The Next Generation AI Pathways
In the second half of this decade, innovative AI capabilities will come from a constellation of closed proprietary[1] and open-source models. The release of ChatGPT, a proprietary generative AI model,[2] in November 2022 was for many their first encounter with an actual AI system. On account of its relatively simple user interface and its general-purpose capabilities, it quickly amassed more than 100 million users and solidified AI’s place as a revolutionary technology.[3] Since then, there has been a surge of AI model development and adoption, as well as applications built on top of the foundation models to advance manufacturing, biotechnology, education, and defense.[4] Leading these efforts are well-resourced companies and countries that leverage massive datasets, vast computing power, and cutting-edge research to cultivate proprietary AI models with exponentially greater capabilities.
Meanwhile, the rise of open-source AI models has enabled the diffusion of multi-purpose AI capabilities among a wide variety of actors. Within open-source models, there is a spectrum of “openness”: for example, individual aspects or some combination of code, data, architecture, and licensing arrangements can be open-source.[5] In certain historical contexts, open-source technology has been a democratizing force, expanding access to knowledge and tools. As a driver of startups and academic research, it is a competitive advantage of the U.S. innovation ecosystem. Open-source AI models thrive on community-driven innovation, fostering collaboration, transparency, and accessibility. Open-source model capabilities, however, can be expected to continue to lag behind proprietary frontier models, primarily due to economic and technical constraints on scaling.
For all their upsides, powerful open-source AI models also entail risks. One of the downsides is giving up the ability to update, constrain, monitor, and withdraw the model if safety concerns emerge following its deployment. Once a model’s weights are released, it is relatively easy to modify, fine-tune to bypass safeguards, or combine with other models, making it impossible to maintain control over the model or its downstream use.[6] The decentralized implementation and development spurred by open-source models could open the door to misuse by adversaries and other malicious actors.
In the remaining years of this decade, we can expect to see rapid advancements in both proprietary and open-source AI models. United States policymakers will need to continue to monitor and govern proprietary AI advancements. At the same time, we need to find ways to effectively govern open-source AI without impairing our own ecosystem’s ability to leverage it. Over-regulating open-source initiatives can stifle domestic innovation and leave the United States at a disadvantage compared to our geopolitical competitors who may not be bound by similar regulatory constraints.
Moreover, ensuring the United States fosters a technology ecosystem with the right balance of proprietary frontier and open-source AI models will have geopolitical implications. Only a very small number of countries and companies possess the resources to develop frontier foundational AI systems. A few will be able to marshal the resources to create these AI systems to solve cutting-edge challenges, but others may have to rely on moderately advanced open models at a fraction of the cost with greater flexibility. Most actors on the global stage will face a tradeoff between power, when using someone else’s closed model, and freedom, when building on top of relatively less capable open-source models. The United States’ AI policy will have to account for both of these trajectories.
The Road to AGI
Based on advances across the AI stack, including improvements to hardware and algorithms, we can expect future generations of AI models to progress toward more generalized and powerful capabilities. Market demand and the direction of current research are pointing toward the advent of agentic AI that can take actions in pursuit of complex, human-directed goals. There may not be a single development that results in Artificial General Intelligence (more on this below), but cumulative progress will likely result in agentic AI that will have certain characteristics: namely, some combination of goal-directedness, longer-term memory, and tool use/ability to take actions and create and carry out plans. AI models already have multimodal capabilities that can learn from and process different types of data — text, image, audio, video — allowing for more than language inputs and outputs.
Over the next few years, we will also likely see AI tools that can process more and more information at once, allowing them to have short-term memory across massive amounts of input,[7] even more so than current LLMs. This capability allows an AI to learn at the time of prompting (“in-context”) from input tokens, recall significantly more information, and reason across data in a prompt.[8] Longer context windows enable users to create specialized models — by inputting large amounts of task-specific data at the time of query — without needing to retrain the model.[9] They also allow AI models to have a sort of memory that enables AIs to create and follow plans in the pursuit of longer-term goals. AI can already use digital tools like calculators, web browsers, coding environments, and digital marketplaces. These tools — which are quickly expanding in scope and sophistication — allow AI to autonomously interact with, impact, and learn from the real world. In time, these task-specific AI agents might be able to interact with one another, but more research is needed on the potential of such possibilities.[10]
Humans will be able to leverage these AI agents to take action using natural language, including to find solutions to pressing problems. When this is available to humans at large, the world will be forever transformed. Humans will be able to employ AI agents that can, for example, write code to take complex actions in the real world.[11] Moreover, the combination of scale, multimodality, and real-world reinforcement learning could lead to one or more centralized model(s) with general capabilities in virtually all tasks, based on knowledge that exceeds all of humanity — a threshold which some consider to indicate Artificial General Intelligence (AGI).[12] The development of AGI could offer never-before-seen benefits, such as the ability to solve critical scientific hurdles by expanding the pool of cognitive labor, and to make humanity ever-wiser by providing us with an improved level of information gathering, communication, and education.
