1. Meta-Organizational Learning Through Digital Consensus
Junyi Li1
, Jungpil Hahn1
, Giri Kumar Tayi2
July 19, 2023
1Department of Information Systems and Analytics, National University of Singapore
2School of Business, State University of New York at Albany
4. Emergence of DAOs
• Decentralized Autonomous Organizations (DAOs): making organizational decision
by token-based voting.
• Meta-Oganization: legally autonomous entities [2]
2
5. Paradoxes in Organizational Design
1. Will DAOs (digital consensus) be an effective mode of decentralized governance?
2. Traditional paradoxes:
• Control vs. Flexibility [3]
• Centralization vs. Decentralization [1]
• Authority vs. Self-Organizing [7]
3
6. Research Question
How will digital consensus affect organizational design?
1. In what ways does consensus-based governance differ from a traditional one,
namely autonomy, and/or hierarchy (what is)?
2. Under what circumstances does consensus-based governance outperform other
modes (what might be)?
3. What is the underlying mechanism accounting for any performance differences
(how)?
4
11. Hierarchy
• Individuals learn from higher performing connected peers with probability η
• Top-Down Coordination: individuals’ solutions must be coherent with managers’
policy
8
12. DAO
• Individuals learn from higher performing connected peers with probability η
• Bottom-Up Coordination: individuals’ solutions must be coherent with policy
consensus
9
14. Experiment 1: Baseline
• Static Reality
• Best Hierarchies as in March’s
model [4]
• Naı̈ve DAOs
• simple majority rule
• uniform token distribution
• full engagement
10
15. Experiment 1: Baseline
Investigating performance variance adds insights.
• DAOs evolve through a staggered process of polarization and homogenization, as
opposed to autonomies’ continuous polarization and hierarchies’ continuous
homogenization. (What is & How?)
11
16. Experiment 2: Turbulence
The performance reversal delays as environments become less turbulent (more static).
“Turbulent” ← → “Static”
12
17. Summary of Baseline
Proposition 1a
In static environments, hierarchies outperform DAOs, and DAOs outperform autonomies.
Proposition 1b
In turbulent environments, DAOs tend to outperform hierarchies and autonomies.
13
18. Summary of Extensions
Proposition 2
Voting threshold has an inverted U-shaped impact on DAOs’ performance.
Proposition 3
Token asymmetry has a negative impact on DAOs’ performance.
Proposition 4a
Contributor incentive has a positive impact on DAOs’ performance when the inactive
rate is high.
Proposition 4b
Contributor incentive has a negative impact on DAOs’ performance when the inactive
rate is low.
14
20. Key Takeaways
? Power A vs. Power B
✓ Polarization vs. Homogenization
? Optimize DAOs’ design
✓ By voting threshold, token asymmetry, and contributor incentive
15
24. Digital Consensus
Digital consensus herein refers to algorithmic coordination rules that are automatically
executed in distributed computer systems and are only modifiable via voting.
• Transparent
• Democratically editable
• Self-executing
26. Genetic Algorithm
m payoff function in March’s (1991) model [4]:
Φ(x) =
m
X
j=1
δj (1)
where δj = 1 if xj corresponds with reality on dimension j; δj = 0 otherwise.
Belief diversity is measured as follows:
Belief Diversity =
2
mn(n − 1)
1
2
n(n−1)
X
i=1
m
X
j=1
ωij (2)
where ωij = 1 if two chosen individuals in the pair i have different beliefs on dimension
j; ωij = 0 otherwise.
27. Model Parameters
Parameter Remark Range of parameter Default
m Problem dimension 60, 90, 120, 150 90
n Number of individuals in the organization 280, 350, 420, 490 350
z The group size of the autonomous group 7, 14, 28 7
α The aggregation degree 1, 3, 5 3
η The learning rate of individuals 0.1 ∼ 0.9 0.3
p1 The probability of learning from the code (Hierarchy) 0.1 ∼ 0.9 0.1
p2 The probability of learning by the code (Hierarchy) 0.1 ∼ 0.9 0.9
p3 The probability of individual reconfiguration 0 ∼ 0.5 0
n′ The number of managers (Hierarchy) 40, 50, 60, 70 50
θ The voting threshold to pass a protocol as consensus 0.4 ∼ 0.7 0.5
Ttb The period of environmental turbulence 20 ∼ 100 -
Itb The intensity of environmental turbulence 0, 10%, 20% 0
ρ The shape parameter in Pareto distribution 1, 2, 3 -
γ Incentive degree & reactivation rate 0 ∼ 0.9 0
pi The probability of individuals being inactive in voting 0 ∼ 0.9 0
Repetition The number of simulation runs 500 500
Search Round The iteration of search in one simulation run 300, 500, 1000, 2000 -
Notes: Search rounds are verified by pilot tests. The default values correspond to the usage in the literature.
