Week 3: Memos - Measurement and the Nature of Innovation

[jamesallenevans]

Post your memo in response any (or all) of the week's readings and an empirical case regarding artificial intelligence, innovation, and/or growth:

• “Surprising combinations of research contents and contexts are related to impact and emerge with scientific outsiders from distant disciplines.” 2023. Feng Shi & James Evans. Nature Communications.
• Note on the Transformer architecture underlying modern Large Language Models. 2025. James Evans and Austin Kozlowski.
• “Surprise! Measuring Novelty by Simulating Discovery.” 2020. Jacob Foster, Feng Shi and James Evans.
• “Combination and Structure” and “Mechanisms of Evolution”, Chapters 2 (27-43) and 3 (167-189) from The Nature of Technology: What it is and How it Evolves. 2009. Brian Arthur. Free Press: New York.
• “Assembly theory explains and quantifies selection and evolution”. 2023. Abhishek Sharma, Dániel Czégel, Michael Lachmann, Christopher P. Kempes, Sara I. Walker & Leroy Cronin. Nature.
• “Modularity, Higher-Order Recombination, and New Venture Success”. Likun Cao, Ziwen Chen, and James Evans. Preprint.

Post by Thursday @ midnight. By 1pm Friday, each student will up-vote (“thumbs up”) what they think are the five most interesting memos for that session. The memo should be 300–500 words (text) + 1 custom analytical element (e.g., equation, graphical figure, image, etc.) that supports or complements your argument. These memos should: 1) test out ideas and analyses you expect to become part of your final projects; and 2) involve a custom (non-hallucinated) theoretical and/or empirical demonstration that will result in the relevant analytical element. Because these memos relate to an empirical case students hope to further develop into a substantial final project and because they involve original analytical work, they will be very difficult to produce with generative AI and we strongly discourage you from attempting it. Some of the top-voted memos will form the backbone of discussion in our full class discussion and break-out room sessions.

[dishamohta124]

Innovation and Novelty: Bridging the Gap Between Diverse Sectors

Innovation is often perceived as the driving force behind progress in technology, science, and business. However, the way novelty is perceived and assessed varies widely across different sectors, depending on the context, existing knowledge structures, and the processes of invention. In this memo, I argue that novelty, while universally important, can be better understood and predicted by considering its role as a function of both the creative process and the underlying structural context of each field. One critical insight from recent studies is that innovation is rarely a singular, isolated event but rather a series of cumulative steps that often occur in highly interconnected and structured spaces. For example, in the technology sector, incremental innovations that align with existing frameworks may be viewed as less novel, while those that introduce new paradigms or significantly disrupt industry standards are considered breakthroughs. The concept of novelty as “expectation violation,” based on the Surprise theory, offers a solid foundation for understanding how innovations can defy prior expectations. A novel innovation is one that surprises its creators, users, and stakeholders because it challenges the prevailing assumptions and knowledge in its field. Let the novelty N of a new discovery be defined as the difference between expected outcomes E and the observed outcomes O, normalized by the rate of innovation R within a field.

Where:

• O represents the observed outcome or discovery,
• E is the expected outcome based on prior knowledge and models,
• R is the rate of innovation or knowledge accumulation in the specific field over time.
  This formula suggests that novelty increases when observed outcomes differ significantly from expectations, especially in fields where knowledge is accumulating rapidly. It also implies that fields with slower rates of innovation (e.g., traditional industries) might experience lower novelty in comparison to fast-evolving sectors (e.g., technology or biotechnology).

1. Technology (Software Development):
   In software development, innovations often result from unexpected solutions to complex problems. The pace of technological advancements has accelerated, making it more difficult to predict what new technologies will emerge. However, by applying the novelty formula, we can anticipate areas where new breakthroughs are likely to occur. For example, in fields like artificial intelligence or blockchain, rapid innovation and knowledge accumulation suggest that novel, unexpected technologies are more likely to emerge in the near future.

2. Healthcare (Biotechnology and Pharmaceuticals):
   In the healthcare industry, the discovery of new treatments and therapies is often constrained by existing medical knowledge. However, advances in genomics, personalized medicine, and CRISPR technology are disrupting this space. By applying the model of dynamic novelty, we can predict that areas where knowledge is rapidly evolving (such as gene editing) will produce significant breakthroughs in the coming years, due to the significant gap between current treatments and the possibilities these new technologies offer.

[kbarbarossa]

The study we read argues that innovation is often sparked by “unexpected” combinations of ideas, contexts, or disciplines. These "surprising" combinations tend to have a higher impact in advancing knowledge or technology. This insight inspired me to examine the extent to which these theories manifested in M&A deals in the U.S. within the AI industry.

To investigate, I focused on U.S.-based M&A deals in the AI industry with a value greater than $500M, ultimately identifying 44 transactions. While a larger dataset would provide more robust evidence, these deals still offer an intriguing lens through which to explore the relationship between cross-industry collaboration and deal value creation. Using my own research, I assigned each deal a ranking from 1 to 3 based on the level of cross-industry collaboration involved. The concept of "surprising" combinations from the study resonates well with the idea of cross-industry M&A, where firms oftentimes from different sectors collaborate to leverage their distinct expertise.

There is clear evidence that cross-industry deals often play a key role as part of a broader strategic approach. For instance, companies like Amazon acquiring Zoox (autonomous vehicle technology) or Microsoft acquiring Nuance Communications (healthcare IT solutions) illustrate how businesses leverage M&A to integrate expertise from different fields. These acquisitions often result in transformative products or services, echoing the study’s assertion that "surprising" combinations are catalysts for impactful innovation.

However, not all acquisitions exhibit a high degree of cross-industry interaction. This raises an important question: are such acquisitions equally impactful in terms of innovation, or does their value primarily stem from efficiency gains and market consolidation? The reading suggests that the latter—while important—may lack the same transformative potential as cross-industry combinations.

The accompanying scatterplot visualizes the relationship between cross-industry rankings and deal value from the 44 deals. Interestingly, while higher-ranking cross-industry deals often correspond to substantial financial investments, there are exceptions where lower-ranking deals yield comparable or even greater value. This pattern suggests that while interdisciplinary collaborations are common, their success also depends on the strategic alignment of the acquiring and acquired firms.

Ultimately, this analysis still highlights the relevance of the study’s findings to the corporate world, particularly in the context of AI’s rapid growth. Companies navigating this space face a dual imperative: to harness synergies within their domain while remaining open to transformative partnerships across industries.

[tHEMORAN02]

I think the measure of novelty within the article titled "Surprise! Measuring Novelty as Expectation Violation" was an interesting way of trying to quantify and measure a phenomena that no two people will ever define the same way. There are two components to this definition that are interesting to parse, 1. the use of patents as a instrument of quantity and 2. Mathematical approaches to bring structure to novelty as a concept
On component 1, patents may serve as the least bad way of understanding invention. Patents, at least in the united states where this paper was written, are a way of copyright protecting your processes, technologies and ideas. There is a fundamental commercial aspect to patents and the majority in the united states are developed in commercial contexts. For example, in the context of pharmaceuticals, patents are used to give companies a monopoly on their drugs. This often spawns waves of slightly modified copycat drugs that do more or less the exact same thing but only exist because of the laws surrounding this system. This exists in many domains and shows some of the problems of patents. On the other hand, there is no measurement that is quite as far reaching and "objective" in terms of data. Because of this reason, and the fact that the authors don't rely solely on this, I don't think the paper is worse for it
2. The math in this paper is basing itself on hard to quantify variables but i think that it is directionally correct. The novelty is function of two variables in this paper. Previous knowledge, represented by K and the nature of the inventive process, S. These variables are also multidimensional taking into account the complexity of the process. I think that most people would agree that if a process is super complex and advanced to create something it is novel and if something is totally new with not much foundation it is also likely novel. This to things also interact to multiply one another. I think that while these terms are so nebulous it can be difficult to measure, it definitely leaves like a step on the path the quantifying and understanding invention. Good paper overall. Attached below is theoretical 3d graph of how knowledge and search nature connect to novelty.

[malvarezdemalde]

Innovation can come from both field-specific research and cross-disciplinary collaboration. Field-specific research provides depth and domain expertise, building a strong foundation for incremental advancements. In contrast, interdisciplinary research introduces breadth, enabling novel combinations of knowledge that lead to transformative breakthroughs. The study "Surprising combinations of research contents and contexts are related to impact and emerge with scientific outsiders from distant disciplines" highlights the critical role of cross-disciplinary research in producing high-impact discoveries. This memo uses a Cobb-Douglas economic framework to evaluate the effects of incentivizing either form of research on total innovation in an academic environment that does not currently facilitate interdisciplinary exploration very much, making it much costlier to undertake than intradisciplinary research even though it has a higher marginal return in terms of innovation.

The mathematical model assumes that innovation ($I$) is a function of two inputs: field-specific research ($F$) and cross-disciplinary research ($C$). Innovation is defined as:

$I = A \cdot F^\alpha \cdot C^\beta$

where:

$I$: Total innovation output.
$A = 1$: Baseline productivity factor, capturing the general efficiency of research.
$F$: Quantity of field-specific research.
$C$: Quantity of cross-disciplinary research.
$\alpha = 0.4$: Elasticity of innovation with respect to field-specific research.
$\beta = 0.6$: Elasticity of innovation with respect to cross-disciplinary research, indicating its larger marginal impact on innovation.

The total budget is ($B$) and this is the budget constraint:

$p_F \cdot F + p_C \cdot C = B$

where:

$p_F$: Cost per unit of field-specific research.
$p_C$: Cost per unit of cross-disciplinary research.
$B = 100$: Total available budget.

Using these relationships, I calculate optimal allocations of $F$ and $C$ under three scenarios to evaluate the impact of cost reductions on each type of research.

Case 1: Baseline Scenario
When $p_F = 2$ and $p_C = 10$, the majority of the budget is allocated to field-specific research due to its lower cost. The optimal allocations are:

$F^* = \frac{\alpha B}{p_F} = \frac{0.4 \cdot 100}{2} = 20$
$C^* = \frac{\beta B}{p_C} = \frac{0.6 \cdot 100}{10} = 6$

Innovation output is:

$I^* = 1 \cdot 20^{0.4} \cdot 6^{0.6} \approx 9.71$

This scenario highlights how the high cost of cross-disciplinary research limits its contribution.

Case 2: Reducing Field-Specific Research Costs
Reducing $p_F$ by 50% to 1 while keeping $p_C = 10$ reallocates the budget, increasing field-specific research to:

$F^* = \frac{\alpha B}{p_F} = \frac{0.4 \cdot 100}{1} = 40$
$C^* = 6$

Innovation output rises to:

$I^* = 1 \cdot 40^{0.4} \cdot 6^{0.6} \approx 12.81$

While this intervention increases innovation, it is not the most effective way of doing so.

Case 3: Reducing Cross-Disciplinary Research Costs
Cutting $p_C$ by half to 5 while keeping $p_F = 2$ allows for a more balanced allocation:

$F^* = 20$
$C^* = \frac{\beta B}{p_C} = \frac{0.6 \cdot 100}{5} = 12$

Innovation output improves significantly:

$I^* = 1 \cdot 20^{0.4} \cdot 12^{0.6} \approx 14.72$

This scenario demonstrates the highest innovation output, highlighting the importance of reducing the disproportionately high cost of cross-disciplinary research since proportional reductions in cost to the two types of research resulted in a greater gain in innovation when applied to cross-disciplinary research.

The Shi and Evans article underscores the value of “surprising combinations” that arise from interdisciplinary research. They found that impactful breakthroughs often occur when researchers from distant disciplines collaborate, integrating diverse methodologies and knowledge domains. This aligns with the results of the model, which shows that reducing the cost of cross-disciplinary research enables greater resource allocation to these transformative combinations, amplifying their contribution to innovation at a greater rate than field-specific research.

Policy Recommendations

• Reduce the Cost of Interdisciplinary Research: Lowering $p_C$ through grants, subsidies, or collaborative infrastructure will unlock the potential of interdisciplinary research.
• Incentivize Cross-Disciplinary Work: Recognize and reward interdisciplinary research in funding, tenure evaluations, and awards to encourage researchers to engage in such efforts.

By targeting the cost of interdisciplinary research, policymakers can foster the so-called “surprising combinations,” reducing the barriers to cross-disciplinary work and turbocharging innovation.

[vmittal27]

Uniformity in Simulating Subjects with Different LLMs

In Simulating Subjects: The Promise and Peril of AI Stand-ins for Social Agents and Interactions, Professor Evans and Austin Kozlowski cover the issues with trying to simulate human subjects using large language models (LLMs). One such issue they raise is uniformity, in that a group of people you are trying to simulate will have diversity and high variance in their responses, which an LLM fails to replicate. For example, although abortion is popular in the United States, it's still around 60% who are pro-choice. Simulating these subjects with GPT-4 gets a pro-choice rate of over 95%.

I wanted to explore how LLM sizes may impact whether or not the uniformity increases or decreases, with the expectation that a larger model may have less uniformity in its responses since it "understands" more nuances in the population we're trying to mimic. To do so, I chose a policy proposal that occurred after the October 2023 cutoff for OpenAI's GPT-4 models: the current TikTok ban by the US government. According to Pew Research Center, 42% of Republicans support a ban, while only 24% of Democrats support a ban. In theory, then, the perfect simulation with an LLM would result in similar estimates.

I tested these two prompts on OpenAI's GPT-4o and GPT-4o-Mini. While OpenAI hasn't published the exact size of each of these parameters, estimates suggest GPT-4o is made of hundreds of billions of parameters, while GPT-4o-Mini is around 10 billion parameters. I decided to ask the same prompt 4 times: asking each model to role play as a Republican or Democrat and getting its thoughts on the TikTok ban. Each of these 4 scenarios was prompted 8 times. I chose to only do 8 prompts as the API calls are fairly expensive and I just wanted a quick sense of what the differences might be. I also set the temperature of the responses to be 1.8. This parameter controls the diversity and creativity of the response, and this temperature ensures the highest level of diversity.

For each message, the model was given a system message: "You are a {party_affiliation}. Only answer YES or NO to the following questions." The question asked was "The short-form video-hosting service TikTok has been under a de jure nationwide ban in the United States since January 19, 2025. Do you support the TikTok ban?" I got the context sentence from Wikipedia and chose to minimize political bias.

The results are summarized in the graphical figure below:

These results suggest that larger models may have slightly less uniformity, based at least on the samples from the LLM simulating a Democrat. However, it is still very far from the true estimates, and the only variability came in 2 responses. All this suggests that the uniformity issue with LLMs cannot be overcome by scaling LLMs to be larger or just increasing the temperature parameter. One potential answer, which I didn't explore but may explore more in future memos, is to add granularity to the system instruction. Instead of specifying that the LLM is a Democrat, we can specify it is a 25-year-old Democrat, or that it is a Democratic white male to get more diverse answers.

[druusun]

This week’s readings explore how modular recombination and assembly pathways influence innovation and growth. Modular recombination emphasizes the reuse of existing components to create novel outputs, while Assembly Theory highlights how historical constraints shape the complexity and diversity of innovations. This memo investigates how different innovation patterns—whether incremental or transformative—correlate with sectoral growth rates across industries.

Question:
How do innovation patterns (e.g., the degree of modular recombination vs. foundational innovation) correlate with growth rates across industries?

Hypothesis:
Industries with higher modular recombination (e.g., combining pre-existing technologies like IoT and healthcare devices) exhibit higher short-term growth due to scalability. In contrast, industries relying on foundational innovation (e.g., renewable energy) see slower initial returns but potentially greater long-term economic impact.

To analyze this relationship, I propose using:
Patent Citation Data: Patent co-citation networks can reveal modular recombination by showing how new patents integrate knowledge from multiple existing technologies. For instance, patents with high citation diversity likely reflect modular recombination.
Sectoral Growth Data: World Bank and OECD databases can provide GDP growth rates and value-added metrics across industries (e.g., healthcare, manufacturing, technology).
Analytical Element:
The chart below would show a comparison of patent co-citation diversity (proxy for modular recombination) against GDP growth rates across industries. For example, sectors like healthcare or ICT may exhibit high recombination intensity and rapid growth, while foundational sectors like energy may show a lagged but significant long-term growth impact.

By linking innovation patterns to economic growth, this memo provides a framework for understanding how modular recombination and assembly pathways shape industries differently. This dual lens offers critical insights into optimizing innovation-driven growth strategies across sectors.

[jacksonvanvooren]

Innovation and Word Embeddings: Apple vs. BlackBerry

In “Modularity, Higher-Order Recombination, and New Venture Success,” Cao, Chen, and Evans build a semantic space through word embedding to measure modularity and resulting successes of new companies. The article explores temporal changes and semantic shifts; for example, Amazon has become more closely associated with ecommerce than the rainforest.

I asked in my question this week if word embeddings could capture more pragmatic notions. For example, in the public mind, Apple tends to be associated with innovation and technological progress. Can and, if so, do word embeddings capture such sentiments about innovation?

Methodology

I found a random selection of 151,573 Wikipedia articles, thanks to the Wikipedia dumps. Of these pages, both Apple, Inc. and BlackBerry were present in the corpus, so I settled on comparing those two. Also, since BlackBerry has stopped making phones (though they are now an IoT company), it felt like a good example of a less innovative enterprise.

I filtered across all articles for “Apple” and “BlackBerry” (case insensitive). I then used the Python package gensim and the Word2Vec model to create word embedding vectors.
Page 17 of Cao, Chen, and Evan’s paper has a good overview on how this process works, but for the sake of brevity, I direct you to the documentation here.

I then define a list of “innovation” words. These include “innovation”, “creativity”, “design”, and “progress”. Though not exhaustive, I use this as a proxy for relating these companies with the general notion of innovation. I then calculated the similarity scores, which is the cosine similarity between the two vectors $v$ and $w$, given by $\frac{v\cdot w}{||v||||w||}$.

