinations of existing knowledge” to find the best solutions. So let’s look at the paths by which innovators come up with ideas. There’s a line of research that assumes that the paths are the essential element to understand the innovation process.
Mathematical formulations of this show you want lots of people searching independently. The broader the better for innovation outcomes. But there is a tendency of the researchers to converge on the initially successful paths. These are affected by decisions about when to disclose.
So, Karim and Kevin Boudreau implemented a field experiment. They used TopCoder, offering $6K, to set up a Med School project involving computational biology. The project let them get fine-grained info about what was going on over the two weeks of the contest.
700 people signed up. They matched them on skills and randomized them into three different disclosure treatments. 1. Standard contest format, with a prize at the end of each week. (Submissions were automatically scored, and the first week prizes went to the highest at that time.) 2. Submitted code was instantly posted to a wiki where anyone could use it. 3. In the first week you work without disclosure, but in the second week submissions were posted to the wiki.
For those whose work is disclosed: You can find and see the most successful. You can get money if your code is reused. In the non-disclosure regime you cannot observe solutions and all communications are bared. In both cases, you can see market signals and who the top coders are.
Of the 733 signups from 69 different countries, 122 coders submitted 654 submissions, with 89 different approaches. 44% were professionals; 56% were students. The skewed very young. 98% men. They spent about 10 hours a week, which is typical of Open Source. (There’s evidence that women choose not to participate in contests like this.) The results beat the NIH’s approach to the problem which was developed at great cost over years. “This tells me that across our economy there are lots of low-performing” processes in many institutions. “This works.”
What motivated the participants? Extrinsic motives matter (cash, job market signals) and intrinsic motives do too (fun, etc.). But so do prosocial motives (community belonging, identity). Other research Karim has done shows that there’s no relation between skills and motives. “Remember that in contests most people are losing, so there have to be things other than money driving them.”
Results from the experiment: More disclosure meant lower participation. Also, more disclosure correlated with the hours worked going down. The incentives and efforts are lower when there’s intermediate disclosure. “This is contrary to my expectations,”Karim says.
Q: In the intermediate disclosure regime is there an incentive to hold your stuff back until the end when no one else can benefit from it?
A: One guy admitted to this, and said he felt bad about it. He won top prize in the second week, but was shamed in the forums.
In the intermediate disclosure regime, you get better performance (i.e., better submission score). In the mixed experiment, performance shot up in the second week once the work of others was available.
They analyzed the ten canonical approaches and had three Ph.D.s tag the submissions with those approaches. The solutions were combinations of those ten techniques.
With no intermediate disclosures, the search patterns are chaotic. With intermedia disclosures, there is more convergence and learning. Intermediate disclosure resulted in 30% fewer different approaches. The no-disclsoure folks were searching in the lower-performance end of the pool. There was more exploration and experimentation in their searches when there was no intermediate disclosure, and more convergence and collaboration when there is.
Increased reuse comes at the cost of incentives. The overall stock of knowledge created is low, although the quality is higher. More convergent behavior comes with intermediate disclosures, which relies on the stock of knowledge available. The fear is that with intermediate disclosure , people will get stuck on local optima — path dependnce is a real risk in intermediate disclosure.
There are comparative advantages of the two systems. Where there is a broad stock of knowledge, intermediate disclosure works best. Plus the diversity of participants may overcome local optima lock-in. Final disclosure [i.e., disclosure only at the end] is useful where there’s broad-based experimentation. “Firms have figured out how to play both sides.” E.g., Apple is closed but also a heavy participant in Open Source.
Q: Where did the best solutions come from?
A: From intermediate disclosure. The winner came from there, and then the next five were derivative.
Q: How about with the mixed?
A: The two weeks tracked the results of the final and intermediate disclosure regimes.
Q: [me] How confident are you that this applies outside of this lab?
A: I think it does, but even this platform is selecting on a very elite set of people who are used to competing. One criticism is that we’re using a platform that attracts competitors who are not used to sharing. But rank-order based platforms are endemic throughout society. SATs, law school tests: rank order is endemic in our society. In that sense we can argue that there’s a generalizability here. Even in Wikipedia and Open Source there is status-based ranking.
Q: Can we generalize this to systems where the outputs of innovation aren’t units of code, but, e.g., educational systems or municipal govts?
Q: We study coders because we can evaluate their work. But I think there are generalizations about how to organize a system for innovation, even if the outcome isn’t code. What inputs go into your search processes? How broad do you do?
Q: Does it matter that you have groups that are more or less skilled?
A: We used the Topcoder skill ratings as a control.
Q: The guy who held back results from the Intermediate regime would have won in real life without remorse.
A: Von Hippel’s research says that there are informal norms-based rules that prevent copying. E.g., chefs frown on copying recipes.
Q: How would you reform copyright/patent?
A: I don’t have a good answer. My law professor friends say the law has gone too far to protect incentives. There’s room to pull that back in order to encourage reuse. You can ask why the Genome Project’s Bermuda Rules (pro disclosure) weren’t widely adopted among academics. Academics’ incentives are not set up to encourage automatic posting and sharing.
Q: The Human Genome Project resulted in a splintering that set up a for-profit org that does not disclose. How do you prevent that?
A: You need the right contracts.
This was a very stimulating talk. I am a big fan of Karim and his work.
Afterwards Karim and I chatted briefly about whether the fact that 98% of Topcoder competitors are men raises issues about generalizing the results. Karim pointed to the general pervasiveness of rank-ordered systems like the one at TopCoder. That does suggest that the results are generalizable across many systems in our culture. Of course, there’s a risk that optimizing such systems might result in less innovation (using the same measures) than trying to open those systems up to people averse to them. That is, optimizing for TopCoder-style systems for innovation might create a local optima lock-in. For example, if the site were about preparing fish instead of code, and Japanese chefs somehow didn’t feel comfortable there because of its norms and values, how much could you conclude about optimizing conditions for fish innovation? Whereas, if you changed the conditions, you’d likely get sushi-based innovation that the system otherwise inadvertently optimized against.
[Note: 1. Karim’s point in our after-discussion was purely about the generalizability of the results, not about their desirability. 2. I’m trying to make a narrow point about the value of diversity of ideas for innovation processes, and not otherwise comparing women and Japanese chefs.]