CHAPTER 8 Scientific Progress in Artificial Intelligence: History, Status, and Futures

Birth and Evolution of Scientific Field

The prospect of automating aspects of human thinking via mechanical systems has been considered for hundreds of years. Modern metaphors and framing of thinking as a computational process have roots in the early twentieth century. Key contributions to the perspective of thinking as computing include the theoretical work of Alan Turing on computability,¹ efforts by John von Neuman, Turing, and others to construct general-purpose computing systems,² and work on computational abstractions of neuronal systems by McCollough and Pitts.³ The 1940s saw the rise of discussions and publications viewing the computer as a metaphor for the brain, including control-theoretic notions referred to as cybernetics.⁴

The modern discipline of AI, per the establishment of a long-standing set of aspirations, harkens back to a research project proposal for a summer workshop held at Dartmouth College in 1956.⁵ The proposal, coauthored by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, outlined a new field of studying how machines could be programmed to perform “every aspect of learning or any other feature of intelligence.” Containing the first use of the phrase artificial intelligence, the proposal described goals of finding “how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.” The summer study is considered as the formal launch of AI as a distinct field of scientific inquiry, setting the foundation for decades of research in computer science.

The maturation of the AI research program saw the evolution of a set of AI subdisciplines with overlapping but distinct research communities, including natural language understanding, problem-solving, planning, vision, robotics, and machine learning. Research areas and communities also formed around distinct foundational approaches to building AI, such as logical reasoning and representations, reasoning under uncertainty with statistical methods, and the use of neural network models versus high-level symbols—a domain of research that had been referred to for decades as connectionist approaches. Further, advances and questions in AI have stimulated efforts in other disciplines, such as cognitive psychology, where cognitive science refers to a subdiscipline of both AI and cognitive psychology that centers on taking inspiration from studies, data, and questions about human cognition to build systems that can perform automated learning and reasoning, and on using computational approaches to modeling and probing human psychological processes.⁶

Representations and Reasoning Mechanisms

Scientific studies of AI are best understood in terms of the technical evolution of different approaches to representing and reasoning with data and knowledge. In the early days of the field, representations and reasoning methods included the use of neural networks, early-on referred to as perceptrons in work on learning to recognize visual patterns,⁷ and symbolic logic applied in both specific instances and in attempts to build general architectures for problem-solving.⁸ Symbolic representations dominated the first several decades of AI research with efforts in statistical methods, including neural networks, continuing but largely taking a backstage position. Work in logic-based systems included rule-based expert systems that came to focus of attention in the 1970s and 1980s. These systems were aimed at capturing specialist knowledge in sets of compact logical rules (e.g., if-then rules) that would be used to compose chains of inferences within an architecture referred to as a production system.⁹

In a paradigm shift in the mid-1980s, attention began to shift from logic-based methods to statistical approaches for handling uncertainties associated with the complexity of real-world problems, such as applications in medical diagnosis and decision support. Representation and reasoning machinery were developed for harnessing probability theory and decision theory,¹⁰ including Bayesian networks¹¹ and, more generally, probabilistic graphical models.¹² Systems were developed using these probabilistic representations for making inferences, such as inferring medical diagnoses from information about a patient’s illness, sets of symptoms, and lab results. In some systems, the collection of additional information to help refine conclusions or diagnoses was guided by computing the expected value of information of additional observations, tests, or data.¹³ In addition, AI research scientists began to incorporate and extend techniques developed in the related disciplines of Operations Research, such as Markov decision processes to support sequential decisions.¹⁴

Despite the rise and fall of excitement in different methods, efforts have continued within and across multiple fundamental representation and reasoning methods. For example, today’s successes and focus of attention on large-scale neural networks extends in a recognizable line from the nascent work in the early 1960s on perceptrons to the most recent developments with methods and systems based on neural networks. Today, studies of symbolic reasoning methods continue, including on mechanisms for integrating symbolic reasoning with neural models to bolster their abilities to perform logic and more general mathematics.¹⁵

Discriminative and Generative Models

Machine learning methods can be broadly divided into two main capabilities, discriminative AI and generative models, each with distinct objectives and application categories. Discriminative models take as input the description of some item and outputs a label, or classification, of the item. For example, in the case of a junk email filter, a discriminative model learns to label each input email as either spam or non-spam by analyzing features derived from the email. These discriminative models directly use the features of the input data to make predictions or classifications, focusing on the relationship between the input data and its corresponding labels.

Discriminative models span classic statistical models of logistic regression, algorithms for learning classifiers from tabular data, and deep learning for diagnosis and classification. Examples of discriminative models include leveraging labeled data drawn from electronic health record systems to predict readmission,²² sepsis,²³ and the onset of infection²⁴ in hospitalized patients.

Generative models have been front and center in the recent excitement about AI and its applications. Such models replicate the process by which data is generated. By learning the probability distribution of output features given input features, generative models can create and output new data instances that resemble the training data (in contrast to the labels output by discriminative models). Multiple methods have been used in generative AI, including techniques named generative adversarial networks (GANs), variational autoencoders (VAEs), diffusion modeling, and more recently, transformers that yield exciting capabilities of generative AI models. Generative models trained on images are now being used to generate novel imagery, as has become popularized in the DALL-E and Midjourney applications. Beyond images, generative methods are being used in a wide range of applications, including the structure and design of protein sequences and the performing of scientific simulations.

Supervised, Unsupervised, and Self-Supervised Learning

The training procedures by which models are constructed in machine learning can be broadly categorized into supervised, unsupervised, and self-supervised learning. Supervised learning relies on labeled datasets. The use of such curated data has been the basis of significant advancements in areas like medical diagnosis, image analysis, and speech recognition. Unsupervised learning refers to methods that find patterns in data without explicit labeling. Traditional variants of these methods include clustering and anomaly detection, which have been particularly useful in exploratory data analysis.