In the evolving discourse on AGI, the United States must ensure that our preparation for and response to AGI adheres to our values, laws, and ethical frameworks. At the same time, recognizing that our main geopolitical competitors aim to reach the “commanding heights”[13] of technological innovation before the United States, it is imperative that the AGI narrative does not predominantly focus on the risks. We must not lose sight of the urgency to bolster our competitiveness, increase the prosperity of our people, and propel our country into the future. If some form of AGI is approaching, the arc of history will bend toward the nations that have understood its potential and have taken appropriate action to prepare for it. The United States has historically marshaled the collective resources of academia, government, and the private sector when the moment has demanded it — such as our mobilizations during the Manhattan Project, the Space Race, the development of the Internet, and the rapid rollout of COVID-19 vaccines. Indeed, our nation’s most ambitious technological achievements have resulted from united national efforts in technological advancement. To maintain global leadership and secure our interests and those of our allies, the United States must proactively shape, develop, and deploy AGI through a comprehensive strategy that integrates the strengths of our entire nation. This approach will not only ensure that we stay at the forefront of technological innovation, but also enable us to shape the governance frameworks that will guide the global deployment of AGI, ensuring its alignment with democratic values and international norms.
AI’s Convergence with General Purpose Technologies
This decade is marked by both the advancement of multiple general-purpose technologies and their convergence. A new wave of general-purpose technologies — AI, advanced networks, compute, biotechnology, next-generation energy, and advanced manufacturing — are emerging. Like the general-purpose technologies that came before them — from steam power and the telegraph to aviation, mass production, and the Internet — they have a wide range of applications across industries and hold the potential to unlock outsized economic and geopolitical benefits.
AI’s acceleration of these general-purpose technologies will further distinguish this technological era by transforming the very process of innovation.[14] While the past several decades of technological change took place primarily in the digital domain, AI and other emerging technologies are unfolding across the physical, digital, and biotechnical domains.
The convergence of these technologies will create new opportunities and unexpected problems that will need to be solved. For example, the convergence between the physical and digital domains allows for the creation of intelligent factories that produce goods in a much more flexible and efficient way. These digitized facilities, however, can also be expected to be much more vulnerable to cyberattacks. Programmable biotechnology will enable personalized healthcare, biomanufactured infrastructure, and more. At the same time, it also risks lowering the barriers to entry for the creation of bioweapons.
Global technology leadership will accrue to the nation(s) best able to harness technological convergence across multiple domains. The leadership of the United States in AI will be necessary but insufficient to unlock the innovation power required to lead in this era. Five other technology battlegrounds — and their convergence with AI — will help shape the geopolitical terms and the destiny of nations in this decade and the next:
- Biotechnology. Biotechnology promises the unique opportunity to grow and manipulate the essence of life as we know it. Applying AI to nature’s source code — DNA — will make the research and development of drugs, foods, and the fermentation of industrial chemicals faster, cheaper, and more accessible. Yet high-quality data will be an essential input for AI-enabled biotechnology, and a limiting factor for the ability to engineer biology. The United States is entering this race with an early lead in terms of innovation, investment, and talent, but public-private partnerships will be necessary to outcompete a determined PRC and secure a “biofuture” that neither the U.S. Government nor industry could achieve alone.[15]
- Advanced Networks. Harnessing the value from AI in real-world situations hinges on the ability to rapidly and reliably transmit data between machines with a latency measured in nanoseconds. Emerging advanced networking standards like 5G advanced, WiFi 7, and 6G will unlock long-anticipated applications like autonomous vehicles, remote human-machine teaming for healthcare, and software-defined intelligent factories. The PRC won the race to deploy commercial 5G networks globally. But as nations and firms compete to shape standards and deploy next-generation networks that underpin cyber-physical systems, the outcome has yet to be decided.[16]
- Advanced Compute & Microelectronics. Continued compute scaling has underpinned rapid progress in AI over the past decade, but Moore’s Law — the prediction that available compute power would double every two years — faces an uncertain future. Compute and energy demand from AI scaling continues to far outstrip the gains from Moore’s Law, creating a bottleneck that threatens AI progress. Scaling breakthroughs in novel computing architectures and post-Moore’s Law microelectronics — such as in-memory computing, reversible computing, and superconductor electronics — would open new possibilities.[17]
- Next-Generation Energy. The global energy sector is undergoing a massive transformation. Clean energy technology is now central to the global techno-economic competition, as nations pursue new ways to power their technological advancement while energy innovations converge with AI, compute, transportation, manufacturing, and other strategic sectors. The United States must catalyze disruptive innovation in technologies like energy fusion, space-based solar power, and long-duration energy storage in combination with policy measures to create new national security, economic, and diplomatic advantages.[18]
- Advanced Manufacturing. A core set of emerging technologies, from AI to additive manufacturing and robotics, are converging to transform how things are made. These technologies harness the United States’ advantage in AI and software to create production systems that are faster, cheaper, and more sustainable. Accelerating the deployment of advanced manufacturing systems could chip away at China’s manufacturing dominance and bolster the United States’ capacity to restore its industrial base.[19]
Progress across each of these general-purpose technologies either builds on or enables transformative change in AI.[20] As states seek to capture the strategic and economic benefits of general-purpose technologies, emerging sectors have become battlegrounds where strategic competition plays out. Commercial competition notwithstanding, the race to shape technological convergence ultimately breaks down across geopolitical and ideological lines. The outcome of this competition will determine whether these technologies are shaped in accordance with democratic or authoritarian values.