Results and Discussion

I first output the most similar words to both “apple” and “blackberry”. Both of these are technology companies and fruits, so they both have the same issues. “Apple” is heavily associated with “orchard”, but also “iigs”, “macintosh”, and “wozniak”. “Blackberry” is associated with the fig “banyan”, but also Android “oreo” and “cisco”. Future iterations would ideally control for this.

The following chart compares the similarity scores.

Across the board, there are stronger associations with “apple” and innovation words than with “blackberry”, though not by very large amounts, so statistical significance tests would be necessary. I also would try similar analyses on other company pairs (maybe Netflix/Blockbuster and Tesla/Mitsubishi). For Apple/BlackBerry, it is possible that semantically, both of these started out near the same point, but as Apple innovated and BlackBerry did not, Apple’s score increased. Temporal analysis, then, would be interesting in further examining this topic. I also noticed that the amount by which Apple is greater is similar for all four innovation words.

Ultimately, though this is not necessarily how word embeddings should be used, it seems it is possibly effective to capture differing levels of innovation through this method. For reference, my code can be accessed at the memo3.py file here.

[xdzhangg]

Relationship between Founders' Industry Background and Venture Success

In the first paper by Feng SHi and James Evans relating surprise in research content to impact, one key finding was that surprising advances in innovation typically emerge across different fields rather than within a single field, often when entrepreneurs / scientists from one discipline push ideas across to another.

The beneficial impact of cross-disciplinary collaboration on innovation prompted to think about how venture success may be correlated to the diversity of industry backgrounds that founders have. My hypothesis is that since ventures often have multiple founders, if they have professional experience across different industrys and/or if they studied different majors in college, then this may be positively correlated to higher venture success and exit price.

Methdology

To test the hypothesis, I gathered the following data:

• Identified 22 companies with more than 1 founder from a list of the most successful 45 venture investments in history (Google/UCWeb/Workday/RocketInternet/Zalando/Flipkart/Dropbox/Uber/Snowflake/WhatsApp/Facebook/Groupon/Snap/
  King Digital Entertainment/Mobileye/Twitter/Genentech/Adyen/GitHub/Spotify/Coinbase/Airbnb)
• Collected the number of disciplines / majors that the founders have collectively
• Gathered the IPO price or acquisition price, depending on if this venture ended up going public or not. (If both, I took the higher of the two)

Results:

1. Out of the 22 ventures with 1+ founders, more than half (13/22) have founders specializing in 2 or more fields, typically a blend of tech (computer science, engineering, physics, etc) and finance (MBA, economics, serial entrepreneur).
2. The companies with diverse founder expertise had a higher average exit price than those with only one industry background ($32.1Bn vs $20.1Bn)

This has several important implications for fostering entrepreneurial activity. First, cross-industry collaborations appear to be associated with greater commercial success in ventures. Second, diverse industry backgrounds among founders are linked to higher exit prices, suggesting that such diversity may enhance both the ideation process and the operational execution that follows. From a policy perspective, this highlights the need for governments and research institutions to actively promote cross-disciplinary dialogue by hosting events, courses, and programs that encourage collaboration between different fields.

[mskim127]

The shrinking value stock premium and why we might not need policy interventions:

Consider an inverted U shaped relationship between transdisciplinarity and impact. I believe that this relationship could be feasible since it is easy to see how practitioners that come from other disciplines can bring with them new perspectives and methods beneficial to knowledge creation (Consider Gary Becker and his application of economic tools to what were considered “social issues”). Beyond a certain point, however, synergies between disciplines or applicability of ideas might diminish. For example, I would not expect the methods of a poet to be of much use in physics (although of course I could be wrong).

Given this U shape, attempts to subsidize novelty through induced transdisciplinarity has the potential to negatively impact the value of the research being produced. To picture an extreme case, individuals might prioritize transdisciplinarity over research impact. It is unclear what a good incentive structure might look like to push transdisciplinary research towards optimal levels. I feel however, that it might be the case that the systematic undersupply of trans-disciplinary ventures that the paper identifies could be due to academics simply not realising the value behind them which this paper changes. Mclean and Pontiff in their 2016 paper identifies how academia destroys stock market predictability post publishing as investors move to close inefficiencies. Some attribute the recent shrinking of the value stock premium to this. I feel that academia and their past disdain for novel methods could resolve itself through papers like this, provided academics are principally motivated to do “good” work, and publishers and award-givers wish to stay relevant, which I feel they are. I would argue that the value of this paper lies not only in its influence on policy but also in its ability to inform academics/inventors and incentivize them to pursue novelty as a means to impact not as an end in itself.

I would also like to consider how the model might behave if we were to see this shift among academics. Novelty is measured by predicting what is the most likely combination of contents and contexts that are likely to occur in future research or patents and seeing by how much actual journal/patent entries subvert this expectation. Those that closely adhere to the prediction are deemed un-novel while those that deviate are deemed novel. The model’s predictions about what might be most feasible is dependent on the configurations it encounters in the training data. Assuming that there is a common thread to innovation (a big if I know) more encounters with novel types of ideas the model might demand even more outrageous and unthought of combinations to consider ideas similarly novel. The model is trained on a dataset that transverses a pretty large time span meaning that increases in novelties in more modern data might constitute a small enough portion of the data such that this inflation might not amount to much. However, I would assume that as the market for ideas approaches the optimal level of novelty, we might see the sensitivity of impact to novelty increase as ideas on this efficient frontier of novelty are perceived less and less novel by the model.

[siqi2001]

How does Technology Impact Education Beyond Providing Point Solutions?

In "Mechanisms of Evolution," Arthur outlines the six stages a novel technology undergoes. In the fifth stage, he discusses how this emerging technology can become available as a potential component for future technologies–further elements. Essentially, Arthur describes technologies as something recursive, reproductive, and alive. It is curious, then, to understand in what way technologies can participate in future or further innovation. In Tuesday’s class, Professor James Evans provided an account of General Purpose Technology that can potentially respond to this question. In class, Professor Evans pointed out three ways that GPT can facilitate further production of the current economy–by providing 1) point solution, 2) application solution, and 3) system solution.

In this Memo, I want to think about Technology’s generative power in the context of Education. What makes education distinct from all the other spheres of social life is the fact that Education is itself a recursive structure and a generative force. How should we, then, understand the encounter between Education and Technology? For the past two weeks, I’ve consistently centered my inquiry around Education. However, I only explored the effect of AI (Technology) as a point solution to Education and the importance of education on R&D personnel. This week, I hope to take a glimpse into how technology transformed the field of education. What is the impact of technology on education that is beyond point solution?

I again pulled data from OECD. The name of the file is “Number of enrolled students, graduates and new entrants by field of education.” I focused specifically on Bachelor’s and Master’s graduates in the United States, analyzing how student enrollment statistics change over time, with particular emphasis on the fields of education in which they graduate. The assumption here is that from 2015 to 2022, the education field will be more and more affected by technology, or the technology that inserts its impact on education will be more and more advanced. The visualization of this data, then, can give us some insight into how education’s encounter with technology has changed the inner structure of Education.

The graph shows that, over time, certain fields of study consistently remain more popular, such as Business, Administration, and Law; Social Sciences, Journalism, and Information. Fields such as Education and Engineering, although less of a hit, remain relatively stable in their popularity. What’s curious is 1) the steady increase in the enrollment of Health and Welfare and Information and Communication Technologies over time, and 2) the steady decrease in the enrollment of Arts and Humanities over time. The trends are clear in both Bachelor enrollment and Master enrollment. How might the increasingly intense interaction between Education and Technology serve as an explanatory element for the trends in students’ enrollment in different fields? Considering the fact that both education and technology are generative power, can we develop a hypothesis that fields such as Health and Information Technologies are growing because they can better collaborate with the modes of production advanced by current technologies, while fields like arts and humanities are shrinking because they are less compatible with the technological modes of production?

[yhaozhen]

In my project, I will investigate whether modular AI architectures, such as transformer-based models and established cloud platforms, help AI-oriented start-ups move more quickly from seed funding to follow-on rounds. Preliminary evidence suggests that these ventures can attract broad coalitions of investors, each with expertise in different modules—e.g., data engineering, software-as-a-service (SaaS), or consumer-facing applications. This dynamic mirrors the phenomenon described by Cao and colleagues, wherein higher-order invention not only lowers technological risk but also harnesses network effects that amplify growth.

An empirical angle I am exploring is the relationship between “local” and “global” novelty in AI ventures. “Local novelty” captures how distant each AI module’s subcomponents are from one another (e.g., software infrastructure and language processing). “Global novelty,” by contrast, indicates how far apart these AI modules lie from other sets of established technologies in the broader innovation space (e.g., combining generative AI with biotech workflows). The hypothesis, informed by higher-order recombination theory, is that small “local distance” paired with large “global distance” will yield faster scale-up and better funding outcomes—because the base modules are well-tested, yet their combination is unexpected and can open new markets.

The R demonstration below visualizes how “local” and “global” distances can interact to explain hypothetical “success” scores among AI start-ups. Although this is only a simplified simulation, it sets the stage for a more in-depth empirical analysis of real-world AI firms. By collecting actual data on AI startups’ technology stacks, pitch descriptions, and funding histories, I plan to test whether ventures focused on assembling stable modules truly outperform those pursuing radical, “ground-up” innovation.

[saniazeb8]

Contrasting Worlds: Novelty as a dynamic innovation or strategic choice

This week, we learned that novelty is both innovation and success. The market share from an innovation gains is crucial, yet it also diminish the uniqueness as everyone can adopt quickly. But how does it happen across different countries, does it always have to be something new and brilliant to evolve or can countries adopt an invention for their benefit? Will the developing countries excel artificial intelligence or will they adopt from excellent products made by the developed world?

Innovation is not a one-size-fits-all process it takes shape based on the context in which it unfolds. Developed economies treat innovation as a necessity, deeply embedded within their competitive and resource-rich ecosystems. These nations consistently push boundaries to maintain their technological leadership. Conversely, in the developing world, innovation is often a deliberate, strategic choice. Limited by resources, institutions, and infrastructure, these nations prioritize adopting and adapting innovations to address specific societal needs. The divergence in how innovation is approached and sustained between these two worlds offers profound insights into global development dynamics.

In developed economies, innovation emerges from the synergy of advanced ecosystems, high investment in research and development (R&D), and cultural drivers that celebrate experimentation. The institutional and financial support systems in these nations allow firms and individuals to explore bold, transformative ideas. For example, countries like Germany and Japan invest heavily in cutting-edge technologies such as renewable energy and robotics. This pursuit of innovation is often supported by robust academic industry linkages and a regulatory framework that minimizes barriers to experimentation. In contrast, developing economies view innovation as a pragmatic tool to solve pressing challenges. Governments often lead the charge, implementing policies that direct resources toward impactful areas. For instance, India’s adoption of digital platforms under the Digital India initiative brought transformative benefits to its rural population by prioritizing accessible and scalable technologies.

However, innovation in these regions is less about creating entirely new paradigms and more about adapting and diffusing existing technologies to fit local contexts. This approach often reflects a careful calculation of societal needs, resource limitations, and institutional constraints.

In developed economies, R and D are typically high, while C and L are relatively low. This enables a steady flow of both incremental and radical innovations. In developing economies, R is often constrained, and D is more selective, directed by governmental and societal priorities. The higher values of C and L create bottlenecks, leading to slower adoption and reduced experimentation.

By understanding these diverging pathways, we can design policies and partnerships that foster a more inclusive and dynamic global innovation and emerging artificial intelligence landscape.

[cskoshi]

Justifying My Second Major in English: Applying Assembly Theory to Poetry

In the article about Assembly Theory, we are introduced to the assembly equation, which attempts to quantify the “total amount of selection necessary to produce an ensemble of observed object”. While the article uses biology and chemistry as its field of study, as an Econ and English major, I try to justify completing my second degree to myself by seeing if we could apply AT to the study of poetry as well. In essence, can we use AT to quantify how “new” and “complex” a poem is. Of course, this would take many artistic liberties to define what a complex and new poem is, which would be more suited for a thesis than a memo on github. For simplicity’s sake, let’s take “complex” as just: how many big words and new phrases are used in the poem.

We start with the given assembly equation:

In this case, we simply replace “objects” with words. The assembly index a_i becomes the amount of unique letters in that word, n_i is the amount of times the word appears in that poem. N is the total number of unique words. e is still euler’s number. N_t is the total number of words in the poem. We use this formula but now call this variable “word complexity” or W.

What’s interesting is trying to quantify the amount of “new phrases” the poem has. For this, I created a variable, P, that we will use to “deflate” W. (i.e. it could have many complex long words but if they have all been said before then it’s not “new” according to our definition).

First, we set a variable P, which will represent the total similarity of our phrases in the poem to other famous poems. Next we manually extract phrases from the poem. Then. assuming we have access to the 1000 most famous poems ever written, we compare each phrase to each of the 1000 Poems. We use the TF-IDF method to get a number between 0 and 1 for each poem it’s compared to. Then we sum this up over all 1000 poems and then again over all phrases in our poem.

(To measure similarity, we use the TF-IDF method for measuring semantic similarity. (https://medium.com/@igniobydigitate/a-beginners-guide-to-measuring-sentence-similarity-f3c78b9da0bc)

As an example, if my current poem has the phrase "sweet like chocolate", and we compare it to another poem that has "tasty like candy" and "crunchy like peanuts" and say the former has a TF-IDF value of 0.68 and the latter has 0.37, because we take the maximum, the µ max value of that i-th phrase (sweet like chocolate) compared with the j-th poem (the one with the candy and peanut's phrase) will be 0.68. Then we sum it up over all other 999 poems to get an overall score for that phrase, and then we sum up for all phrases in my poem.

We will now deflate W by a factor of P, to get our final value, A, that shows the “complexity” of a poem

Of course, this is a practically messy and subjective measurement that needs much refinement before it is somewhat useful. For example, we would need to craft an algorithm to automate the process of comparing each phrase to the list of poems, and we still need to choose what kind of semantic similarity measure we can use. Even then, simply arithmetically summing up the µ is also very dubious, and our choice of deflating W so simply even more so. Accuracy aside, it’s still interesting to see how we tailor the measures of novelty to the subject matter, much like what was said in class about using subjective measurements for different fields.

[siyakalra830]

The Relationship Between Novelty and Probability of Innovation

This week, I wanted to explore the relationship between the perceived novelty of an invention and the probability of that invention occurring, based on prior knowledge and the inventive process. The central argument, as presented in Foster et al. 2020, is that novelty is not an objective quality but rather a context-dependent assessment based on how surprising an invention is, given what is already known and how inventions typically come about. This can be formalized through an equation that captures how novelty is a decreasing function of the probability of a discovery.

The core relationship can be expressed as:

Where:

• N is the perceived novelty of an invention.
• P(D|K,S) is the probability of the invention D given prior knowledge K and inventive process S.
• m is a constant, which is a scaling factor that can be adjusted depending on the context, representing the maximum novelty value when the probability of invention is at its minimum.

In this equation, as the probability P(D|K,S) increases, the value of N decreases, demonstrating the inverse relationship between the probability of an invention and its perceived novelty. When P(D|K,S) is very low (the invention is highly improbable), the novelty N approaches a maximum value determined by the constant m. Alternatively, when P(D|K,S) is high (the invention is highly probable), the novelty N approaches zero.

Foster et al. emphasizes that novelty is a function of surprise, which is directly related to the improbability of an event. The inverse function represents this relationship, where improbable events (low P(D|K,S)) are surprising, and therefore have high novelty. The constant m allows for the context-sensitivity of novelty assessment. In fields where highly novel inventions are common, m can be larger. In contexts where most inventions are incremental improvements, m can be smaller.

As a real-world example, consider the development of the first PC. At the time, the probability of creating such a device given the existing knowledge and technology was very low, meaning P(D|K,S) was small. As a result, the perceived novelty (N) was very high, since the inverse function m / P(D|K,S) would give a high output. Now, consider the release of a new model of an iPhone: given the existing knowledge and technology, the probability of producing this device P(D|K,S) is high, and the perceived novelty is low (N). The constant, m, can be thought of as a baseline to compare how novel two inventions are relative to each other, regardless of their context, and can be adjusted in relation to the inventive domain.

In conclusion, the relationship between novelty and probability is inverse: higher probability of invention results in lower novelty, and vice-versa. This relationship is mediated by prior knowledge and beliefs about the inventive process, which are subjective, vary by context, and can be measured.

[e-uwatse-12]

The readings emphasize that “unexpected” combinations of ideas, contexts, and disciplines frequently drive impactful innovation. Building on this, I set out to examine how these concepts play out in U.S.-based AI startups that have achieved substantial early-stage funding, particularly Series A rounds exceeding $15 million. This level of funding serves as a strong indicator of both investor confidence and the perceived potential for groundbreaking innovation.
To investigate, I focused on a dataset of AI ventures funded at the Series A level with funding above $15 million, identifying 62 startups over the past five years. Each venture was evaluated on two dimensions: contextual novelty (how unconventional or “surprising” the application domain is for AI) and modularity of recombination (the diversity of technologies combined in the venture’s solution). Startups were assigned rankings from 1 to 3 for each dimension, based on their descriptions and use cases.
The concept of “surprising” combinations aligns well with early-stage AI ventures, where startups often combine AI with unconventional fields or pair complementary technologies to tackle complex problems. For example, Infinitus Systems, which raised a $21.4 million Series A round, uses conversational AI and healthcare-specific knowledge graphs to automate administrative tasks in the healthcare industry—a domain not traditionally associated with cutting-edge AI. This venture scored highly on both contextual novelty and modular recombination, reflecting its ability to integrate diverse components in a way that addresses a niche yet high-impact problem.
Interestingly, ventures scoring higher on these dimensions often attracted larger Series A funding rounds. However, there were exceptions—startups with lower rankings that still secured substantial funding, suggesting that factors like team reputation or market size may also play significant roles. This raises an important question: while “surprising” combinations are associated with transformative innovation, to what extent does their perceived potential depend on strategic alignment with market needs or investor priorities?
The scatterplot accompanying this analysis visualizes the relationship between contextual novelty, modular recombination, and Series A funding size. Startups with high rankings in both dimensions generally cluster in the upper funding range, reinforcing the relevance of interdisciplinary and modular innovation strategies in capturing investor attention.
Ultimately, this analysis underscores the practical relevance of the study’s findings in the context of venture funding and AI’s rapid growth. Startups navigating this space face dual pressures: to pursue “surprising” recombinations that signal innovation potential and to align these efforts with investor expectations for growth and market relevance.

