人工智能系统建模创新

Executive Summary Recent advances in artificial intelligence (AI), particularly around large languagemodels (LLMs), have made OpenAI’s Generative Pre-trained Transformers (GPTs),Google’s Gemini, Anthropic’s Claude, and Meta’s Llama household names. Thesequential releases of foundation models from a small number of big players tell alinear story of innovation, but history suggests a more complex narrative. Just aselectricity in the Second Industrial Revolution would eventually spur more reliable anddistributed power systems, foundation models are paving the way for what someresearchers have termed “compound AI systems.” These systems encompass a set ofdistinct components, including at least one AI model and other components thatmanipulate the data in ways not learned by the model.*The components mayaggregate multiple calls to a model, insertcontext with retrievers, or extendcapabilities with tool use. Compound AI systems improve performance on tasks fromthe base model and can enable the model to perform entirely new tasks. Consider an autonomous vehicle. The vehicle itself is a complex system of systems,including power, propulsion, sensor systems, and more. The vehicle is notsubstitutable for a single AI model, but subsystems involve model and compoundsystem design decisions. For example, some manufacturers do not include LiDAR(Light Detection and Ranging)in their self-driving vehicles, opting solely for cameras,while others use a sensor suite, including cameras, LiDAR, and radar. Processing thesesignals for a task, such as object detection, could take at least two approaches:Trainseparate networks for each sensor or sensor type and combine the predictions (acompound AI system approach) or train a single model that fuses all sensor inputs (anAI model approach). At the level of object detection, the model and compound systemapproaches can be direct substitutes. This interchangeability allows lessons from oneapproach to spur improvements in the other. New model advances could enablecompound systems to take advantageof that capability; a compound system approachcould add a new sensor network, which might inspire a more efficient single networkdata fusion model that shares representations between sensors for faster objectdetection. System-to-modelinnovation is an emerging innovation pathway that has drivenprogress in several prominent areas over the last decade.System-to-modeladvancesinclude DeepMind’s incorporation of policy and value calculations in a single networkwith AlphaGo Zero;chain-of-thought prompting leading to OpenAI’s o1 model; theOneGen single pass network and Cohere’s Command R family models in retrieval- augmented generation (RAG); and circuit breakers for AI safety training in languagemodels.System-to-modelinnovations close the feedback loop between model andcompound system advances, opening frontier AI breakthroughs to more diversecontributions beyond the top labs. Instead of onemodel-to-modelpathway for progress, recent trends highlight thedynamic interplay between AI model and compound system innovation, whereprogress along one pathway leads to breakthroughs in the other pathway. Policies that foster AI progress are likely to benefit model and compound systemdevelopers, but today’s debates have focused more on improving model capabilities. Ifrecentsystem-to-modeladvances are indicative of future trends, then policymakersshould also devise tailored solutions for compound system developers to encouragemultiple bets on the future of AI. System-level innovations advance with the diffusionof AI and expand the baseof contributors to leading-edge progress in the field. Whilethe spread of general-purpose technologies across societies and economies is oftenmore consequential than the innovations themselves, the diffusion of AI will nothappen in a geopolitical vacuum.1Countries that can identify and harness system-level Center for Security and Emerging Technology |2 innovations faster and more comprehensively will gain crucial economic and militaryadvantages over competitors. This issue brief suggests a three-part framework tonavigate the policy implications ofsystem-to-modelinnovation: Protectsmaller companies as incubators of tacit knowledge that could seed futurefoundation model advances. Governments may need to consider subsidized services,such as safety testing and red team assessments at a system level, for smaller andmedium-sized enterprises that lack the resources and expertise to remain compliantwith the evolving policy landscape or withstand adversarial attacks. This processshould include a detailed mapping of the companies, infrastructure, researchorganizations, data analytics firms, and technical and nontechnical talent base thatcomprise U.S. and allied compound AI system innovation networks to determine theactors involved and resources needed for protection. Diffuseinnovation outside of leading foundation model providers by providing modelsand inference compute access to develop