From the Atomic Bomb to the Transistor: The Twilight of National Systems and Corporate Labs
How can we promote breakthroughs in science and technology? After World War II, many notable models emerged.
One is represented by the Manhattan Project and the Apollo program. The film “Oppenheimer” provides an excellent interpretation of the Manhattan Project, where almost the world’s top nuclear physicists gathered (many scientists fled Europe to come to the United States), and the United States invested enormous resources, demonstrating that a national system focused on areas at the edge of technological breakthroughs, with clear goals, can achieve tremendous success. It’s important to note that the success of the Manhattan Project was largely thanks to a unique entrepreneurial culture created by physicist Oppenheimer—young, flat, cross-disciplinary, free discussion, and mission-driven. This provided a prototype for future high-tech project management (like Silicon Valley startups).
An enterprise model famous for foundational research is the BlueSky model, with Bell Labs being the most notable. After World War II, AT&T monopolized the telecommunications industry in the U.S. and enjoyed overwhelming advantages, enabling it to secure stable and substantial monopoly profits, providing Bell Labs with long-term stable R&D funding, allowing it to invest in basic research without the pressure for short-term profits. From the perspective of corporate R&D, Bell Labs emphasized research over development. Managers explicitly demanded, “Create something remarkable for five to ten years from now.” If someone focused too much on current demands, they might be reassigned. More importantly, Bell Labs gathered experts from physics, chemistry, mathematics, and engineering in its relatively dense parks in New Jersey, along with personnel with market and manufacturing experience. This successfully balanced the openness of science with the practicality of industry within the same organization, creating foundational achievements ranging from transistors, lasers, to UNIX systems, and the C programming language. Similarly, PARC at Xerox in Silicon Valley was once widely praised. However, PARC later became a counterexample of corporate research, as it birthed epoch-making innovations like graphical interfaces and mice, but its parent company Xerox focused on “making better copiers” and failed to capitalize on these innovations, ultimately benefitting Apple and Microsoft. This reflects that corporate labs can generate significant “spillover effects” but also warns us that in rigid hierarchical organizations, top leaders might completely misunderstand the true value of innovative lab scientists.
The third model is the DARPA model. DARPA stands for the Defense Advanced Research Projects Agency, established after World War II and viewed as an efficient national funding model to promote high-risk, high-return frontier technological innovations. It has three characteristics: First, its founding purpose is to break through bureaucratic red tape, preventing scientists from spending too much energy on lengthy funding application processes. Secondly, it takes cues from venture capital (VC) practices that emerged in the U.S. during the 1950s and 1960s, emphasizing high-potential yet high-risk projects. Thirdly, it seeks to fill the gaps in the market and academia: focusing on investing in research fields that are too risky for private enterprises to invest in and whose cycles are too long for academia to support. A typical case is the early 21st century when scientists from Moderna, who would later shine during the COVID-19 pandemic, managed to obtain $10 million from DARPA for RNA research, which was not regarded favorably by academia, in just half an hour. This rapid decision-making mechanism ensured the agility of scientific research and laid a solid foundation for RNA applications in COVID vaccines twenty years later.
Ending the Monopoly with Prizes: The “Space Odyssey” and “Uncharted Terrains” of Civil Heroes
In addition to these three models, there is actually another, which DARPA loves—the public competition model, most famously illustrated by the three autonomous driving challenges held by DARPA from 2004 to 2007. Technological breakthroughs mainly solve two types of problems: one type is cutting-edge scientific problems, such as whether machines can recognize images like humans (machine learning problem), and also complex protein folding problems that have troubled humanity for a long time; the other type is promoting technological breakthroughs, such as whether private aerospace can make headway instead of being monopolized by the state, or how to realize autonomous driving? Over the past thirty years, four globally impactful competitions have taken place surrounding the aforementioned four problems, and each competition has successfully achieved breakthroughs.
1. Industrial Grand Prix
The three autonomous driving challenges held by DARPA from 2004 to 2007 marked a key starting point in the global development of autonomous driving technology. To greatly stimulate technological innovation, the prize money increased from $1 million to $2 million. The first desert challenge in 2004, though no team finished, successfully verified the feasibility of autonomous driving. Subsequently, in 2005, multiple teams successfully completed the race, establishing the core technological architecture for perception, localization, planning, and control. By 2007, the urban challenge further advanced, achieving breakthroughs in autonomous driving under complex traffic environments.