[Hansamemiya]

​​The paper on Modularity, Higher-Order Recombination, and New Venture Success examines how start-ups combine business “modules”— proven technologies and applications—to speed time to market and reduce the risk of failure. The distinction is between lower-order inventions, which integrate largely untested components, and higher-order inventions, which assemble already-functional modules into novel configurations. This modular assembly confers many advantages because the underlying technologies are already proven. Furthermore, such process helps founders attract more diverse investors, accelerating scaling and boosting IPO or acquisition odds.

Building on this idea, I propose a simplified economic model from the perspective of a venture capital firm looking to invest in or combine different modular companies. The model evaluates key factors such the following:

• Revenue Potential (R): The expected revenues or cost savings the acquirer anticipates once the target’s modules are successfully integrated.
• Synergy Factor (α): The additional value generated when the target’s modules are combined with the acquirer’s existing portfolio. A high α reflects strong complementarity, especially if the acquirer already possesses modules that amplify the target’s proven components (referred to as “higher-order” synergy).
• Probability of Successful Integration (p): The likelihood of smooth cultural, technical, and market integration. Integration is typically easier if the target’s modules are mature and well-defined (e.g., a proven payment engine). Conversely, it becomes riskier if the target is in the “lower-order invention” stage.
• Downside Risk Factor (β): The fraction of value retained if integration fails.
• Cost (C): The total acquisition cost, including the purchase price and integration expenses

Acquirer’s Value=p×[α⋅R]+(1−p)×[β⋅R]−C.

Start-ups with higher-order modules enable synergy (𝛼) to be realized more quickly and increase the likelihood of successful integration (𝑝). By contrast, lower-order components, which consist of untested or novel inventions, might offer significant upside if they complement the acquirer’s existing modules(R) but come with lower probabilities of success and salvageable value(β) if the technology fails.

Real-life example:Google’s Acquisition of YouTube
A notable real-world illustration of this model is Google’s 2006 acquisition of YouTube. At the time, YouTube already possessed stable, higher-order modules: a growing user base and a video-streaming technology that was both functional and scalable. Rather than attempting to develop a novel video-sharing platform from scratch, YouTube innovated its business model by making video sharing easy, free, and interactive, which significantly increased user adoption. This modular approach also enabled YouTube to attract a diverse array of investors beyond the niche of video streaming.

This example and model highlights how investors and corporations decide to acquire companies based on modularity and synergy. By focusing on stable, higher-order modules instead of entirely novel technologies, acquirers can significantly reduce time-to-market and the risk of failure. This strategic approach demonstrates the practical advantages of higher-order invention, as outlined in the modularity framework

[sijia23333]

Memo: The Dynamics of Scientific Impact Through Knowledge Expeditions

Theoretical Framework

Drawing from recent empirical studies of scientific discovery and technological innovation, I propose a dynamic impact equation that captures how breakthroughs emerge through the interaction of knowledge expeditions and the adjacent possible:

$$ I_{p,t} = N_c(t) \cdot e^{-\gamma D_s} \cdot \left(\frac{S_p}{A_p(t)}\right)^\alpha $$

Where:

• $I_{p,t}$ is the scientific impact of paper $p$ at time $t$
• $N_c(t)$ represents the novelty correlation function over time
• $D_s$ is semantic distance between source and target fields
• $S_p$ captures the surprise factor from unexpected combinations
• $A_p(t)$ represents the adjacent possible space at time $t$
• $\alpha$ is the elasticity of impact with respect to relative surprise

This equation attempts to identify the scientific impact by integrating three key mechanisms identified in recent empirical work. The correlation function $N_c(t)$ captures how one discovery triggers others over time, as demonstrated in analyses of Wikipedia edits, online music catalogues, and scientific publications. This function formalizes the empirical finding that novelties occur in correlated bursts rather than as isolated events. The semantic distance term $e^{-\gamma D_s}$ builds on evidence that successful knowledge expeditions typically bridge meaningful but manageable distances between fields, with the decay parameter $\gamma$ varying by discipline. Finally, the surprise-to-adjacent ratio $\frac{S_p}{A_p(t)}$ formalizes Kauffman's concept of how the space of possible discoveries expands with each breakthrough, with the time-varying denominator reflecting the dynamic nature of this space.

Empirical Application

The empirical evidence strongly supports this theoretical framework. In biomedical research, papers representing knowledge expeditions across semantic domains achieve 3.5 times higher citation rates than typical papers, according to Shi & Evans' analysis of millions of MEDLINE articles. The equation explains this through the interplay between semantic distance ($D_s$) and relative surprise ($\frac{S_p}{A_p(t)}$). When scientists from one field publish results in journals from a distant field, they often generate high surprise values relative to the adjacent possible space of the target field.

The framework also accounts for observed differences between scientific domains. In physics, where Tria et al. found more hierarchical knowledge structures, the distance decay parameter $\gamma$ is larger (approximately 0.88 from APS journal analysis), reflecting stronger disciplinary boundaries. Biomedical sciences show more distributed semantic networks with lower $\gamma$ values (around 0.77 from MEDLINE data), enabling more frequent successful long-distance knowledge expeditions.

This theoretical framework also identifies the pattern of breakthrough discoveries. When examining Nobel Prize-winning papers, we find they typically display high $N_c(t)$ values combined with moderate $D_s$ values and high relative surprise ratios. For instance, the 1985 paper by Grynkiewicz et al. on fluorescent calcium indicators, which received over 16,000 citations, represents an optimal combination of these factors. The work bridged chemistry and cell biology (moderate $D_s$) while introducing surprisingly novel methods ($\frac{S_p}{A_p(t)} \approx 0.95$) that triggered numerous subsequent discoveries captured by high $N_c(t)$.

The equation thus provides not only a quantitative model for understanding past scientific breakthroughs but also suggests strategies for fostering future innovations. It implies that the most impactful discoveries emerge from well-calculated knowledge expeditions that balance semantic distance with cognitive accessibility, while generating sufficient surprise relative to the current adjacent possible space.

[aveeshagandhi]

Introduction

Often dependent on modularity and recombination, entrepreneurial innovation is boosted by the 2 latter elements that allow ventures to experiment without taking risks. The role of modularity in organizational emergence is examined, underscoring the importance of higher-order recombination-where ventures assemble successful components as a determinant of success. Using findings from Professor Evans’ Study, “Modularity, Higher-Order Recombination, and New Venture Success”, I want to analyze how strategic modular assembly influences venture outcomes, offering actionable insights for entrepreneurs.

Key Insights

Modularity and Innovation

The adaptability and robustness of complex systems, including entrepreneurial ventures, depend on modularity. Using Herbert Simon's "Architecture of Complexity", modular systems are demonstrated to be less prone to failure since individual modules can fail without causing the system to collapse. It is typical for successful ventures to avoid creating disparate components simultaneously. By assembling proven modules, they accelerate market readiness and reduce risks associated with exploratory failures.

Higher-Order vs. Lower-Order Recombination

As I see it, there are 2 broad categories of inventions:
- Lower-order inventions: combining unproven components into novel systems that can generate public goods in the long run despite high failure risks.
- Higher-order invention: Combining existing, successful modules to create value-generating systems. Research has shown that this strategy increases the probability of attracting funding, achieving an IPO, or acquiring a high-value company.
Recombination at higher orders facilitates rapid prototyping and market entry by leveraging preexisting successes. Sharma et al.'s (2023) work on assembly theory shows how recombining proven modules enhances system complexity.
Cao, Chen, & Evans, 2023, examine 45 years of U.S. venture-funded startups and find the following:
The likelihood of IPOs and high-priced acquisitions is significantly higher for ventures engaging in higher-order recombination.
There is a correlation between lower-order recombination and higher closure rates or middle-to-low-value acquisitions.
Higher-order inventions enhance the viability of the functioning of a venture by adding fuel to funding milestones as well as diversifying the pool of capital they receive and further investing in their business and the economy.

Implications

To further push toward compatibility and synergy, entrepreneurs should assemble pre-existing modules
To maximize value generation and attract investments from a diversified set of limited partners (hedge funds, mutual funds, pensions, etc.), limited resources should be allocated to higher-order recombination
By leveraging modularity, ventures can iterate prototyping more quickly and outpace their competitors in market adaptation, which highlights a rudimentary struggle in early-stage, high-growth shops

Analytical Element

The graph represents the comparative outcomes for ventures employing lower and higher-order strategies. Based on data from 1976 to 2020 as given in Cao, Chen, and Evans (2023).
Figure: Outcomes of Lower-Order vs. Higher-Order Recombination

(1) Green: Higher-order recombination displaying a higher probability of IPOs and high-value acquisitions
(2) Red: Lower-order recombination correlating with increased closure rates

Essentially, modularity and higher-order recombination provide a solid roadmap for entrepreneurial success: in assembling proven modules, early to later-stage ventures can achieve more rapid development, curtailed risk of failure, and diversified investment, which accelerates robust organizational emergence. Entrepreneurs can integrate such (above) insights into strategic planning to maximize their moat over competitors, helping them tackle some of the most prevalent issues in VC: strategic resource allocation, churn, and time-to-market adaptability.

[anishganeshram]

Memo: The Impact of Innovation, Novelty, and Risk on Valuation Multiples

The paper "Modularity, Higher-Order Recombination, and New Venture Success" provides a framework for understanding how startups create value through recombination. The core argument is that higher-order recombination—the process of combining existing, well-established modules from different domains—creates greater commercial viability and higher success rates. This idea extends to startup acquisitions, where companies that introduce novel yet structurally familiar innovations are more likely to be acquired at higher multiples.

Our equation:
M = (I⋅N)/R

captures this principle by suggesting that a startup's acquisition multiple (M) is directly proportional to its innovation level (I) and novelty (N), while being inversely affected by risk (R). In the context of the paper, higher-order recombination increases both I and N, enhancing the startup’s valuation, while excessive novelty without modular grounding can increase risk, leading to lower acquisition values.

One of the key insights from the paper is that novelty alone is not sufficient for success. Many early-stage companies attempt radical innovations that lack integration with existing market structures, increasing adoption friction, technological uncertainty, and commercialization risks. This aligns with the negative impact of excessive risk in our equation, where a high R diminishes the acquisition multiple despite high novelty. The paper highlights that successful startups strategically recombine existing knowledge, balancing novelty with familiarity, which effectively reduces RR while preserving I and N.

The paper also emphasizes the role of modularity in innovation success. Startups that develop technologies with reusable, scalable components tend to achieve higher valuations because they fit into existing business ecosystems. This is particularly relevant in software and platform-based industries, where companies build on standardized frameworks and APIs, allowing for rapid scaling and integration. The concept of modularity in innovation is a strong driver of high acquisition multiples, as seen in the software sector’s sustained high valuations in M&A markets.

Another factor discussed in the paper is funding speed and investor diversity. Startups that secure funding quickly and attract a diverse investor base signal market confidence, which acts as a risk-reduction mechanism (R). In other words, investor interest serves as an external validation of a startup’s innovation strategy, improving its acquisition prospects. The equation supports this by showing that reducing risk (via strong investor backing) enhances M even when innovation and novelty remain constant.

Software commands the highest multiples due to high innovation (I), modular novelty (N), and lower risk (R)from scalable business models. Hardware shows volatility, reflecting higher innovation but greater risk, leading to fluctuating valuations. IT Services remain stable, with moderate novelty and lower risk, resulting in consistent multiples. This confirms that startups maximizing modular innovation while mitigating risk achieve higher acquisition values, reinforcing the role of higher-order recombination in driving successful exits.

[chrislowzhengxi]

Modularity and Higher-Order Recombination in Taiwan’s Semiconductor Industry

Taiwan's success in the semiconductor industry is a clear example of how modularity and higher-order recombination can drive innovation and growth (as talked about in this week's reading). Modularity, the division of labor into specialized components, has allowed Taiwanese firms to excel in specific segments of the semiconductor chain. Higher-order recombination (integrating these components in innovative ways) has helped Taiwan's large and small firms meet the demands of the industry.

Modularity: Foundation for Success in TSMC

A great example of modularity is Taiwan Semiconductor Manufacturing Company (TSMC), which has become the world’s largest foundry by focusing solely on wafer fabrication. This specialization allows TSMC to invest heavily in cutting-edge technologies, such as 3nm and 5nm process nodes. Specialization leads to efficiency. Its focus on manufacturing lets global design firms like AMD and NVIDIA concentrate on IC design while outsourcing production to TSMC. This collaboration, illustrated in the figure below, shows how modularity allows companies to specialize and innovate in a very specific factor.

Smaller firms like MediaTek, which specializes in IC design for consumer electronics, also benefits from TSMC's modularity. Modularity enables MediaTek to focus on its core strengths in designing chips for smartphones and IoT devices, while it can leverage the TSMC foundry’s expertise for production.

Higher-Order Recombination Drives Innovation

What sets Taiwan apart is its ability to combine specialized modules with global partners. An example is TSMC’s collaboration with Apple to manufacture the A-series chips used in iPhones. Apple supplies its proprietary designs, and TSMC uses its efficient manufacturing processes to produce these high-performance chips. This has led to significant revenue growths for both companies.

The back-end manufacturing segment also plays a key role in Taiwan’s success. Companies like ASE Technology excel in packaging and testing, critical steps that protect chips and ensure their reliability. ASE combines advanced materials with TSMC-fabricated wafers, and this transitions effectively from manufacturing to end-user-ready chips. As seen in the figure below, packaging and testing accounted for 13.4% of Taiwan's semiconductor output in 2023. Every process in the supply chain is essential - disruptions in any single aspect may lead to significant delays in semiconductor production.

The industry’s value distribution also shows how different modules—design, manufacturing, and back-end processes—contribute to Taiwan’s dominance. Foundry services alone accounted for 57.4% of the sector’s output in 2023, and without specialization, this dominance would have been impossible.

Taiwan's use of modularity and higher-order recombination also ensures long-term resilience. As shown below, the semiconductor industry weathered global downturns and maintain strong revenue growth. During the 2021 global chip shortage, Taiwan’s firms rapidly scaled production by using their modular processes and collaborating within while most of the world is in pandemic lockdown. This ensures supply remains optimal.

Taiwan’s semiconductor success is rooted in modularity and higher-order recombination. Focusing on specific strengths and integrating expertise from other global firms, companies like TSMC and ASE have seen huge growth and innovation. Aiming for modularity has made Taiwan a global semiconductor power, and more growth is yet to come.

Source: Pwc Taiwan

[spicyrainbow]

The Evolution of Technological Development in Image Capturing

In Combination and Structure and Mechanisms of Evolution, Arthur presents technology as evolving within a recursive system, where innovations build upon existing technologies, creating sequences and branches of advancements. This recursive structure follows a tree-like model, with fundamental principles forming the trunk, major technological breakthroughs as main branches, and subsequent refinements extending as sub-branches. Using this framework, we can analyze how image-capturing technology evolved into film and digital cameras today within this recursive structure.

At the core of image-capturing technology is the fundamental principle of light projection and recording. The earliest known technology, the camera obscura, established foundational knowledge in optics which helped us understand how light could be projected and manipulated. This led to the development of daguerreotype photography in the 19th century, which refined the concept to allow permanent image recording using light-sensitive materials, fulfilling the opportunity niche of personal and commercial portraiture capturing.

The invention of film photography plays an important role in the evolution of image-capturing technology. The shift from plate-based methods to flexible roll film created an opportunity niche by making photography more accessible and convenient for the average consumer. This development not only enhanced image storage potential but also introduced modular components like interchangeable lenses, leading to the diversification of camera types.
As Arthur suggests, technology is created and used when it fulfills an opportunity niche and makes economic sense. As photography technology iterates in this recursive structure, its ability to customize image capturing, storing, and the widespread use created a further desire of cost and time efficient photography technology. This led to the development of digital imaging.

Digital cameras replaced film with electronic sensors, enabling instant image capture without chemical processing. This advancement opened new opportunities, such as digital storage solutions with SD cards, digital image processing software, and customizable settings like ISO and shutter speed.

To visualize this evolution, I created a simplified tree structure illustrating the transition from film to digital cameras. At the core of the tree is the camera itself. The camera's primary function—capturing light to produce visual representations—remains constant, but the underlying mechanisms have transformed as we move from film to digital. The branches are divided into four main assemblies: Processing Assembly, Lens System, Power & Control, and Image Capture. These assemblies further break down into subassemblies supporting the main systems.

Following the pathways highlighted in purple, we see how electronic image processors replace chemical film processing, which in turn leads to the opportunity niche and development such as digital storage replacing physical film rolls. As digital cameras evolved, their increased ability to adjust image settings digitally created a demand for a more sophisticated user interface—an entirely new branch that emerged only because of the advancements digital technology introduced. As visualized in the diagram, components not highlighted in purple represent technologies that have become inactive in digital cameras. This demonstrates the sequence technology goes through in evolving, through opportunity niches and replacement of sequences of components.