28. Baseline & Turbulence: Conclusion
Proposition 1a
In static environments, hierarchies outperform DAOs, and DAOs outperform
autonomies.
Proposition 1b
In turbulent environments, DAOs tend to outperform hierarchies and autonomies.
29. Experiment 3: Voting Threshold
• Voting threshold has an inverted U-shaped effect on DAOs’ performance.
30. Experiment 3: Voting Threshold
• Staggered process: polarization before consensus, homogenization after consensus.
• The staggered process can be adjusted by θ.
31. Experiment 4: Token Asymmetry
• Asymmetry restrains the polarization before consensus formation.
• Causing homogenization into minority views.
32. Experiment 5: Contributor Incentive
• Due to the intense reward, there will be additional asymmetry, resulting in worse
performance (–).
• Due to reactivation, incentives are beneficial when individuals rarely vote (+).
35. Sensitivity Analysis
• We conduct sensitivity analyses for non-focus parameters, including the group size
z, the number of individuals n, the number of managers n′, the aggregation level
α, the learning rate η, and the reality dimension m. Qualitative results are
consistent across these parameter values.
• A variety of hierarchies with varying supervision efficiencies are also examined. In
particular, the hierarchies’ performance will linearly decrease with increasing p1 or
decreasing p2, consistent with March’s (1991) model. This means naı̈ve DAOs
certainly outperform some deficient hierarchies.
• With individual reconfiguration rate p3 [8], the performance of DAOs and
hierarchies largely restore to static levels. Yet, autonomies may outperform when
p3 > 0.2, resonating with the literature that autonomous teams excel in intensive
organizational experimentation [5, 6].
36. Robustness: Supervision (i)
• p1: learning from code, p2: learning by code
• Left side: p2 = 0.9, right side: p1 = 0.1
• These two figures represent two specific segment lines across the entire 3D surface
(p1, p2, Z)
40. References i
[1] Christina Fang, Jeho Lee, and Melissa A. Schilling.
Balancing Exploration and Exploitation Through Structural Design: The
Isolation of Subgroups and Organizational Learning.
Organization Science, 21(3):625–642, June 2010.
[2] Ranjay Gulati, Phanish Puranam, and Michael Tushman.
Meta-Organization Design: Rethinking Design in Interorganizational and
Community Contexts.
Strategic Management Journal, 33(6):571–586, 2012.
41. References ii
[3] Daniel A. Levinthal and Maciej Workiewicz.
When Two Bosses Are Better Than One: Nearly Decomposable Systems
and Organizational Adaptation.
Organization Science, 29(2):207–224, 2018.
[4] James G. March.
Exploration and Exploitation in Organizational Learning.
Organization Science, 2(1):71–87, February 1991.
[5] Peerasit Patanakul, Jiyao Chen, and Gary S. Lynn.
Autonomous Teams and New Product Development.
Journal of Product Innovation Management, 29(5):734–750, 2012.
42. References iii
[6] Phanish Puranam, Harbir Singh, and Maurizio Zollo.
Organizing for Innovation: Managing the Coordination-Autonomy Dilemma
in Technology Acquisitions.
Academy of Management Journal, 49(2):263–280, April 2006.
[7] Marlo Raveendran, Phanish Puranam, and Massimo Warglien.
Division of Labor through Self-Selection.
Organization Science, 33(2):810–830, 2022.
43. References iv
[8] Timo Sturm, Jin Gerlacha, Luisa Pumplun, Neda Mesbah, Felix Peters, Christoph
Tauchert, Ning Nan, and Peter Buxmann.
Coordinating Human and Machine Learning for Effective Organization
Learning.
MIS Quarterly, 45(3):1581–1602, September 2021.