[1] “Proprietary AI refers to artificial intelligence technologies developed and owned by specific companies, often made available to customers through licenses or subscription services.” See Open Source vs. Proprietary AI: A Comparative Analysis, Medium (2023).
[2] “Generative AI is a category of algorithms that finds patterns in training datasets and extrapolates from them to generate content such as text, images, or audio, given natural language or multimedia input.” See Generative AI: The Future of Innovation Power, Special Competitive Studies Project at 31 (2023).
[3] Krystal Hu, ChatGPT Sets Record for Fastest-Growing User Base – Analyst Note, Reuters (2023).
[4] The State of AI in 2023: Generative AI’s Breakout Year, McKinsey (2023); Lauren Coffey, Harvard Taps AI to Help Teach Computer Science Course, Inside Higher Ed (2023); Sydney J. Freedberg Jr., Pentagon Tested Generative AI to Draft Supply Plans in Latest GIDE 9 Wargame, Breaking Defense (2024).
[5] On one end of the spectrum, fully open-sourced models may release their weights and code, thereby giving insight into the inner workings of the model; on the other end, models might be accessible only through an API with no access to the underlying model. In-between are various levels of partial openness, including models with published architecture and training code, but unpublished weights. See Zoë Brammer, How Does Access Impact Risk? Assessing AI Foundation Model Risk Along a Gradient of Access, Institute for Security and Technology (2023).
[6] Note there is ongoing research into watermarking model weights to help trace where open-source software originally came from (tracking the system, rather than the outputs) and projects focused on preventing model weights from being further tuned. See Simon Lermen, et al., LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B, arXiv (2023).
[7] Context windows have been getting longer, and this work could lead to continuous context windows in which AI has significant short term memory; however, this area remains a novel challenge. See Machel Reid, et al., Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens of Context, Google DeepMind (2024).
[8] For example, the entire grammar manual for a specific language and some examples of sentences in that language can be fed into a context window, and the AI model can learn to translate text from English into that language at a level similar to a person learning from the same content. See Chaim Gartenberg, What is a Long Context Window?, Google The Keyword (2024).
[9] Fine-tuning (further training of the model on a possible set of tasks) and prompting (particularly with longer context windows) are two independent ways to make a specialized model. Developers of closed AI models can fine-tune their own models to create specialized models via further training. However, for closed-source models which are not publicly available to fine-tune, a longer context window provides a mechanism for third parties to fine-tune the model in effect. Open-source models can both be fine-tuned and employ long context windows.
[10] When agents can talk to each other and the authors of their respective systems are different, the ensuing consequences could be catastrophic, in cases of unstable interaction dynamics. We expect that terms of service and licensing agreements, at least initially, will prohibit this interaction. However, well-constructed agent-agent interactions could be beneficial for problem-solving and quick iteration.
[11] Progress toward AI that can execute action is supported by work integrating foundation models with robotics to enhance a robot’s ability to understand and respond to complex commands and environments. See for example, Figure Status Update, Figure AI (2024); NVIDIA Project GR00T, NVIDIA (2024).
[12] Sam Altman, Planning for AGI and Beyond, OpenAI (2023).
[13] Full Translation: China’s ‘New Generation Artificial Intelligence Development Plan’ (2017), DigiChina (2017).
[14] Generative AI: The Future of Innovation Power, Special Competitive Studies Project at 61 (2023).
[15] National Action Plan for U.S. Leadership in Biotechnology, Special Competitive Studies Project (2023).
[16] National Action Plan for U.S. Leadership in Advanced Networks, Special Competitive Studies Project (2023).
[17] National Action Plan for U.S. Leadership in Advanced Compute & Microelectronics, Special Competitive Studies Project (2023)
[18] See National Action Plan for U.S. Leadership in Next-Generation Energy, Special Competitive Studies Project (2024).
[19] National Action Plan for U.S. Leadership in Advanced Manufacturing, Special Competitive Studies Project (forthcoming).
[20] For example, scientists are building large-scale models trained to understand the language of biology. Generative AI is being used to design AI chips used to power future compute advancements. In energy, scientists are leveraging generative AI for projects that range from designing novel high-performance battery materials to stabilizing fusion reactors — all in a race against time to usher in a new paradigm of abundant, low-cost energy that will, among other things, support future AI scaling. The convergence of novel and existing forms of AI with other general-purpose technologies has contributed to the development of advanced manufacturing technologies.