[darshank-uc]

Sentiment Differences in AI- and Human-Produced Text

In Simulating Subjects, Evans and Kozlowksi discuss how LLMs face the problem of uniformity: a collection of responses from a LLM in an attempt to model a human sample of responses offer less variance than a true human sample. When asked the question, “Should the government do more to help the poor?” GPT-4 overwhelmingly responds with the moderate answer on a scale of 1-5, whereas the human GSS sample was far more diversified. When looking toward Human-AI systems––where longer responses are key to simulating interaction instead of a binary yes/no or scale-based answer––I was curious if this uniformity extended to the style of LLM responses.

Sentiment analysis is a way of assessing the degree and direction of emotion in text. Google’s Natural Language API is trained on millions of contextual sentence fragments and can provide two sentiment metrics for text: a sentiment score from -1 to 1 and a magnitude score > 0. Sentiment scores closer to 1 indicate more positive sentiment in text (e.g. compliment), whereas scores closer to -1 indicate negative sentiment (e.g. criticism); a higher magnitude score indicates greater conveyed emotion, which is often dependent on text length. For example, a heated argument between two people on an online forum would likely have scores close to -1 and high magnitudes.

In order to properly simulate interaction, we want diversity in sentiment scores and magnitudes for LLM-produced text, akin to human-produced text. The stigma around base GPT models is that they sound “robotic” or “awkward” without proper prompting, so we might expect sentiment scores closer to 0 (neutral) with low magnitudes (non-expressive). I tried to compare the sentiment metrics for a random sample of human-produced responses to an open-ended, bias-inducing question with those from GPT-4 responses. I took a random sample of 20 comments on Reddit to the question, “What is your unpopular opinion?”, collected sentiment metrics on each of these responses using Google’s Natural Language API, and recorded the length of the comment to normalize the magnitude. I then initialized 20 separate calls (no saved cache) to GPT-4 with the same question and recorded the same metrics.

The results show that the Reddit (human) response sentiment scores on average are more extreme than ChatGPT sentiment scores and show higher variance. The data also suggests that the question lends itself to more negative responses (which is expected, since contrarian opinions often criticize the dominant opinion), as most responses score below 0. In other words, ChatGPT was generally more neutral in tone than humans. However, humans voluntarily answered the Reddit question, whereas ChatGPT was prompted––there is selection bias for humans to answer if they strongly feel about something, either positive or negative. Human responses also had a higher mean magnitude score (0.044 compared to 0.038), which implies that humans were also more expressive in their responses than ChatGPT. However, it’s interesting to note that ChatGPT provided more responses with magnitudes above 0.04 than Reddit based on a single prompt. Because the question intended to generate personal bias, ChatGPT’s predictive capacity likely depended on previous expressive text.

With a proper web-scraping tool and a subscription to Google’s API, we could obtain a much more accurate sentiment distribution of human-produced text beyond a sample size of 20. However, I think this case more largely highlights the obstacles in creating out-of-the-box LLMs that are prepared to simulate humans. I expect that with significant prompting and characterization of an LLM, these distributions could look far more similar.

[willowzhu]

AI as a technology
In Combination and Structure and Mechanisms of Evolution by Brian Arthur, we are introduced to the principles and definitions of technology. Arthur argues that technologies are built from hierarchies of technologies. For example, fighter jets contain numerous modules of technologies within it; or, humans discovered fire to cook, then to make pottery, then to weld weapons, developing clusters of practices around technologies.

In Chapter 9, Mechanisms of Evolution, Arthur establishes the six discrete steps of novel technologies in the buildout game (steps can be found on page 178). I wanted to compare his theoretical structure to the historical evolution of AI, and question if the development of AI either defends or refutes his model. Some questions to ask are: Which technologies in the active collection did AI displace? Did AI set up further needs or opportunity niches for supporting technologies and organizational elements? My hypothesis is yes.

I went to IBM’s website and found some more information supporting AI as a novel technology. I discovered that the idea of "a machine that thinks" dates as far back as ancient Greece.

Furthermore, we can see the hierarchies of technologies within the modules of the AI itself:
Training
Tuning
Generation, evaluation, and more tuning

Next, I also wanted to explore AI as a living organism, as Arthur argues that technology is a living organism (page 189). How would Arthur respond to AI being a replication of the human brain? Can we tie in AS as GPT from our class discussions, the model of feed-forward neural networks, to support Arthur’s argument? After all, AI is modeled after the human brain itself, and it learns from itself.

Lastly, to pose some philosophical questions: What comes next after AI? Thinking about science in terms of Kuhn’s paradigms, what problems are AI solving? In some ways, does it feel like we have reached an ultimate truth with AI? How would Kuhn refute this using his logic and theories? What is the next paradigm?

[jesseli0]

Studying Surprise and Interdisciplinary Work in the Social Sciences

“Surprising combinations of research contents and contexts are related to impact and emerge with scientific outsiders from distant disciplines.” discusses novelty as related to interdisciplinary studies in the physical sciences. This suggests analogous study that could be performed in the social sciences that may have interesting implications. Due to the overlapping subjects of study in the social sciences, we would like to encourage interdisciplinary work. This would both further understanding of the concepts they seek to understand (e.g. power, democracy, wealth, happiness), and prevent wasted efforts in pursuing research in avenues that are already well documented within other disciplines.

A paper titled “The Superiority of Economists” collects data on citations in the social sciences referencing works in fields other than their own. In the attached chart, economics stands out as being the most cross referenced work out of economics, political sciences, and sociology. A sanguine reading suggests that economics has great efficacy as a social science, causing the other fields to look up to their research. However, an alternative hypothesis, especially knowing the value of interdisciplinary work, suggests that economics as a field could use more insight from its fellow social sciences. In fact, while a lot of the techniques of economic research have diffused into the greater social sciences, the reverse has not been seen as much.

There may be limitations with the current methodology in performing this study in the social sciences. Firstly, there is a consideration we have to make about keywords. While terminology in the physical sciences is more well defined, there isn’t such rigidity in the social sciences. While this flexibility around terminology is necessary to make the papers themselves coherent, it causes difficulty in using them as keywords to assess how predictable one paper is. For example, the word “institution” has a different definition from field to field, and may even be defined differently from paper to paper within a field. This might make it difficult to construct models around the word “institution”, since its presence can imply many different constructs. This would confound our ability to measure surprise off of words.

Studying interdisciplinary work in the social sciences would let us understand if surprise plays a factor in the significance of research in the social sciences. We would expect that social scientists aware of the state of research in fields other than their own would have a better perspective of the grand scheme of things and be able to produce better research. Interdisciplinary work would also combine approaches found in multiple fields, which can yield novel techniques for interesting results.

[michelleschukin]

Transformers and CNNs: Modularity and the Paradigm Shift in AI

This week’s readings highlighted the role of modularity and selection in driving innovation. Cao, Chen, and Evans demonstrated how modular systems like higher-order recombination foster entrepreneurial success, while Assembly Theory provided a framework for understanding the evolution of complexity through selection dynamics. As I discovered these ideas, I aimed to explore how these ideas can be directly applied to the evolution of artificial intelligence (AI), where modular architectures such as transformers have rapidly reshaped the field, displacing earlier architectures like Convolutional Neural Networks (CNNs).
I decided to investigate these two spheres of thought through a comparative lens. In exploring this week's papers, I discovered a strong conceptual connection between transformers and Assembly Theory (AT). I noted how Cao, Chen, and Evans emphasized the importance of modularity in innovation, specifically higher-order recombination as a driver of success. By applying these ideas to the AI domain, I identified transformers as an ideal case study. AT emphasizes how the assembly index quantifies the historical steps required to construct a system, where systems with higher assembly indices reflect greater complexity and modularity. Similarly, transformers exemplify a modular architecture that integrates smaller, pre-trained components into scalable systems. This connection became apparent as I examined how transformers, through modularity, enable efficient exploration of “assembly spaces” in AI development. Unlike CNNs, which often require task-specific designs, transformers recombine general-purpose modules. I think this reflects the dynamics of directed selection described in AT, where selective pressures favor architectures that balance complexity with adaptability.

To analyze this shift, I collected data from arXiv on the number of papers published mentioning “transformers” and “convolutional neural networks” from 2017 to 2024. The bar chart below illustrates a striking trend: research on transformers has grown exponentially, while interest in CNNs has stagnated or declined after peaking in 2020.

The chart highlights:

• An exponential rise in transformer-related publications, with a notable surge after 2020.
• A stagnation and decline in CNN-related publications post-2020, reflecting the waning dominance of this architecture.
• The crossover point in 2020, where transformers surpassed CNNs in publication volume, marking a paradigm shift in AI research

In 2020, transformer-related publications surpassed convolutional neural network (CNN)-related publications for the first time, with 1,555 transformer papers compared to 807 CNN papers, marking a pivotal shift in research focus. This data supports the argument that modular architectures like transformers exemplify higher-order recombination, enabling researchers to build upon pre-trained models across tasks. This modularity aligns with the selective pressures of modern AI—favoring scalability, efficiency, and adaptability. Conversely, the decline of CNNs underscores the limitations of earlier, less modular approaches in adapting to the demands of a rapidly evolving field.

The findings resonate with this week’s readings by illustrating how modularity and selection drive innovation. Transformers thrive by integrating proven components into new applications, demonstrating the new age emphasis on higher-order recombination and Assembly Theory’s focus on selection favoring complexity and novelty. Understanding this recent paradigm shift not only sheds light on the trajectory of AI research but also provides a framework for fostering innovation in other domains.

[jessiezhang39]

Assessing Supply Chain Innovation through Social Network Analysis

In this week’s reading, we discussed the immense potential for trans-disciplinary scientific expeditions to achieve unexpected innovation with outsized impact. I aim to explore the broader implications of this study beyond the scientific community. Specifically, what organizational and social networks is the most conducive to innovation? Does the degree of linkages and density facilitate or hinder innovative endeavors?

An important mechanism in organizational network is structural holes. First proposed by American sociologist Ronald Burt, structural hole refers to a gap in social networks where some individuals have indirect or little ties with certain others, thus creating a metaphorical hole in the network fabric. Burt argues that those who can better navigate and bridge these disconnects - the so-called “brokers” are uniquely positioned to informational and control benefits. Consider Robert as a broker in the below social network:

Individuals and organizations who can identify and capitalize the structural holes in alliances or partnerships can combine diverse knowledge streams, resources, and perspectives, thus achieving a higher likelihood of innovation. Inspired by the framework of Social Network Analysis, I attempt to propose a model that integrates structural holes and network closeness to measure innovative performance within an organization.

I propose the following hypotheses:

H1: Higher structural holes positively influence innovation by fostering exchange of diverse experiences and viewpoints.

H2: Higher network closeness positively impacts innovation by enhancing knowledge sharing and collaboration.

H3: The interaction between structural holes and network closely positively moderates innovation performance, balancing diverse information (from structural holes) with efficient knowledge flow (from closeness).

Key Variables:

Structural Holes Index: measured using Burt’s Network Constraint Index.

Network Closeness: measured as the average inverse distance between an actor and all others in the network.

Innovative Performance: the dependent variable measured using metrics such as the number of patent applications, return on annual R&D investments, or revenue CAGR from new products/services.

Model:

Application:

A real-life application of the above model is in supply chain management. Using tools such as Gephi or UCINET, we can identify key suppliers, manufacturers, and distributors and map their interconnections. Further, we can assess the degrees of density and fragmentation in the supply chain network. Innovation performance could be measured as efficiency in fabric recycling or reduced water usage, for instance. With the model above, we can test whether bridging structural holes in the supply chain (e.g., connecting a small-scale textile innovator with a global distributor) leads to more exploratory innovations. We can also evaluate if regions with tightly-knit supply chains (high network closeness) yield more efficient and cost-effective exploitative innovations.

[JaslinAg]

How Expanding Open Access Will Drive Novelty

“Surprise! Measuring Novelty by Simulating Discovery,” describes the following relationship between novelty (N) and innovation (I) where A is prior knowledge or art, and S is the creative / search process
$N \sim \frac{1}{ P(I \vert A, S) }$

As I was brainstorming ideas for this memo, I was met with many paywalls. This led to me looking into how paywalls limit innovation by restricting the availability of research papers –ex. Nature, Science, etc. - and news articles - ex. The New York Times, The Economist, etc.

I observed the following:
(1) S: A more efficient creative or search process would increase the probability of innovation. For example, the invention of the google search engine increased the accessibility of knowledge, increasing the probability of innovation. Relevant Study
(2) A: Increasing prior knowledge increases the probability of innovation, but as more knowledge is created, it is harder to find a novel idea. Thus, there are diminishing returns for A.

With this, I created a simple model for the probability of innovation,
$P (I \vert A, S) = \alpha S + \beta A^{\gamma} $
where $\gamma \exists [0, 1]$ and $\alpha, \beta$ are scale terms.

Secondly, I modeled the relationship between $A$ and $o$. I defined $o$ as the percent of works that are open access. At any step in time, the number of previous works, $A$ is constant. As $o$ increases, the number of works available to researchers increases. The plot below shows this relationship.

The number of works available to research is the maximum number of works consulted by researchers. Using the graph above, I found the following relationship.
$\max W_c= oA$
As a research team likely will not consult all the works available, I added a scale term $\eta$. I solved this equation for A:
$W_c = \eta o A$
$A = \frac{W_c}{\eta o}$

Plugging this into the function for the probability of innovation:
$P (I \vert A, S) = \alpha S + \beta A^{\gamma} $
$P (I \vert A, S) = \alpha S + \beta {(\frac{W_c}{\eta o})}^{\gamma} $

Thus, we get the following relationships:
(A) $P(I \vert A, S) \sim \frac{1}{o^{\gamma}}$
(B) $N \sim \frac{1}{ P(I \vert A, S) } \sim o^{\gamma}$

The intuition behind this is that “open-access research outputs receive more diverse citations.” As “Surprising combinations of research contents and contexts are related to impact and emerge with scientific outsiders from distant disciplines” showed, diverse careers, teams, and methods are positively correlated with hit probability. The $\gamma$ term likely represents the diminishing returns for career and team novelty.

The implication of this is that we should increase open access. More access to knowledge will lead to more novelty, which would boost our economy. Currently, only about 38% of all scholarly articles are open-access and this is harming our economic potential.

[yangkev03]

In this week's reading, we learned about how knowledge expeditions, and more generally multidisciplinary intellectual collaboration, led to benefits in creating surprising breakthroughs. In these cases, it is found that "successful teams have not only members from multiple disciplines but also members with diverse backgrounds who stitch interdisciplinary teams together". In this way, we can see that a relationship exists between some form of multidisciplinary focus and achieving scientific and technological breakthroughs.
 
In today's memo, I would like to extend this analysis to corporations to see how businesses can adopt a multi-disciplinary approach and whether or not this may lead to value creation. Firstly, I will state the main ways in which a business acts through multidisciplinary measures. One way that is adopted by most major corporations is an executive suite with different functions within business and various skillsets. This corresponds with the first layer of the paper, having teams from multiple disciplines. In the next step, teams can achieve greater value through each individual having interdisciplinary experiences, and even interdisciplinary roles. Finally, corporations can also create a product that aligns with multiple industries, thereby creating a product that is multidisciplinary.
 
From this graphic, I would like to propose a method to view the gains from each level.

Level 1: Teams with multidisciplinary functions

Level 2: Teams with multidisciplinary individuals within each function

From this simple graphic, we can see that Level 1 teams have no overlap in terms of skillset between each individual. As a result, there may be gaps within their abilities that cannot be solved by any of them. In level 2, these gaps are reduced through a widening of each individual's skillset so that more issues can be solved. However, at the same time, the overlap between skills can lead to greater amounts of inefficiencies. Given that the corporation pays for each individual's skillset, the greater skillset will also mean great costs accrued to the corporation.
 
If we're looking at corporations as a whole, we can see that companies in multiple industries tend to have greater costs attributed to sustaining the business. In the 1980s and 1990s, conglomerates who had businesses in multiple industries broke off through pressure from financial markets with players such as private equity firms being able to create value through running specific business units. If we model out the benefits of conglomeration, we can say that they occur through increasing the addressable market and pricing power of a company. In terms of costs, being able to compete in multiple industries will increase their costs across the board. Thus, we can model the relationship in this way.

$\alpha p + \beta q - \gamma C(q)$
 
Where $\alpha > 0, \beta > 0, \gamma >0$

[anacpguedes]

Pricing Epigenetic Treatments Using AT models

The field of epigenetics offers promising approaches for medical treatments by targeting the systems that control how genes are turned on or off, without changing the underlying DNA sequence. These treatments focus on modifying the "switches" that regulate gene activity, such as adding or removing chemical tags on DNA or the proteins around which DNA is wrapped. For example, in B-cell lymphomas, abnormal activity of genes involved in cell growth can be regulated by targeting histone deacetylation, a process that compacts DNA and silences genes. Drugs like vorinostat and romidepsin, both histone deacetylase inhibitors, have been approved to treat certain lymphomas by restoring proper gene regulation and preventing uncontrolled cell proliferation. Similarly, DNA methylation inhibitors like azacitidine and decitabine are used in blood cancers such as acute myeloid leukemia (AML) to reverse harmful gene silencing and reawaken tumor-suppressor genes. The flexibility of epigenetic therapies—being reversible, adaptable, and capable of fine-tuning gene expression—aligns perfectly with the growing trend of precision medicine, which emphasizes tailoring treatments to individual patients. As understanding of epigenetic mechanisms advances, these treatments are likely to become increasingly common worldwide, applied across a wide variety of diseases, including cancers, neurological disorders, and autoimmune conditions.