This competition not only explored early mainstream technological routes for autonomous driving but also cultivated the first generation of core talent in the field. After the competition, these researchers entered the industry, directly giving rise to leading companies in autonomous driving in the U.S. such as Waymo and Cruise. Although this competition had no direct connection to the founding of Tesla, the technological foundations and industry consensus it established profoundly influenced global development in autonomous driving, providing important support for future technological route competition, commercialization, and talent storage.
In 2004, the X Prize competition was also held in the United States. This competition required civilian teams to complete a manned suborbital flight twice within two weeks, ultimately won by the Scaled Composites team with its “SpaceShipOne”. The initiator of the X Prize, Peter Diamandis, announced in 1996 a $10 million prize, inspired by Japan’s Asahi Shimbun, which set a $25,000 reward in 1930, recruiting the first person to complete a “non-stop trans-Pacific flight.” The prize from Asahi Shimbun leveraged participants’ investments of millions of dollars, prompting Diamandis to see the leverage effect of prizes.
The X Prize not only broke the long-standing government monopoly on aerospace but also established the direction for low-cost, commercial space development and directly promoted the rise of private space companies centered on suborbital travel such as Virgin Galactic—whose predecessor was the “SpaceShipOne” that won the X Prize. The industrial environment it created also laid a foundation for the later rise of commercial rocket companies like SpaceX.
Notably, it took Virgin Galactic a full 17 years from winning the prize in 2004 to achieving its first manned commercial flight in 2021. Waymo took even longer from its inception to fully implementing driverless taxi services in the U.S. Bay Area. This indirectly reflects that after technological breakthroughs, commercial implementation still requires a lengthy accumulation and continuous technological iteration.
2. Academic Contests
In the long journey of artificial intelligence development, whether machines can recognize cats in images like humans has been one of the core problems. Before 2010, the performance of machine learning technology in image recognition was unsatisfactory, with low accuracy rates and inability to meet practical application needs, leading artificial intelligence into a developmental valley. To promote breakthroughs in machine learning technology in image recognition, Stanford University’s Fei-Fei Li team launched the ImageNet project and held the annual ImageNet challenge, which became a key event for driving the deep learning revolution in artificial intelligence.
The competition ran eight times from 2010 to 2017. In the first competition, the accuracy rate of the winning system was 72%, while the average accuracy rate of humans was 95%. This significant gap reflects the inadequacy of machine learning technology in image recognition at the time. Yet, this competition provided a unified testing platform and comparative standard for global researchers, allowing them to see their technological gaps clearly and identify their research directions.
A historic turning point occurred in 2012 when a team led by University of Toronto professor Geoffrey Hinton won with AlexNet, based on neural networks, raising the accuracy to 85% and igniting the deep learning revolution in AI, marking 2012 as the dawn of this wave of AI. The breakthrough of AlexNet completely changed the landscape of artificial intelligence development. Following this, the ImageNet challenge became a global event in the field of artificial intelligence, attracting top teams from around the world, significantly accelerating the pace of technological iteration. In 2015, the ResNet model achieved a breakthrough in the ImageNet challenge, reaching an accuracy rate of 96%, surpassing human levels for the first time; in 2017, the SENet model was crowned the champion in the last ImageNet challenge, with performance far exceeding human levels, approaching saturations in image classification tasks. Following this landmark achievement, the ImageNet challenge ceased operations, shifting focus to harder directions such as video understanding, few-shot learning, and unsupervised learning. Fei-Fei Li, as one of the founders of this competition, earned the title of “Mother of AI”, while Geoffrey Hinton was revered as the “Father of AI”, and will receive the Nobel Prize in Physics in 2024 for his deep learning contributions.
Interestingly, the victor in 2024 for the Nobel Prize in Chemistry also emerged from a competition. This is the CASP (Critical Assessment of protein Structure Prediction) competition initiated by structural biologist John Moult in 1994, regarded as the “Olympics of blind testing” in the field of protein structure predictions, aimed at pushing forward the solutions to the “protein folding problem” that has troubled biologists for fifty years. Proteins are the core carriers of life activities, and the three-dimensional structure of a protein determines its function, making the resolution of the protein folding problem critical for the development of biomedicine and genetic engineering, helping scientists better understand disease mechanisms and develop new drugs.