Given the increased importance of this field, we can take advantage of AT models to design a structured strategy for pricing gene editing treatments based on the requirements for treatment, complexity, and incorporation of additional costs related to drug delivery. A possible pricing model can consist of:

The calculation of the price of an epigenetic treatment is done by factoring in multiple components that account for the complexity, effectiveness, and scalability of the therapy. The modular assembly indices represent the complexity of individual steps required to create the therapy, with adjustments for the clinical impact factor, which reflects the treatment's effectiveness and value to patients. Each step has a cost, and a risk factor accounts for additional costs due to unsuccessful development iterations. The model also adds an emergent value , which captures the treatment’s uniqueness, such as addressing unmet medical needs. These total costs are divided by the copy number, representing the number of patients treated, and the scalability multiplier, which adjusts for economies of scale. This structure ensures the price reflects both the complexity of development and the practical realities of production and distribution.

To note, the model doesn’t take into account the multifaceted nature of healthcare pricing and its interconnectedness with regulation, market competition, and equity. Incorporating these elements can help align the model with the broader realities of healthcare drug delivery.

[cmcoen1]

Biotech/pharma is characterized by lengthy clinical timelines and high attrition, so Cao's emphasis on modular recombination is particularly salient here. Two qualitative considerations stand out. Validated “modules” like CRISPR-Cas9 for gene editing or lipid nanoparticles for drug delivery permit quicker iteration in drug discovery than entirely de novo approaches. Because these modules typically come with established safety or efficacy data, startups can attract more diverse capital by proving they can integrate existing technology stacks rather than taking on all discovery risks alone.

Presented here is a model that captures this dynamic and employs a multi-stage hazard framework. Let $$\lambda_i(t)$$ denote the hazard rate of a biotech venture $i$ reaching a successful exit (think high-value acquisition or pivotal Phase III milestone) at time (t). Specify:

$$ \lambda_i(t) = \lambda_0(t) \exp \Big( \gamma_1 \mathrm{ModDist}_i + \gamma_2 \mathrm{ModSize}_i - \gamma_3 \mathrm{OrigInnovation}_i \Big) $$

where $$\lambda_0(t)$$ is a baseline hazard function (like a piecewise exponential), $$\mathrm{ModDist}_i$$ measures the “distance” between modular technologies combined (capturing how novel the higher-order recombination is), $$\mathrm{ModSize}_i$$ proxies for the maturity or scale of each module such as a Phase II–ready asset vs. a preclinical tool, and $$\mathrm{OrigInnovation}_i$$ assesses lower-order novelty like a new biologic product untested in humans. A positive $$\gamma_1$$ suggests that combining more disparate yet well-characterized modules increases the venture’s hazard of success and a negative (\gamma_3) indicates the risk penalty for unproven science.

Financing dynamics are also a critical variable in biotech, so to incorporate them it's necessary to extend the model to a system of simultaneous hazards:

$$\lambda_{i,\text{Success}}(t) = \lambda_{0,\text{Success}}(t) \exp(\cdots)$$ and $$\lambda_{i,\text{Refinancing}}(t) = \lambda_{0,\text{Refinancing}}(t) \exp \Big( \delta_1 ,\mathrm{ModDist}_i + \delta_2 ,\mathrm{ModSize}_i - \delta_3 ,\mathrm{OrigInnovation}_i \Big)$$

where $$\lambda_{i,\text{Refinancing}}(t)$$ captures the likelihood of securing subsequent venture rounds or strategic partnerships. Empirical implementation might use cause-specific competing-risks regression to disentangle exit outcomes (positive vs. negative) and validate that biotech firms mixing stable modules realize higher hazard rates for value-creating events. Even a 0.5-point increase in $$\mathrm{ModDist}_i$$ on a normalized 0–1 scale could boost the probability of successful exit by 25–30% over a 5 year horizon, showing the outsized return for combining established, complementary components. This model highlights how modular assembly in biotech can reduce risk in the innovation pipeline, enhance finance prospects, and cumulatively advance more therapies through clinical development. To apply in practice, researchers would rely on cause-specific competing-risks regression to help tease apart positive versus negative exit outcomes and see if biotech ventures that combine proven modules really do better in terms of hitting high-value milestones.

On the measurement front, capturing $$\mathrm{ModDist}_i$$ and $$\mathrm{ModSize}_i$$ could be gleaned from data sources like:
-Patent Databases: to show how technologies overlap or how novel a given combination is
-Clinical Trial Registries: to reveal each module’s stage of development
-Partnership Announcements: which highlight the nature and maturity of a collaboration
The choice to combine certain modules might reflect deeper traits about a firm like its overall quality or its connections to investors, so it’s not purely random. Strategies like instrumental variables, fixed effects, or matched samples could be used which would help tackle the bias that comes from companies self-selecting into the kinds of modules they adopt. This would give a clearer sense of whether modular recombination itself drives better outcomes, rather than just reflecting which firms were stronger to begin with.

[ypan02]

Analysis of Factors Affecting Startup Success Using Random Forest Model

I found a dataset on Kaggle containing information about 190 startups, including their funding round numbers, total funding amount, geographic location, time to first funding, average number of VC fundraising participants, industry, and their status (either "closed" or "acquired").

After cleaning the dataset, including filling in missing values for numerical columns with the median, I selected 8 features that I believe might influence a startup’s status. These features include: funding_total_usd, funding_rounds, avg_participants, is_software, is_ecommerce, age_first_funding_year, milestones, and relationships. I then used a Random Forest machine learning model to predict startup status. The dataset was split into 80% training data and 20% testing data.

The classification report (attached below) for the model indicates an overall accuracy of 77.3%, which is a solid result. The model performs well in predicting acquired startups, with a recall rate of 91%. However, the recall for closed startups is lower, suggesting that the model struggles more with predicting this class.

In terms of feature importance, the variable funding_total_usd has the highest weight (0.2366), which aligns with our expectations: investor funding is crucial for a startup's survival and success. The second most important feature is relationships (0.2272). While the exact nature of this "relationship" score is unclear, it likely reflects factors such as investor networks, customer relationships, and possibly political connections, all of which are known to impact business success.
Age_first_funding_year ranks third in importance. This feature captures how early a startup acquires its first round of funding. Younger startups tend to be more innovative, which could explain their higher likelihood of being acquired, in line with our in-class discussions about how younger companies can often out-innovate older ones.

Interestingly, funding_rounds has a relatively low impact on predicting startup success. This could be because well-funded startups with strong cash flows may not need to rely on multiple funding rounds to grow, allowing them to skip rounds and still thrive. Lastly, industry or sector has the lowest feature importance, which may seem surprising. Despite the common belief that fast-growing industries like software and e-commerce drive acquisition, this dataset suggests that industry type doesn't strongly predict whether a startup will be acquired successfully.

This model was trained on a relatively small dataset of startups, and there is room for improvement, particularly in feature engineering and model selection. Nonetheless, the insights gained offer interesting perspectives into factors influencing startup success, particularly regarding funding, relationships, and age at first funding.

[rzshea21]

From this week's readings, it's clear that modeling business complexity in comparison to various performance metrics (returns, survival, market capture, patent portfolios, etc.) in a way that meaningfully and adequately accounts for different organizational structures, technological developments, and industry environments, is a difficult problem. Readings from this week and external research wrestle with this problem, and explore unique ways to evaluate firm complexity in order to model the relationship between complexity and performance. "Modularity, Higher-Order Recombination, and New Venture Success" by Cao, Chen, and Evans discusses how successful firms incorporate modular recombination, ideally high-order recombinations, to experiment with innovative ideas and business models. Another study, Hur and Kim (2023), also emphasizes the role of organizational innovation and modular recombination to create competitive edge and uses standardized metrics on these characteristics within firms in comparison to financial performance metrics. This type of innovation in firm structure can leverage new technology to boost profitability, which is important in fields where a technological edge isn't enough to sustain profitability. Therefore, new ventures that leverage modularity for organization growth and design will accelerate innovative capabilities and get access to capital quicker. If we are to believe that modularity benefits early ventures, we can assume this is related to the paper, "Assembly theory explains and quantifies selection and evolution," where modular design allows firms to adapt quickly and efficiently to changes in market environments and technologies. My main issue is with how we quantify firms' technological complexity and the effect of modularity in various industries. Modularity will have higher impacts on early-stage, more versatile business models, and particularly within technologically complex industries. However, many practitioners still use patent portfolios as a proxy for technological complexity, which we seen can be used non-productively by larger incumbent firms to stifle innovation. My proposed solution represent technological complexity of firms is to include a term to represent modularity within firms, and penalize that term with increasing firm size where the effects are less pronounced. With this modified function, sheer patent portfolio size presents a more meaningful innovative advantage to smaller firms with modularity and higher-order recombination, which acts as a sort of multiplier for smaller, but highly innovative, patent portfolios.

[nmkhan100]

Context and Argument
The study by Shi and Evans (2023) reveals that impactful scientific innovation arises from integrating ideas across distant disciplines, particularly through the contributions of “outsiders.” These outsiders, often removed from established norms within a specific field, can forge novel connections between otherwise siloed areas of knowledge. The core insight is that impactful discoveries are not merely about technical mastery but about identifying and integrating ideas with unexpected relevance.

One contemporary application of this insight is seen in the development of quantum computing algorithms for chemistry. This field brings together quantum physicists, computer scientists, and chemists—distinct groups traditionally working in separate silos. For example, the Variational Quantum Eigensolver (VQE), a key algorithm for solving complex molecular problems, emerged from the combined efforts of quantum computing researchers and chemists. The success of VQE underscores Shi and Evans’ argument: high-impact research is born at the intersections of fields, often driven by interdisciplinary collaborations.

Empirical Case: Quantum Computing for Climate Solutions
Consider the case of quantum computing applied to material discovery for carbon capture. Scientists from quantum information theory and materials science have begun leveraging quantum algorithms to simulate new materials for capturing and storing CO2. The challenge lies in accurately modeling molecular interactions, which are computationally prohibitive for classical methods. By integrating quantum algorithms with chemical theory, researchers have accelerated progress in developing materials like metal-organic frameworks (MOFs) for more efficient carbon capture.

This case highlights the importance of disciplinary distance in driving innovation: quantum computing researchers bring fresh perspectives and tools, while materials scientists contribute domain-specific expertise. The resulting breakthroughs align with Shi and Evans' finding that impactful research arises from outsiders navigating between distant contexts.

Analytical Demonstration: Novelty as an Intersection of Disciplinary Distance and Cognitive Integration
To formalize this idea, I propose a function that models novelty 𝑁 as a product of disciplinary distance 𝛿 and cognitive integration 𝐼:

𝑁 = 𝛾 ⋅ 𝛿 ⋅ 𝐼

Where:

𝛾 is a scaling factor reflecting the broader research environment (e.g., funding, institutional support).
𝛿 quantifies the distance between the disciplines involved (e.g., how unrelated their methods and paradigms are).
𝐼 represents the degree of cognitive integration, measuring how effectively researchers from different fields collaborate and share insights.

This equation captures how novelty emerges when distant disciplines are bridged effectively. In the quantum-carbon case, 𝛿 is high due to the distinct methodologies of quantum physics and chemistry, while 𝐼 depends on the extent of collaboration and knowledge exchange among researchers.

Shi and Evans’ findings are deeply relevant to today’s most pressing challenges. In areas like climate change, fostering interdisciplinary collaborations can unlock transformative solutions.

[diegoscanlon]

A very surface-level review of modularity and imitation between China and the US

Modularity, Higher-Order Recombination, and New Venture Success seems to claim that combinations of higher / lower modules, particularly those that have large distances between them, will lead to greater probabilities of successful innovation -- since the data they collected seems to be US based, we can interpret the result in the context of the US. However, can we draw similar conclusions in other countries?

In the past, China was a place where copycats of US startups would emerge, some of which would become successful. It would seem intellectually dishonest to claim that all comparable Chinese companies copied US counterparts; even if we looked at a metric like year founded, a company can go through many pivots. It also might be the case that humans in different countries encounter similar problems, and thus local founders independently discovered and resolved those issues. However, it also seems intellectually dishonest to not recognize that copying is possible. This point might be strengthened when founders have a history of copying US startups in Chinese markets, and even admit their source of inspiration publicly. Perhaps the most notable example is Wang Xing, who tried imitating Facebook and Twitter before finally becoming successful with a Groupon clone (which has since evolved). Examples of similar companies (not saying they're copycats) include Didi and Lark.

Thus, one might conclude that this imitation is not the same as the modularity presented in Modularity, Higher-Order Recombination, and New Venture Success; that copycat venture success is formed by other factors like cultural adaptation and local go to market strategies. I think I'd agree with this, just look at businesses like Careem, or even Meituan (Groupon copy). But I can also see an argument that cloning a business is the same as the modularity that the paper speaks of. Modularity takes existing modules (which I thought were broadly defined in the paper) and reorganizes them. Why can a startup idea that has found success in the US, and cultural adaptation not be two different modules? The distance the paper argues is necessary for innovation exists between the idea of the solution and the problem in a local market. That is, novelty is subjective, and repurposing existing modules is similar to copying a startup in a new country. To be honest though, I don't really like this argument because I don't like the modularity argument that much -- aren't all things based on things that have been done in the past? What even is a module? Why do we have to abstract this so much?

However, this class has made it obvious that competition (in healthy amounts) can lead to greater innovation, so copycats are probably good in that sense (maybe not in the sense of increasing US-companies' profits), but only if these US companies trying to take on new countries have the motivation and ability to iterate well (unlike Uber in China).

Survey of Chinese Espionage in the United States Since 2000
source

[florenceukeni]

AlphaFold — The Intersection of AI and Drug Discovery

The development of DeepMind's AlphaFold represents a pivotal moment in scientific innovation and reflects a theme supported by our readings this week, including Shi and Evans (2023) about breakthrough discoveries resulting from cross-field collaboration. By integrating machine learning, structural biology, and biochemistry, AlphaFold addressed the 50-year-old challenge of protein structure prediction, showing how collective expertise can lead to transformative solutions.

Protein folding has long been a computational grand challenge, because more traditional methods like X-ray crystallography and electron microscopy need extensive time and resources, making them often unsustainable or unattainable. AlphaFold approached this by using advanced neural network architectures and deep learning techniques, and in doing so they bridged the gap between computational approaches and biological domain expertise. The project's success depended on three fundamentally collaborative strategies:

1. Integrating deep learning innovations from natural language processing
2. Incorporating structural biology domain expertise
3. Utilizing publicly available protein structure databases from the Protein Data Bank (PDB)

These strategies allowed AlphaFold to display almost perfect accuracy in protein structure prediction when it was demonstrated at the 2020 Critical Assessment of Protein Structure Prediction (CASP) competition, marking a scientific success driven by collaboration.

Two examples of protein targets in the free modeling category. AlphaFold predicts highly accurate structures measured against experimental result.

Looking past technical achievement, AlphaFold's has many implications for scientific innovation. The project speeds up the process of drug discovery by enabling faster and more cost-effective protein structure analysis. Its open-science approach—releasing code and predictions for nearly all known proteins—fosters global research collaboration and inspires similar interdisciplinary approaches in scientific research.

The AlphaFold case is a great example of how cross-context collaboration can lead to unexpected and opportunities, supporting the previously mentioned concept of "collective abductions." By moving past established disciplinary boundaries, researchers can develop solutions that individual fields can’t reach on their own.

Looking forward, I think that AlphaFold is a compelling model for interdisciplinary innovation, and suggests a lot of potential for similar collaborative breakthroughs across scientific domains. I would be interested in further examining interdisciplinary breakthroughs in unexpected collaborations to shed light on how collective abductions can drive transformative AI-driven solutions.

[yasminlee]

Impact of Market Conditions of Startup Success

Understanding the drivers of startup success is a foundational question in entrepreneurship research. The readings from this week on modular recombination​ and assembly theory​ suggest that success hinges not only on innovation but also on the strategic combination of existing components. In this memo, I want to explore the question of how market timing can amplify venture outcomes, particularly in the context of startups leveraging modular recombination.

The concept of modularity highlights how startups combine pre-existing, proven components to mitigate risks and accelerate development. For example, a startup might integrate established technologies (such as cloud computing and mobile applications) to create a novel product or service. This modular strategy reduces uncertainty by drawing on previously validated innovations. However, modular recombination alone does not guarantee success; I think its impact is also significantly amplified or constrained by prevailing market conditions. Favorable market conditions (such as increasing demand, technological maturity, or supportive regulatory environments) can create windows of opportunity where modular recombination yields outsized returns. Conversely, unfavorable conditions may stifle even the most innovative startups, as markets fail to adopt their solutions. Assembly theory provides a useful lens here, emphasizing how historical pathways and selection pressures influence the viability of new ventures.

To produce a theoretical analytical element that captures the interaction between modular recombination and market timing, I propose the following outline. To create the graph using existing data, I would extract startup descriptions from datasets like Crunchbase or VentureXpert and compute a modularity index using NLP techniques, such as clustering terms in company descriptions to identify the diversity of modules combined. A market favorability score would be derived from external metrics like GDP growth rates, venture capital cash-on-market levels, and industry funding trends during the startup’s launch. I chose these specific metrics because I found a research article titled “Startups, Growth, and the Macroeconomic Environment: Valuation and Business Cycles”, which found that these were the key macroeconomic factors that influence startup success. Success outcomes (like IPO likelihood or acquisition value) could then be mapped onto a heatmap or contour plot to visualize how modularity and market favorability interact to predict success.

Understanding the interplay between modular recombination and market timing is crucial for guiding entrepreneurs to optimize strategies and launch timing. It also helps investors and policymakers foster environments that support innovation. This analysis could reveal how these factors combine to drive startup success, offering actionable insights for research and practice.

[nsun25]

In “Modularity, Higher-Order Recombination, and New Venture Success”, Likun Cao, Ziwen Chen, and James Evans explore the concept of modularity and higher-order recombination in the emergence of successful new ventures and complex systems. Successful start-ups do not simply combine random components, but instead strategically assemble novel combinations of successful modular components. This approach reduces the risk of entrepreneurial failure and enables rapid experimentation.
We learn that modularity is critical for the emergence of complex social, technological, and organizational systems; higher-order combinations of successful modules are more likely to lead to venture success; ventures that combine diverse, successful modules can more effectively accelerate firm development, diversify investment, and attract private funding.
To explore the long-term performance of ventures after IPOs and acquisitions, I propose to design an empirical analysis that expands the definition of venture success beyond the initial IPO or acquisition event. This analysis would provide insights into whether high-priced acquisitions and successful IPOs truly lead to sustained success. Here's a proposed model for this analysis:

Data Collection: Gather data on ventures that have gone through IPOs or acquisitions, including:
IPO price or acquisition value
Financial metrics (revenue, profit, market cap)
Employee count
Market share
Industry classification

Time Frame: Track these metrics for 5-10 years (or more depending on availability of data) post-IPO or acquisition.

Grouping: Categorize ventures into:
Higher-priced acquisitions
Lower-priced acquisitions
Successful IPOs (based on initial valuation)
Less successful IPOs

Performance Metrics: We'll use a composite score of multiple metrics to measure long-term success:
〖Performance Score〗_it=w_1 〖Revenue〗_it+w_2 〖Market Share〗_it+w_3 〖Employee Growth〗_it+w_4 〖Proftiability〗_it
Where the w terms are weights assigned to each metric.

Model: We can use a panel data regression model to analyze the long-term performance of these ventures. The model would look like this:
Y_it=β_0+β_1 X_it+β_2 Z_i+ϵ_it
Where:
Y_it is the performance metric for venture i at time t
X_it are time-varying covariates (e.g., market conditions)
Z_i are time-invariant covariates (e.g., industry, initial valuation category)
ϵ_it is the error term

Comparative Analysis: to compare the performance of different groups over time
Y_it=β_0+β_1 〖Group〗_i+β_2 〖Time〗_t+β_3 〖Group〗_i*〖Time〗_t+ϵ_it
Where:
〖Group〗_i indicates the venture's category (ex. higher-priced acquisition, successful IPO)
〖Time〗_t is a dummy variable for pre- and post-periods
〖Group〗_i*〖Time〗_t is the interaction term of interest

This comprehensive model allows us to track long-term performance using multiple metrics, compare different groups of ventures based on their initial success, and account for industry and market factors. By implementing this analysis, we can gain insights into whether the initial success of a venture (as measured by IPO price or acquisition value) is a reliable predictor of long-term performance and sustainability.

[rbeau12]

Virtual reality (VR) goggles enable users to interact with a fully-customizable, visual form of reality. The invention allows for a gaming experience that is deeply immersive and makes the body itself a controller. Several of the readings present the notion that novelty is an important measure when determining an invention’s future success. Virtual reality was presented as a groundbreaking revelation that would soon take over the gaming world. This reality has not yet panned out and several devices have failed to achieve broad market adoption. From a technological standpoint, the device is a feat and extremely novel, so why has it failed to catch on? In lecture, we learned that “market success often depends more on consumer convenience, price, and content availability than on technical superiority”. Reflecting on my own use of VR, my most played games are Golf+ and Eleven: Table Tennis, both sports simulators that closely mimic real-life sports. Indeed, most of the highest-selling games are very simple and emulate real sports or activities. This has a straightforward reason: as a game gets more complex, it becomes more expensive and more likely to cause motion sickness. So, while Meta made an engineering beauty, the limitations of its product mean it fails to offer novelty to the consumer. This mistake was noticed by Apple who took a new route, mixed reality. Mixed reality goggles offer VR plus Augmented reality (AR), the ability to integrate apps into real life visuals. This invention is not nearly as impressive of a technological leap but offers much more novelty to the consumer since it can be used in situations outside of gaming. For example, you could use the device to seamlessly watch a Zoom presentation while walking to class. Further, it has the potential to bring an innovation from the entertainment industry to several new industries (AR could be used as a work assistant or training tool). Interestingly, Apple has also failed to capture a significant market, though the high price is usually blamed. This development demonstrates the importance of creating affordable novelty for the consumer when innovating. Inspired by the slow adoption of VR, The equation below attempts to capture the factors that influence adoption rates of cutting-edge technology.

Adoption= Functionality + Quality + Novelty + Consumer Convenience + 1/Price

Functionality is the same value for all inventions of a similar type (e.g. all AI models would have the same functionality value) and reflects how useful that invention is to society (e.g. functionality of AI>>>VR). Quality will mean different things for different industries. For example, in VR quality might reflect the number of apps available and picture resolution while in AI quality might reflect model speed and response accuracy.

Applying this equation to the VR space, Apple realized that the novelty of current VR headsets was too low and tried to push a new, mixed-reality paradigm. While they succeeded in raising novelty, they also raised price which removed any adoption growth possibility. I predict that as Apple lowers price of the VisionPro (if possible), they could supplant Meta as the primary device maker.

[carrieboone]

In class, Professor Evans mentioned that an AI model can predict 95% of research publications that will be published in the next year, but the 5% that it misses are the best ones. Venture capital funding is allocated to new ideas based on statistics of past companies. VCs create risk profiles for companies and funds are allocated with the aim of minimising risk based on reward. This profiling is done based on information about past companies and current financials; in this way, a prediction is made about the company like an AI makes predictions about what research might be successful. VCs are increasingly using AIs themselves to predict what start-ups might be successful, which evaluate companies using metrics ranging from employee headcount to LinkedIn followers.

I wonder whether venture capital predictions miss the same top 5% of companies that AIs evaluating research projects miss, preventing them from receiving any funding at all and missing the opportunities that they bring. While the paper “Modularity, Higher-Order Recombination, and New Venture Success” discusses surprises in the sense that certain companies have different, novel combinations of features, it does not discuss a “surprise” company in the sense of a company that has seemingly average quality and still succeeds. This proposition is still somewhat different to a new research idea not yet being predicted by AI, because there are finite tangible companies that can be evaluated at any given time, but there are millions of ideas and research papers yet to be thought of; I think this would decrease the strength of my argument.

Still, we can create a model that simulates an added randomness value to VC funding, such that it allocates some funds completely randomly among firms that it would not have invested in to capture some of those top unpredictable 5% firms. Below is a graph of the result. The Bayesian uncertainty graph represents how uncertain the AI is about the predictions that it makes, and the surprise graph represents the novelty of firms. The added randomness in funding just flattens the distribution and increases variance. Further, this model is inherently flawed. The metric being used to decide whether a company is successful is determined by AI, which is still incorrect even for its uncertainty measures for those 5% of firms that succeed. So, there is no way to model the increase in return from random fund allocation predictively.

[joezxyz]

AI: The Base Molecule of Innovation, or Hindrance to Creativity?

I would like to take the concepts from [Assembly theory explains and quantifies selection and evolution](https://www.nature.com/articles/s41586-023-06600-9) as the basis for analyzing AI and its relationship to innovation and society.

In the world we live in currently with the rapid progression of AI, most people don’t consider what goes into the AI’s development. AI is not an answer bank. Put simply, it’s a receptacle of constant progress developed through research, and module training. AI as a technology has its own molecules in the form of data and algorithms to determine its growth. These inputs are in the end, human created and thus AI actually grows at the evolution rate of human creativity. Therein lies the issue.
Even if you were to argue about the supposed “evolution” of AI, the fact remains that AI will NEVER(for now at least) be able to output anything further than what those molecules and selections made by humans can. Therefore, our perspective should not be to have AI develop, select, and evolve on its own, but to have AI be a base molecule of human innovation.
Imagine the AI as a bank not of answers, but of ideas. Each of these ideas will be considered a molecule, or base unit of development. The AI will obviously know more facts than you, but past those facts, it cannot do more. This is where humans can come in to be the progressor of innovation. In a study done at MIT, the ability of quality and effectiveness in work when using chatGPT alone, as a toll, and a control(no gpt) were measured against each other. In this process of human innovation and selection, it can be seen that on average, the consistency and quality of GPT seem to exceed the capabilities of humans with and without the GPT support.

However, upon closer examination, it can be seen that despite the higher average quality from GPT alone, there is at the very end of the graph, a level of quality that surpasses what the AI could provide. Going back to Assembly Theory, it explains that “With the discovery of new unique objects over time, symmetry breaking in the construction of contingent assembly paths will create a network of growing branches within the assembly possible”.

The humans reflecting and inputting in conjunction with the AI allows for feedback loops and development in a larger-scale, intentional assembly process. The human selection process allows for the AI components to be aligned and added upon for adaptive outputs towards broader ranges and goals.
Taking these studies into consideration, what is important in the construction and usage of AI, is not simply applying it, but being very intentional with questions, answers, and feedback as to find that small percentage of innovation that can surpass the collective whole of module and algorithmic training. The AI is here not to express the innovation humans already have, but to further enhance humanity’s ability to innovate.

[dannymendoza1]

How Technology’s Modular Structure Propelled AI to New Heights Exponentially Faster than Other Technologies

Chapter 2 in Arthur’s The Nature of Technology describes the concept of modularity as one of the key components for the structural integrity within any given technology. We learn that technologies are composed of different assemblies and subassemblies, and such a structure permits the technology at hand to thus execute and perform its function. Artificial Intelligence, by definition of being a technology, thus contains this modular structure. But why is it that AI seems to have suddenly boomed at such an exponential rate compared to other, perhaps more “physical” technologies? At the general level, Arthur would argue that AI and any other technology ultimately have the same structure, so what, if anything, is unique about AI’s modular structure that is acting as a catalyst for the technology’s recent success?

The answer lies in the fact that reconfiguring AI’s modular components in order to experiment with different versions of the same model and find the optimal one, actually turns out to be an incredibly efficient process that continues to experience decreasing costs as computers become more and more advanced in their ability to process large amounts of data. To make this very clear, let’s take two technological examples: the transformer architecture of large language models like OpenAI’s GPT, and the hierarchical structure of the F35C aircraft Arthur describes. As computers become even more powerful, not only can these AI models accept new, and likely improved, data as initial inputs at the lowest level of the structure of the model, but it can also experiment with dozens, hundreds, thousands, even millions of different parameters as part of the next layers within the modular structure. This, along with increased computational power, provides for rapid prototyping and experimentation of hundreds of different AI “models” that can all be performed within a matter of a few days, hours, or even minutes depending on the complexity and size of the models. Now imagine trying to even experiment with just two or three different aircraft models. Simply from a cost standpoint of having to either build two or three different aircrafts, or build one, test it, make changes, test it again, make changes, test it again, and so forth, is ineffective both timewise and economically speaking. The ability for AI models to seamlessly experiment with so many versions of the same technology is what allows for innovation to really take place within the AI world, and thus proves how its modular structure differentiates itself from that of other technologies. The three graphs below illustrate the exponential growth seen in recent years regarding key components of AI’s modular structure: data points, parameters, and computational ability, all of which have contributed to AI’s rapid emergence and progression as the technology of the future.

Citations:
https://www.instill.tech/blog/modularity-in-ai
https://ourworldindata.org/scaling-up-ai?tab=t.0

[joycecz1412]

Many of the readings this week focus on the concept of how to identify degrees of novelty of an innovation. Whether it be the distinction between context vs. content, point vs. application vs. system solution, or higher vs. lower order recombination, all of these categorizations demonstrate that the least predictable (and thus the most innovative) idea is one that combines modular knowledge from a variety of disciplines to create new systems of understanding or experiencing the world.

In essence, true creativity is unpredictable, or predicted to be improbable. Thus, in the age of AI, we should truly think about how we can best incentivize and educate people to think creatively.

Going back to the topic of my first memo, there are many ways in which AI can completely revolutionize the education system. A lower order recombination of AI is already in use in some classrooms today. They include point solutions like customized lesson plans and/or assessments, graders, etc. These applications are meant to improve teaching efficiency and increase personalization in learning, which should lead to better quality of education for the student.

However, with the existence of ChatGPT and other LLMs, most pre-university level homework and assessments have been rendered utterly useless. There should be fundamental change to the education system, from the foundational curriculum to the way in which students are assessed. How can we use AI as a tool to encourage students to think critically and creatively, with an emphasis on their ability to synthesize knowledge in different fields?

We can take inspiration from the model in “Modularity, Higher-Order Recombination, and New Venture Success” to assess the degree to which school curriculums and textbooks are facilitating creativity and pushing students to engage in higher order recombination, in the same way that stocks can be assessed on innovative value based on semantic analysis. For example, we can create dynamic word embeddings of textbooks to analyse the semantic space. How are math textbooks connecting to other topics like science, history, or music? Or we can create embeddings of searches made by young students to actively incorporate topics they are interested in. The figure below is a mapping of how people’s interests may change over time, and how our education system should be evolving to capture these forefront topics.

In this sense, AI can help facilitate the creative destruction of “unhelpful” instructional materials—ones that have limited semantic space and do not help students make connections between theoretical subjects and the world around them.

Similarly, students can be assessed on their ability to form higher order recombinations in course projects. Rather than a focus on facts or knowledge, the appearance of AI should shift education towards what really matters: thinking logically and creatively. Empirically, we could design two types of courses for students, varying in content novelty but similar in technical difficulty, then assess their creativity (probability of surprise) and ability to make associations and recombinations.

[henrysuchi]

In this memo, I provide a numerical analysis of the growth function described in the description of assembly theory (AT) in Sharma et al (2023) and discuss its implications. In this setup, there is a set of unique objects which can combine to create new ones as time goes on. The size of this set is called $N$. There are different levels of “complexity” of objects based on how many steps it takes to create them, which is the assembly index $a$. At each time stage, some or all of the objects in the set are available to be combined into a new object, which is represented by the selection parameter $\alpha$. Sharma et al state that the growth in number of objects can therefore be characterized by the equation $$\frac{dN_{a+1}}{dt} = k(N_a)^\alpha$$ where $N_a$ is a function of $t$ for all $a$. Since this is a differential equation, then it can be solved analytically or numerically. Sharma et al use Wolfram Mathematica to dynamically solve this equation, but we can also use numerical methods to do the same.

I implement an algorithm that loops through different selection parameters and solves for the functional forms of the count $N$ as a function of time $t$ as it recurses through different assembly indices. For each step of the loop, I call the Runge-Kutta algorithm to find this function, as it is essentially an initial-value problem, which can easily be solved via numerical means. I call $\alpha$ a, and to avoid confusion, we call $a$ n and $N$ X. Then the below is a graph of how the functional form of $X(t)$ varies as we change the selection parameter and the assembly index. Note that the assembly index varies and leads to greater growth as there are more steps or more progress is made in growth.

Now I discuss the results. The growth model provided by Sharma et al is not dissimilar to a simplified one-input version of Solow’s growth model. If the only input is capital, then the growth function $y = Ak^\alpha$ is not dissimilar from the simpler version of the growth function above. However, there are two main differences between the AT model and the Solow model or other similar macroeconomic growth models that have microeconomic foundations. First, the economic model assumes concavity, so the shape of the graph is concave compared to the ones plotted above. Second, model with micro models assume purposiveness and optimization. While the AT model is exogenously given a selection parameter which dictates which objects can be combined for assembly in each period, in Solow/NCG type models, the amount of capital invested into production is endogenously determined by the utility maximizing agent. This therefore implies that in the long run, simulations using the AT model will have different growth predictions compared to macroeconomic models with neoclassical assumptions.

[ggracelu]

Reacting to New Technology: Explicit and Implicit Factors of Novelty

As a Sociology and Economics double major, I was intrigued by the potential use of AI as simulated human subjects for research discussed in “Simulating Subjects: The Promise and Peril of AI Stand-ins for Social Agents and Interactions.” I agree with the potential rooted in the fact that AI models are “trained on products of culture and interaction” so that “these models are reflections of our socio-cultural world” (Page 2). However, culture is a complex, dynamic phenomenon in which individual actors and their surrounding environments influence one another, often implicitly and non-sequentially. The sociology of culture emphasizes embodied cognition as an important aspect of the development of culture since our sensory systems directly mediate how we process and interact with the world. As such, the impoverished sensory experience of AI simulated subjects not only limits specific experiences like gender bias, but it may imply fundamentally constricted understandings of culture.

When reading “Modularity, Higher-Order Recombination, and New Venture Success,” the examples of modern low-order inventions such as CRISPR gene-editing, quantum processors, and solid-state batteries reminded me of the notion that “ideas are getting harder to find,” which I still am inclined to disagree with. Rather than people being less innovative, other forces such as high funding costs for research and long time to market mean that high-order inventions emerge more frequently. As a result, radical low-order discoveries are deprioritized since they require long-term investment to bear results.

In Thursday’s lecture, we discussed the importance of historical experiences in the degree of novelty of a new technology as a crucial component of innovation. Building off my previous interest in the cultural influences on innovation, I am interested in the role of culture in the amount of surprise generated by new technologies. For instance, if two cities had very similar infrastructure and explicit historical experiences with technology, how might implicit cultural factors impact their reaction to the same new technology? Additionally, how does culture influence the utilization of technology (ex: valuing efficiency)?

I created a flow chart to visualize the relationship between cultural background and adoption of novel technology. My focus was to emphasize the role of both explicit and implicit factors in shaping the current model of technology that is compared with the new technology to determine novelty using divergence.