In the over twenty years before the CASP competition was established, while the number of participating teams continued to grow and technology iterated, the protein folding problem remained unsolved fundamentally, with prediction accuracy persistently low. Until the CASP13 competition in 2018, DeepMind’s AlphaFold1 first participated and won the championship by leading the second place by about 15% in the most difficult “free modeling (template-free)” category, marking the largest historical leap in progress and offering hope for solving the protein folding problem. In 2020, AlphaFold2 won again, achieving predicted accuracy recognized by the scholarly community as having “essentially solved the protein folding problem.” This breakthrough brought DeepMind’s founder Demis Hassabis the Nobel Prize in Chemistry in 2024. The impressive performance of AlphaFold2 made Moult lament, “We who aspire to put crystallographers out of a job are now starting to worry about being outperformed ourselves.” Moreover, the free open use of AlphaFold underscores Hassabis’s original intention as a scientist.
3. Leverage of Competitions
Looking back at these four competitions, whether it is the DARPA autonomous driving challenge and X Prize driving technological breakthroughs, or the ImageNet and CASP competitions pushing scientific frontiers, all have played significant roles in advancing scientific and technological progress, forming important paradigms in driving breakthroughs in science and technology. A thorough analysis of the operational logic and results of these competitions reveals their immense value rooted in attracting global scientists and technicians to participate through public means, forming an atmosphere of “global collaboration and exploration,” even evolving into internationally recognized academic events. Their fundamental value lies in the leverage principle of “greater good from small means”—achieving far greater resources invested through small amounts of prize money or honors, maximizing the return on investment.
Specifically, the common value of these competitions mainly illustrates three aspects.
First, the open and transparent platform attracts global participation, breaking the limits of geography and institutions, allowing wisdom and strengths distributed across the world to converge. Whether from universities, research institutions, or enterprises, all can compete and communicate on the same platform. This open model can maximize innovation vitality, making the most exceptional technologies and ideas stand out. As the events evolve, they become annual sectoral highlights, concentrating global attention on a single goal and forming a powerful synergy.
Secondly, competitions can leverage more resources, achieving a leverage effect of “greater good from small means.” Be it the prize amounts of $100,000 to $200,000 from the DARPA autonomous driving challenges, or the $10 million prize from the X Prize, comparatively speaking, these are minor figures relative to the R&D investments of participating teams and the subsequent scale of industrial development. For instance, the X Prize leveraged over $100 million in R&D investments from participating teams, while the entire commercial aerospace industry today has reached a scale of $469 billion; although ImageNet set no prizes, its academic influence and industrial appeal mobilized massive investments from global tech giants, rapidly advancing deep learning technologies’ implementation and sparking a funding frenzy in the AI sector.
Thirdly, whether in scientific or technology breakthroughs, competitions can serve as a source for entrepreneurship, gathering core talents related to respective fields to establish a complete industrial ecosystem. The DARPA autonomous driving competition nurtured the first generation of core talent in autonomous driving, fostering top companies such as Waymo and Cruise; the X Prize spurred Virgin Galactic, setting the stage for the rise of SpaceX, a private space giant; ImageNet prompted the transition of AI talent from academia to industry, igniting a vigorous entrepreneurial wave in AI, causing the salaries of AI professionals to break through ceilings; as for CASP, it provided the best opportunity for DeepMind, a startup acquired by Google but still operating independently, to validate its value—both to humanity and the wealth of innovation itself. Competitions essentially catalyze scientists into scientist-entrepreneurs and embody a decentralized organizational model that encourages cross-sector exploration while pursuing clear objectives, forming an industrial ecosystem and driving continuous iteration of technology and science.
Interestingly, the most effective measure of these competitions is whether they effectively “outdo” themselves. The ImageNet competition and CASP both concluded due to “successfully” meeting their goals, whereas the competitions in autonomous driving and private space ensured the completion of industrial prototype verifications, reassuring investors while passing on the later development to entrepreneurs, despite the daunting journey from prototype to industry realization.
These competitions that once attracted global attention over the past thirty years also prove that scientific breakthroughs are never the result of lone efforts but require open platforms, collaborative powers, and efficient resource empowerment. Competitions serve as the vital bridge connecting these powers. They embody the spirit of open science and public exploration, reflecting a philosophy of crowdsourcing. Competitions also create beneficial communication platforms, showcasing not the veiled works leading up to end results but the open records of the winding exploration processes and mistakes while inviting peers to participate, critique, and seek breakthroughs through collision. This represents the scientific spirit of the era.
Looking forward, we expect more competitions similar to these to emerge, focusing on scientific frontiers and technological bottlenecks, mobilizing more resources and talents while driving human technology forward and unlocking more unknown potentials.