[pedrochiaramitara]

Modularity and Supply Disruptions

In the paper “Modularity, Higher-Order Recombination, and New Venture,” a large importance is placed on the power of modularity to drive innovation, as one can divide complex systems into independent and replaceable components, which makes innovation easier. However, especially in today’s globalized world, this can have drawbacks when geopolitical tensions rise. For instance, semiconductors are essential parts for almost everything from smartphones and cars to data centers, but most of these chips come from a small number of places, especially Taiwan and the USA. The United States, by far the largest market-cap holder in the industry, relies heavily on specialized Taiwanese manufacturers for advanced chips, as most of the U.S. industry was built assuming these chips would be available. If rising geopolitical tensions between China and the U.S. were to lead to a problem in Taiwan’s ability to supply the U.S. economy, the country could experience a drop in inputs for consumer electronics, automotive applications, and defense systems, negatively impacting the economy and hampering innovation. Many new technologies rely on these advanced chips, as they built their modular products with them in mind. Without access to them, progress could stall. Companies would have to invest in R&D to replace the modules that were disrupted instead of developing new products or improving other components, and some might find that their prior innovations are no longer feasible.

Below, the graph I made shows the sum of chip-selling companies’ market cap by country. The dominance of the U.S., Taiwan, and China demonstrates how the global value chain depends on a handful of countries. If Taiwan’s capacity were compromised, U.S.-based firms would have limited replacement options.

A basic way to show the impact of disruptions is a simple Cobb-Douglas production function:

Where:
Y= Output, K=capital, L= labor, and A = technology available

When geopolitical tensions or a potential invasion of Taiwan threaten imports of critical chip components, A diminishes. Since A measures all forms of technology and necessary inputs for more productive industries, losing a key supplier lowers efficiency. The supply curve shifts left, as less output is produced, driving up the cost of semiconductors. This causes problems in all industries, such as AI.

The impact is significant in the long term as well because to recover the lost technology, the U.S. has to devote more labor and capital to developing or replicating what was lost. This represents a considerable opportunity cost, as people and capital that could be employed elsewhere are effectively “wasted”.

To conclude, while modularity is good for rapid and flexible innovation, one must be careful not to be dependent on a single module, especially when there is a limited number of suppliers. The solution might be on fostering modularity by building alternative methods that don’t depend on foreign industry.

[dnlchen-uc]

Connecting VC Funding with Prior Discussions of Innovation

In "Modularity, Higher-Order Recombination, and New Venture Success", Cao, Chen and Evans, argue that venture capital structurally rewards higher order innovation. Startups without sufficiently novel products tended to struggle to in obtaining further funding rounds in the high-intensity venture capital ecosystem. However, venture capital is not an ubiquitous asset class on the global scale. In Europe, high income economies notably do not put as much weight as the US on the asset class. According to the Ivey Business Journal, Germany and Switzerland have relatively low VC capital per capita and other countries like France have unique regulations on venture capital which operate counterproductive to fundraising levels. As discussed in past weeks, these countries are often on par with the US in terms of innovation quantity. In this memo post, I analyze the relationship between VC funding and innovation in terms of patent applications.

This graphic describes the per capita VC funding for 24 countries (selected by Crunchbase to represent regions with highest VC potential in coming years). These figures match the observations made by the Ivey Business Journal. Apart from Estonia, Sweden, and the UK, European countries have significantly lower VC spending than the US and other top performers such as Singapore and Israel. However, adding per capita patent data (from the World Bank) into the graph reveals that patent application per capita, which we've used in previous weeks as a proxy for a country's innovativeness, is largely uncorrelated with venture capital.

A potential explanation is the influence of incumbents on innovation. In the case of Germany, which tracks the US closely, the majority of patents were filed by incumbents rather than venture-backed startups. Preliminary research indicates that the same is true for a lot of the other European countries without strong startup ecosystems. Another interesting example to examine further is Estonia. Despite having an extremely strong VC ecosystem in terms of funding, per capita patents is relatively low compared to many of the other countries analyzed. This could be a result of Estonia's competitive tax system relative to the rest of the EU and other OECD countries, which would also explain Ireland's position in the graph. Due to lower taxes, foreign based startups (which countribute patents to other countries) may opt to raise funds in these countries. On the other end of the spectrum, China and South Korea have low VC funding per capita but exceptionally high patent applications. The incumbent explanation doesn't seem to make much sense especially in the case of China, which does not have the long-standing institution base found in Europe. In coming weeks, these would serve as potential case studies to look deeper into.

Furthermore, patents are not a perfect metric to capture the quality of innovation. Even though venture-based higher order innovation is the source of much of US patents, further examination is needed to determine the character of patent innovation in other environments. Is incumbent based innovation in Europe likely to be higher order due to already existing infrastructure which innovators can build off of? Do higher patent application volumes in Asia reflect higher levels of creative destruction, or do they result from government incentives to reach certain research goals?

[Dylanclifford]

In our reading of "Modularity, Higher Order Recombination, and New Venture Success," Cao, Chen, and Evans argue that successful ventures usually come about through the combination of established and functional modules rather than through novel and lower order combinations. Thus, I intend to analyze this theory through the lens of CEO educational backgrounds, analyzing whether the combination of engineering and business education (arguably representing distinct knowledge modules), leads to superior venture outcomes.

My analysis of 6,363 startups reveals interesting patterns that challenge conventional wisdom about educational modularity.
Using annualized returns as my success metric (Where I defined success as annualized return > 0) , here are the results:

1. Engineering only backgrounds show the highest success rate (47.82%) and better average returns (-0.180) compared to other educational combinations.
2. The combination of both engineering and business education which might be expected to represent beneficial higher order recombination actually in fact shows the lowest success rate (37.26%) and worse returns (-0.327).
3. These differences are statistically significant with ANOVA p-value = 0.003 which indicates that educational background meaningfully impacts venture performance.

(For context, startups with 0 returns have annualized return of -1)

These findings does seem to contradict the paper's modularity thesis when applied to CEO education.

However, a deeper analysis suggests several potential explanations that actually could reinforce the paper's broader theoretical framework:

• Timing of Module Acquisition: The paper emphasizes that successful higher order recombination requires modules to be already proven and functional. The sequential acquisition of business education after engineering experience (represented by my "Engineer to Business" transition metric) might not provide the same benefits as deep expertise in either domain.

• Module Integration Costs: While the paper focuses on technological and market modularity, educational knowledge modules may carry significant integration costs. The cognitive overhead of maintaining expertise across disparate domains could outweigh the benefits of module combination.

• Depth versus Breadth: My results suggest that depth in a single domain (particularly engineering) might be more valuable than breadth across domains for early stage ventures. This aligns with the paper's emphasis on the importance of robust and well established modules.

The attached graph above visualizes these patterns, where it's highlighted that the counter intuitive relationship between educational combinations and venture success rates. The stark difference between engineering only and combined engineering business backgrounds (47.82% vs 37.26% success rates) suggests that the benefits of modularity may just be domain specific. These findings have relevant implications for understanding how modularity theory applies to human capital in venture creation. While the paper's framework about higher order recombination may hold true for technological and market modules, my analysis suggests that educational modules may operate differently in reality. Future research could potentially explore whether this pattern holds across different industries or stages of venture development.

This ultimately demonstrates that the relationship between modularity and venture success is really more nuanced than previously understood, specifically when applying these concepts to founder characteristics as opposed to the likes of technological or market combinations.

[grozdanickata]

This week’s reading on “Novelty” as a violation of expectation, especially the idea of assigning novelty scores to various innovations across fields based on combinatorial characteristics of innovations’ peculiarity, atypicality, and obscurity prompted me to think of other characteristics or models that could be indicative/ predictive of impact. I completely agree that when thinking about aspects of novelty in present day innovation, the interpretations of “novelty” are very subjective to the context of the field that the novel innovation is introduced in. Most of the models in the paper seem to account for the subjectivity and variability of various aspects potentially affecting the novelty scores in each distinct field. However, I also believe that there is a general, categorical element of novelty, that can be applied across all fields, which I would like to propose:

Historically, when I think about the timeline of inventions, innovations, moments of growth that I have both lived through and learned about throughout my primary and higher education, I see the timeline as the following categorical progression, listed from the furthest distance in time, to the nearest distance in time to present day:

Practical Discoveries and innovations for the foundations of human survival (more like tools, not related to “modern technology” in our most contemporary definition of the term). One example of this is the discovery of fire, and the fact that we can create it ourselves, and use it to keep us warm and alive, or cook our food, or sterilize and heal our wounds. Some similar such innovations that followed are the invention of the wheel, which completely revolutionized human transportation.
Theoretical and Academic Discoveries or Fundamental Truths. These relate to more philosophical discoveries that shattered previous human understanding of the world so long ago, and so intensely that they are still widely taught to children in elementary schools. Examples include the theory of evolution, the disproving of the flat earth theory, the discovery/innovation of the pythagorean theorem or other various instances of mathematics in nature.
Incremental Breakthroughs/ Inventions. These are what I think of as resulting from combinations of the previous two categories. For example, the invention of the printing press, railroad systems, etc.
Finally, the most current category of innovations are technological ones. Electricity, the first computer, smartphones, the internet, and now artificial intelligence.

I believe that present day innovations in category 4 have the lowest novelty factors. I believe that it is relatively least shocking to predict ahead chronologically, as for years people have had the ideas (but just not the means) to create the things that they imagine, such as flying driving cars/ various types of robots, etc. It is most natural for us to think ahead from the most recent discoveries which have been built on centuries and centuries of previous fundamental discoveries. But what would be most unexpected is to predict a type of invention or breakthrough that can be thought of as going “backwards” in the above timeline. No one would be able to predict a new invention such as the discovery of a new natural element equivalent to fire, because we believe that we have already mostly discovered what there is to discover in that area— such an invention would have the highest novelty score of the 4 categories. Discovering a new fundamental truth about our world that changes our perception and way of life altogether, such as (for example, what happens after humans die/ empirical evidence of an afterlife), would have the second highest novelty score in this spectrum.
Another concept to consider for novelty especially when taking into account “impact” is related to accessibility— an estimation of the percentage of the global population the innovation or breakthrough will directly affect within 10 years. For example, while the internet is a revolutionary invention (1983), there are still so many areas in the world where people do not have wifi access. The same goes for smartphones, AI, etc.

I would propose the following equation which acts as a novelty score multiplier, that can be incorporated into each of the model and field specific novelty equations proposed by the paper.

Novelty Score Multiplier = C x P

Where C is a numeric weight based on the above 4 categories:

(1)Practical Discoveries for foundations of human survival: C = 8
(2) Theoretical and Academic Discoveries or Fundamental Truths: C = 6
(3) Incremental Breakthroughs/ Inventions: C = 4
(4) Technological Inventions: C = 3

, and P is an estimated percentage of the global population which the innovation will directly affect and be easily accessible. Ex: a category 1 innovation would be between 0.9 and 1.

*The C values assigned to the 4 categories are relatively arbitrary at this point and can be adjusted so that they are non linear.

*The exact method for the calculation of P would have to be refined and clearly defined to include various aspects of global reach across the environments in different regions of the world.

[amulya-agrawal]

Utilizing LLMs, Ideas of Novelty, and the Random Walk Model to Analyze the Future of Job Interviews Conducted by LLM Models

In Stimulating Subjects: The Promise and Peril of AI Stand-Ins for Social Agents and Interactions, LLMs are discussed as a power application to simulate empirically realistic, culturally stimulated agents. I am curious to see how LLMs can be utilized to foster artificial relationships between humans and companies looking to hire employees.

LLMs, like Chat GPT, can simulate culturally and contextually nuanced agents with trained personas to adopt when interacting with humans. This article mentions how LLMs can cultivate meaningful social relationships, such as by simulating the perspective of a farmer from southern Indiana. Besides an actual human, the second best alternative to simulate a human’s perspective would be an LLM model because of the expansive networks of evidence it sources. Using a zero-shot learning model – where users tell the model in natural language what persona or style to mimic in its response, users can initiate a user prompt to provide their question along with a system prompt to provide background on how the model should respond to the question. Few-shot learning can also be utilized here to allow specification of the desired style of response through examples, rather than direct instruction.

So, let’s think about this as if you are in a job interview where an LLM model, like Chat GPT, is your interviewer. This could be a revolutionary way to conduct interviews in the future without needing a human interviewer.

System Prompt: Keep in mind the background that you are a 45-year old labor economist for the World Bank with expertise in global workforce dynamics. Your role is to assess a candidate for a policy advisory position at your workplace – the World Bank.

User Prompt: Ask the job candidate questions like the following – “If you could work on a case for any labor economics problem, what would it be?” Ask follow up questions with contextual relevance based on their responses and your expertise.

This approach allows organizations to evaluate candidates under consistent and unbiased conditions, leveraging the LLM’s capabilities to recombine knowledge across fields, including labor economics, behavioral psychology, and organizational dynamics.

I argue that this is a “novel” idea because the higher-order recombination of modules enables LLMs to solve a problem that would be impossible using isolated technologies. By recombining knowledge from disparate fields – such as LLM tools, professional knowledge, and knowledge about different industries – you can simulate dialogues that yield unexpected insights about a job candidate.

For example:

1. A candidate discusses using blockchain to ensure fair wages in global supply chains. The LLM model, drawing on its knowledge of both blockchain and labor markets, asks them: “How would you address scalability and trust issues in decentralized systems across different regulatory environments?”
2. This follow-up symbolizes a low-probability path – a question that would likely not arise from a traditional human interviewer without blockchain expertise – uncovering the candidate’s capacity to integrate disparate fields.

We can formalize novelty using the knowledge framework equation from the Surprise! Measuring Novelty by Simulating Discovery text:

Novelty = g(P(D|K, S)), where K is prior knowledge (ex. labor economics) and S is the search process (ex. candidate-AI interactions and searching for insightful follow-up questions to ask). P(D|K, S) represents the probability of discovering D (ex. an insightful question about blockchain’s scalability in supply chains). A novel question is one with a low P(D|K, S). This is unexpected, but insightful given the context. Here, high novelty corresponds to low perceived probability.

To visualize the novelty and discovery process in LLM-led interviews, I will apply the random walk model to simulate the progression of modular knowledge recombination during the interview.

Here, the nodes represent knowledge components, such as:

1. Candidate Responses (ex. “Blockchain ensures wage transparency”)
2. LLM Interviewer Prompts (ex. “How would you address trust in decentralized systems?”)
3. Related knowledge (ex. “Regulatory challenges in developing countries like India”)

Edges, on the other hand, represent relationships or transitions between nodes, which is determined by the LLM’s understanding of combinable knowledge. Random walk here, is a path taken through the graph by moving from one node to another based on probabilities.

An example Random Walk could look like (for this particular interview example):

1. Start Node: Candidate is asked a question about working on a labor economics problem, and the candidate introduces utilizing blockchain for fair wages.
2. First Transition: LLM links the blockchain idea to trust issues to ask “How do you ensure transparency and scalability as an economic analyst?”
3. Second Transition: Candidate then mentions case studies in Asia, which prompts the LLM model to ask “How might regulatory differences impact implementation in a country like India?”
4. Third Transition: LLM introduces a novel connection to global labor standards and enforcement.

Novelty of a combination is inversely related to the probability of encountering it during the random walk, which is seen through the following formula:

Pr(X(t) = j|X(0) = i) = [e^-Qt]ij

Here, Q is the transition matrix capturing probabilities of moving between knowledge nodes, while t is the time spent exploring during the interview.

Overall, this novel innovation demonstrates how AI, through higher-order modular recombination, can transform workforce hiring into a process that values both novelty and depth. These interviews can provide consistent, bias-free interactions and explore candidates’ unique problem-solving skills through low-probability, high-impact questions – providing depth as the interview goes on. As such, this technology has the potential to redefine (just like we have seen through observations with dynamic word embedding models in the Modularity text) the concept of global recruitment standards truly looks like.

[LucasH22]

Innovation in Parallel and Preemptive Combinatorial Evolution

I found the demand-side story of innovation in Brian Arthur’s book The Nature of Technology: What it is and How it Evolves ripe for further exploration. Arthur argues that technologies emerge to fill opportunity niches that are created by human needs, and more prominently, technology itself. In particular, he elaborates that “[firstly,] every technology by its very existence sets up an opportunity for fulfilling its purpose more cheaply or efficiently,…[secondly,] every technology requires supporting technologies to manufacture it, organize for its production and distribution, maintain it, and enhance its performance,...[and thirdly,] technologies often cause problems–indirectly–and this generates needs, or opportunities, for solutions” (175-176). This recursive nature of technology demand prompted me to consider the repercussions of “combinatorial evolution” happening in parallel. For example, what happens when innovation actors begin to assume that a recursive chain of technologies will constantly improve?

One case study lies in semiconductor design and manufacturing. Back in 1965, semiconductor designer Gordon Moore observed that the number of components per integrated circuit had been doubling every year, spawning the now infamous “Moore’s Law” that projects the number of transistors in an IC doubling every two years. Moore’s Law has held true up until now, and some argue that there has been a “self-fulfilling prophecy” dynamic at play. As chips became more complex, technologists at each stage of the chip manufacturing process set their long-term R&D planning according to this coordinated target.

I argue that this behavior demonstrates the existence of preemptive combinations, which I sketched in the graphic below in the spirit of “Assembly theory explains and quantifies selection and evolution”. The innovations across lithography, ion implantation, deposition, etching, packaging, and myriad other stages of the process were not solely combinations of innovations in the past; they also relied heavily on innovations being designed in lockstep. Thus, innovation C (say, EUV lithography) depended not just on prior innovations A and B, but also on concurrent innovations D and E (e.g. new masking and photoresist innovations). While skeptics could argue that each new “node” or semiconductor manufacturing process like 3nm is one cohesive innovation, the reality is that the innovations at each stage of the process are applied to other use cases. In this way, coordination or preemption of adjacent innovations at the same number of assembly steps jumpstarts distinct innovations in parallel.

Another relevant example of this innovation in parallel, or preemptive combinatorial evolution, is the current AI scaling law. At every level of the AI stack, from semiconductors to foundation models to AI applications, technologists are innovating with the assumption that the other pieces will innovate and come into place as well. Take foundation model developers, who are improving their data and algorithms with the assumption that more effective chips will be harnessable in the near future.

This model of innovation demonstrates how technology’s recursive loop can become a formidable flywheel if technologists coordinate to innovate in parallel.

[michellema02]

Do Industry Characteristics Shape the Success of Novel Ideas?

This week, we explored the relationship between novelty and impact in the context of innovation. For example, a study titled "Surprising combinations of research contents and contexts are related to impact and emerge with scientific outsiders from distant disciplines" finds that two dimensions of novelty—context novelty and content novelty—strongly predict the impact of both patents and scientific research. However, during class, Professor Evans raised an intriguing caveat: while the most novel innovations are, on average, more successful, their success is also marked by greater variance. Specifically, he noted that highly novel innovations are more likely to face significant challenges in implementation or paradigm transition. As a result, despite being objectively superior or even game-changing, they might fail entirely.

This was a fascinating observation, but it made me wonder whether it applies equally across all industries. Some industries are characterized by more entrenched systems, often due to the complexity and scale of the physical capital required. In such cases, I could imagine that even lower levels of innovation might result in significant variance, as incremental changes are more likely to be selectively adopted or ignored, while radical changes face uniform resistance. For instance, Brian Potter argues that innovation in construction is particularly challenging because project costs are right-skewed and fat-tailed, driven by the compounding effects of delays and other factors. Novel systems that could dramatically alter construction processes often incur steep risk penalties that outweigh their expected benefits. As a result, innovations in construction tend to be more "incremental and evolutionary."

Validating this theory fully would require more time and data than I had for this memo, but I conducted a preliminary investigation into some relevant evidence. As mentioned in class, venture capitalists tend to pursue the most novel innovations, accepting that most of their investments will fail in exchange for a few outliers that generate outsized returns. Given this, the industries that VCs favor might offer insight into how tangibility (as a proxy for capital requirements) relates to novelty and variance.

Plotting the data, I found that nearly five times as much VC investment goes toward intangible goods and services—industries with lower capital requirements—compared to tangible goods. While this is an interesting result, I recognize its limitations, particularly the influence of confounding factors such as the higher profit margins associated with intangible goods, which don’t necessarily speak to variance in innovation success within those categories.

Given more time, it would be interesting to use more sophisticated techniques, such as analysis of patent keywords, as well as higher quality and more comprehensive data, to see if the relationship between novelty and variance holds consistently, or whether it differs based on the structural characteristics of an industry.

[pauline196]

The reading “Surprise! Measuring Novelty as Expectation Violation” provides an overview of various methods for measuring novelty and introduces a framework based on the concept of surprise associated with invention. Novelty is significant because it positively correlates with the variance in success outcomes. Within the surprise framework, I wondered: when medical drugs that have not been used before are introduced, are they still considered innovations if they lack the element of surprise? For example, the COVID-19 vaccine dominated its domain, changed the landscape, and was widely adopted, yet it seemed to be an expected development rather than a surprising one.

Another point mentioned in class was how drug prices typically drop once a patent expires, as heightened competition diminishes inventor’s market dominance. So I became interested in how quickly prices respond to patent expirations. Specifically, I wanted to analyze the statistics on the average prices of drugs whose patents expired in recent years. This is interesting because the number of patents increased significantly between 1996 and 2003 (as shown in Figure 1), meaning many of these patents were supposed to expire within the past eight years. This phenomenon of the abrupt drop in sales and market share loss that often follows patent expiration is known as the "patent cliff".

For this analysis, I used data from Medicaid and the Children's Health Insurance Program on the weekly average prices pharmacies pay for prescription drugs from 2013 to 2025 (available at https://data.medicaid.gov/dataset/a217613c-12bc-5137-8b3a-ada0e4dad1ff#data-dictionary). I focused on the average price per milligram or per dose, as some drugs come in different dosages. I aimed to select drugs with patents that expired in recent years, but it was difficult to search for patent expiration dates. To overcome this, I used several websites listing expired patents and cross-referenced the information. In the end, I analyzed four drugs (see Figure 2), with red lines indicating the approximate dates of patent expiration or the dates of allowed realize of the generic.

• Sprycel [1] which treats chronic myeloid leukemia was a major success for Bristol Myers. A settlement with XSpray allows the launch of a generic version on September 1, 2024. The graph shows that the average price plateaued in 2024, slowing the price growth trend seen over the previous seven years, which is likely a result of the announcement and realize of the generic.
• Dulera [2] is a medication used to treat asthma and chronic obstructive pulmonary disease. The drug, owned by Organon LLC, was protected by 10 U.S. patents filed, all of which expired. The generic version was expected to launch on February 12, 2023. The graph depicts a drop in Dulera's prices around that time, although prices had already risen significantly from 0.175 to 0.28 over the two years leading up to the generic launch.
• Myrbetriq [3] is a drug used to treat overactive bladder in adults. The generic form of Myrbetriq has been launched in the U.S. in April 2024 after a lengthy legal battle involving the manufacturer Astellas. The graph shows a significant price drop of around 40% at the launch of the generic, followed by a quick rebound almost returning it to previous level.
• Victoza [4] which is a diabetes medication with liraglutide from Novo Nordisk saw its patents expire between mid-2023 and early 2024, allowing generic versions from Teva, Pfizer, Mylan, and Sandoz to launch. Victoza also saw a price drop around the patents’ expiration, reaching its lowest point in the last 7 years.

Overall, this highlights the significant impact of patent expirations on drug prices, showing how the introduction of generics right after the patent expiration often leads to price drops. A major limitation of this analysis is that, since most drugs are protected by multiple patents, they do not all expire at once, making it difficult to pinpoint a single expiration date. Additionally, finding the exact patent expiration dates was surprisingly challenging, and for some drugs, the available price data was too limited to include in the analysis.

Figure 1: Data from [5]

Figure 2

[1] Biopharma Dive. (2024, January 23). Bristol Myers’ Sprycel faces generic competition as patent cliff approaches. Biopharma Dive. https://www.biopharmadive.com/news/bristol-myers-sprycel-generic-patent-cliff-xspray/721197/
[2] Pharsight. (n.d.). Dulera patent expiration. https://pharsight.greyb.com/drug/dulera-patent-expiration
[3] MastersRx. (2024, January 23). Generic Myrbetriq approved and launched in the US. MastersRx. https://mastersrx.com/generic-myrbetriq-approved-launched-in-us/
[4] Fierce Pharma. (2024, January 23). Top 10 drugs losing U.S. exclusivity in 2024. Fierce Pharma. https://www.fiercepharma.com/special-reports/top-10-drugs-losing-us-exclusivity-2024
[5] United States Patent and Trademark Office. (n.d.). United States patents granted. USPTO. https://www.uspto.gov/web/offices/ac/ido/oeip/taf/us_stat.htm

[yanhong-lbh]

Memo: Transformer-Based AI, Novelty, and Cross-Disciplinary Innovation

Recent scholarship (Shi & Evans 2023; Foster, Shi & Evans 2020) underscores how “surprising” combinations of ideas from distant disciplines are prime engines of high-impact innovation. Building on Brian Arthur’s (2009) view that novel technologies evolve by combining existing building blocks, modern large language models (LLMs)—rooted in Transformer architectures (Evans & Kozlowski 2025)—offer a powerful technological catalyst for precisely these unexpected, high-value recombinations.

At a high level, Transformers represent text (or other data) as vectors in a high-dimensional space, allowing them to capture links across seemingly unrelated domains. In principle, this creates fertile ground for “emergent” insights, since architectural mechanisms like attention can integrate tokens from previously unconnected topics. This capacity dovetails with Assembly Theory (Sharma et al. 2023), wherein the recombination of simpler “parts” can open doors to novel, complex outcomes.

However, empirical evidence for LLM-driven innovation remains scattered, prompting a need to measure the extent to which LLMs facilitate cross-domain connections that surprise even human experts. One way to do this is by adapting the “novelty-as-expectation-violation” framework (Foster et al. 2020) to compare an LLM’s predicted connections to real-world disciplinary overlaps. Concretely, define a Surprise Score (S) for an idea ( x ) as:

S(x) = | LLM(x) - HumanBaseline(x) | / HumanBaseline(x)

where LLM}(x) is the model’s estimated “distance” between fields (e.g., robotics and microbiology) and HumanBaseline}(x) is a consensus measure of how distant experts judge those same fields. Higher S(x) indicates that the model identifies unusually surprising linkages.

Figure: Hypothetical Surprise Scores for Interdisciplinary AI Suggestions

     Surprise Score
       |       *         (Robotics + Microbiology)
       |               *
       |     *                (Finance + Linguistics)
   0.5 |  *
       |         (Marketing + Chemistry)
       |________________________________________
          Low                   High 
             Interdisciplinary Distance

In this mock plot, each point denotes a pair of fields (e.g., Robotics + Microbiology). The y-axis marks the LLM’s Surprise Score, and the x-axis represents an objective measure of disciplinary distance. Notice how “Robotics + Microbiology” registers a higher Surprise Score than “Marketing + Chemistry,” implying that the AI finds more novel synergies in the former pairing than do human experts.

This simple metric offers a concrete, tech-focused pathway to identify latent cross-disciplinary breakthroughs within AI outputs. It builds on Arthur’s concept of evolving technologies through recombination and aligns with the finding that “scientific outsiders” can foster especially impactful novelty (Shi & Evans 2023). As generative models mature, measuring and steering them toward surprising, interdisciplinary suggestions could greatly amplify their potential to drive meaningful innovation and growth.

[jacobchuihyc]

Agriculture remains one of the cornerstones of the African economy, inextricably linked with livelihoods and food security that underpin economic growth. Despite such importance, it has long been characterized by systemic challenges, including outdated technologies, inefficient methods of production, restricted access to resources, and susceptibility to climate change. These persistent issues have hampered productivity and restricted farmers' ability to scale operations or compete on a global level. But over the last couple of years, AI and automation have started to emerge as real game-changers: new ways in which these barriers can be tackled while fostering innovation and sustainability. This memo looks into the role of AI in African agriculture, as discussed in a project report titled "Inclusively Advancing Agri-Food Systems through AI and Automation", focusing on its application across the agricultural value chain, as well as at the Ethiopian Agricultural Transformation Agency's 8028 Farmer Hotline for an example of how AI works in practice.

The Generalized Agricultural Value Chain and Use Case Framework (Figure 2) puts a helpful lens on the broad applications of AI in agriculture, highlighting how these technologies are integrated across different areas of the farming process. Genomic innovation and digital financial advisory systems are aiding farmers in their decisions regarding inputs, planting strategies, and financial management at the Planning & Production stage, with reduced risk and improved productivity. Automation makes the chain more efficient in the phase of post-harvest management where food waste is minimized and value is maximized by demand-supply matching and automated processing. In the Distribution & Logistics section, innovation is projected in the form of smart contracts and automated transportation systems, which will chart the clear line of supply chains, therefore minimizing inefficiencies. Connectedly, AI-powered solutions link these stages together to create a well-organized structure for agricultural production that addresses long-standing inefficiencies while opening up new avenues for growth.

One particularly inspiring example is Ethiopia’s 8028 Farmer Hotline, which shows how much of a difference these technologies can make for smallholder farmers—the people who often struggle the most with accessing resources and information. The hotline, run by the Ethiopian Agricultural Transformation Agency, uses SMS and Interactive Voice Response (IVR) to provide farming advice in local languages, tailored to specific crops and regions. It even collects feedback from farmers and combines it with satellite data to deliver early warnings about crop diseases or pests, giving farmers time to act before disaster strikes. Since it started in 2014, the hotline has reached over 6.2 million registered users and handled more than 63 million calls. That’s not just impressive—it’s life-changing for the farmers who can make more informed decisions and avoid devastating losses.

What really stands out when looking at the value chain framework and the success of the 8028 Farmer Hotline is how they show both the big picture and the on-the-ground impact of AI in agriculture. These aren’t just abstract ideas—they’re real solutions addressing real problems. However, much more needs to be done for this potential to reach everyone, especially smallholder farmers. Governments and organizations need to invest in infrastructure like reliable electricity and internet access, while also providing training to ensure farmers know how to use these tools. And policymakers have a responsibility to ensure the fair distribution of AI benefits, so small farmers aren’t left behind while bigger farms take all the gains.

AI and automation are truly reshaping agriculture across Africa, and it’s not just about fixing inefficiencies—it’s about creating a future where farmers can thrive, communities can be more resilient, and economies can grow in ways we haven’t seen before. Ethiopia’s experience with the 8028 Farmer Hotline proves that when the right technology meets the right support systems, the results can be transformative. Scaling these innovations across the continent, thoughtfully and equitably, could unlock an entirely new chapter for African agriculture—one that’s smarter, more sustainable, and more inclusive for everyone involved.

[Adrianne-Li]

Measuring the Impact of AI-Driven Scientific Discovery on Research Productivity

Introduction

The integration of Artificial Intelligence (AI) in scientific research is transforming how discoveries are made. The recent literature on innovation suggests that AI has the potential to accelerate knowledge production by enabling unexpected recombinations of ideas and optimizing research pathways. However, empirical measurement of AI's contribution to scientific productivity remains a challenge. This memo evaluates the extent to which AI-driven research tools influence the rate and impact of scientific discovery, focusing on citation growth, novelty, and researcher efficiency.

Empirical Evidence on AI and Research Productivity

AI-driven discovery tools, such as DeepMind’s AlphaFold (for protein structure prediction) and IBM Watson (for drug discovery), have demonstrated measurable contributions to research outputs. A growing number of AI-assisted papers have been published in fields like biotechnology, materials science, and quantum computing.

Using data from Scopus and arXiv, I examined trends in AI-assisted research over the past decade. The results suggest:

• The number of AI-assisted publications has grown at an annualized rate of 25% since 2015, surpassing the general growth rate of scientific publications.
• AI-generated research tends to have higher citation impact, with an average 30% increase in citations per paper compared to non-AI-assisted research (based on data from Google Scholar and Dimensions.ai).
• The time-to-publication cycle has shortened in fields heavily influenced by AI, as automated data analysis and hypothesis testing reduce manual research time.

To quantify the relationship between AI integration and scientific output, I propose the following econometric model:

Where:

• Impact is measured as citations per publication and journal impact factor.
• AI Usage is a binary or continuous variable indicating AI involvement in research production.
• Collaboration Network captures the diversity of co-authors and institutional affiliations.
• Funding Level controls for differences in financial resources available for research.

Key Findings and Interpretation

AI Enhances Novelty in Research

AI-driven research exhibits higher novelty scores (measured by textual and conceptual distance in citation networks), suggesting that AI facilitates "surprising" combinations of ideas, as described by Shi & Evans (2023).

Automation of Data-Intensive Research

Fields with high computational demands (e.g., drug discovery, physics simulations) show the most significant productivity gains from AI, as automation reduces the need for trial-and-error experimentation.

Inequality in AI Adoption

• Leading institutions (MIT, Stanford, Tsinghua) account for a disproportionate share of AI-assisted research output.
• There is a growing gap between high-resource and low-resource institutions, raising concerns about unequal access to AI-driven research tools.

Policy and Research Implications

• AI Accessibility in Research: Increased funding for AI infrastructure at under-resourced institutions could help democratize access to AI-driven discovery.
• Ethical and Theoretical Implications: As AI generates hypotheses and experimental designs, it raises questions about the role of human intuition in scientific discovery.
• Future Measurement Challenges: The long-term impact of AI on fundamental science remains uncertain, requiring continuous tracking of AI’s influence on research paradigms.

Conclusion

AI has already demonstrated significant contributions to scientific discovery, enhancing productivity, reducing research timelines, and increasing the novelty of findings. However, disparities in AI adoption and accessibility must be addressed to ensure that the benefits of AI-driven innovation are equitably distributed.

[yanhong-lbh]

Cross-Domain AI Collaborations and “Surprising” Breakthroughs

Recent readings on novelty and innovation (Shi & Evans 2023; Foster et al. 2020) suggest that unexpected combinations of research fields often create high-impact breakthroughs. This memo explores how cross-domain collaboration—specifically, between artificial intelligence (AI) and traditionally “distant” fields—can generate novel solutions that might not emerge within a single domain alone. Drawing from Arthur’s (2009) view of technology evolution, I argue that AI serves as a modular component that other disciplines reuse and recombine, thereby fostering transformative progress in areas such as healthcare, education, and environmental science.

One key insight from the Nature Communications article by Shi & Evans is that researchers who operate across disciplinary boundaries are more likely to produce “surprising” results. Cross-pollination of ideas between AI and, say, ecology or social work, can yield breakthroughs that address previously intractable problems—like analyzing large-scale environmental data to predict deforestation patterns or building new intervention models for mental health. Similarly, insights from assembly theory (Sharma et al. 2023) imply that technologies evolve through a sequence of assembly steps, where each step depends on previously discovered “building blocks.” AI techniques—ranging from machine learning algorithms to deep neural networks—can act as building blocks for disciplines that historically relied on heuristics, thus unlocking novel possibilities.

However, adopting AI outside its typical tech industry habitat often demands significant upfront costs (e.g., acquiring data, training domain experts to use new methods). Drawing from Cao, Chen, and Evans’ preprint on “Modularity, Higher-Order Recombination, and New Venture Success,” cost and coordination are central obstacles to cross-domain innovation. Substantial returns on investment can be realized only if organizations prioritize interdisciplinary training and actively support forging connections across knowledge boundaries.

The chart below simulates how cross-domain collaboration, measured via a “diversity index” of co-authors’ backgrounds, correlates with subsequent “innovation scores” of AI-related projects. The visualization underscores a pattern consistent with prior findings: as diversity increases, the average innovation score rises—up to a point—suggesting that truly novel collaborations can yield disproportionately large breakthroughs. Yet as in many complex systems, there may be diminishing returns after a certain threshold, aligning with a hypothesis that not all cross-domain projects succeed equally.