Over the last decade, a special form of unsupervised learning,²⁵ named self-supervised learning, has become very important. Self-supervision is a simple yet powerful idea that has enabled AI systems to learn from vast unlabeled datasets, such as massive corpora, crawled from across the web. One approach to self-supervision is to generate labels automatically by a “fill in the blanks” process of hiding words in text or other types of tokens in datasets and then trying to predict the hidden information. As an example, a model might predict the next word in a sentence or the next frame in a video sequence based on previous words or frames.

Self-supervised learning represents a significant shift in machine learning, moving away from heavy reliance on human-labeled data. This paradigm is unlocking new possibilities across various fields, enabling models to learn from vast untapped datasets and driving innovation in areas where labeled data is scarce or expensive to obtain.

Key Concepts and Research Directions

Several key directions have come to the fore as important developments, requirements, and directions in work on DNNs, including interest in automated ways to learn rich representations rather than curate them with expert guidance; efforts on robustness; the weaving together of multimodal datasets; and the critical value of hardware, innovative algorithms, and programming platforms for research and development.

Learning representations. Early supervised machine learning required the identification of handcrafted salient observations or features of the input and based its predictions on those handcrafted features. Researchers have explored how deep learning can identify such features automatically or, more generally, rich representations directly from fine-grained data, in processes referred to as representation learning. The automated learning of rich features to represent an image, starting with the lowest-level input pixel features and becoming progressively more complex and abstract at successive layers of neural networks has been a celebrated aspect of modern deep neural models for vision.

The ability to automatically learn, or discover, candidate features enables systems to discover how best to organize the structure of machine learning problems, often yielding more accurate and robust performance on complex tasks than human-defined attributes, albeit at the cost of increased data requirements. Neural models leveraging such representation learning have been developed for natural language processing, computer vision, speech recognition, and health care.

Robustness and generalization. Efforts in the field of DNNs have increasingly focused on achieving robustness and generalization to ensure accurate performance in varying real-world environments that can be assumed to require robust capabilities, for example, accurate diagnoses, classifications, and predictions on new, previously unseen data, that is not contained in training datasets. Efforts in this realm push DNN training procedures to seek universal patterns from their training data so as to reduce their being overfit to training data and to be more adaptable to diverse real-world scenarios.

Hybrid strategies. Successes have been found with combining DNNs with other computational methods such as coupling the neural models with scientific simulations, integrating the methods with Markov decision processes (e.g., reinforcement learning), and integrating DNN approaches with symbolic approaches to reasoning. As an example, the AlphaGo systems rely on an integration of deep neural models for making predictions with reinforcement learning for guiding the choice of actions.³²

Multimodal modals. Most DNN efforts have focused on the singular modalities of language or visual tasks. In the spirit of pursuing more human-like intelligence, researchers have pursued the development of multimodal models that bring together language, imagery, sounds, and other modalities. Multimodal DNNs include early efforts to do image captioning and more recent efforts to make inferences about language and images for such tasks as writing radiological reports.

Tools, methods, and platforms. With the advent of the importance of DNNs and growing focus of attention on using increasingly large datasets, methods have been pursued for introducing new forms of efficiency via hardware and algorithmic innovation, and for developing programming environments doing exploratory work with architectural designs for neural networks. At the hardware level, graphical processing units (GPUs) have provided speed-ups via parallel processing of matrix and vector operations that are central to deep learning.

Algorithmically, efforts span methods for introducing new forms of speed-ups in distributed computing at the hardware system level as well as on higher-level software innovations aimed at speeding-up the core back-propagation procedure to identify parameters that specify weights of connections in neural models. For example, efforts have focused on adaptation of mathematical optimization procedures like stochastic gradient descent.

Programming environments such as TensorFlow and PyTorch were created to ease the design and testing of DNNs, providing engineers with computing libraries, methods for accelerating GPU computation, and efficiently specifying and revising the structure of neural networks.

Models as platforms. For decades in machine learning, researchers have studied methods for adapting models trained on a source task to perform well on other domains via processes of fine-tuning the models with specialized data. This process, often referred to as transfer learning, leverages the knowledge that the model has gained from the initial training to perform well on a related, but different, task.³³ Large-scale neural models can serve as platforms for extending via fine-tuning with specialized datasets drawn from target task domains. Given the myriad uses of the large models as platforms that can be extended via domain-specific data, they have been referred to as foundation models.³⁴ Foundation models can be seen as an extension and scaling-up of transfer learning to DNNs that are trained on extremely large datasets, often encompassing a wide range of topics, languages, or modalities. Their versatility lies in the ability of pretrained models to be fine-tuned with smaller, task-specific datasets, thereby reducing the need for training a model from scratch for each new application. This approach not only saves significant computational resources but also allows for building upon the model’s base capabilities and knowledge. The term foundation reflects their role as a fundamental base upon which more specialized or fine-tuned models can be built, similar to how a foundation supports a structure. Their general-purpose nature and scalability make them akin to a utility or resource that can be tapped into for numerous AI systems. Fine-tuning pretrained foundation models has become a standard methodology for adding new capabilities, such as adding multimodal capabilities to language-only models³⁵ and for extending the power of generalist models to specialist performance.³⁶

Architectural Innovation: “Attention Is All You Need”

A seminal paper introduced the Transformer architecture,³⁷ the foundation of today’s generative AI. This particular design of DNN delivers surprising competencies via a mechanism called attention, which allows neural language models to learn to focus on different parts of an input sequence when generating each part of the output. In short, transformers learn during self-supervised training how to weight the importance of different parts of the input data. Such a broad ability to learn where to look and what to consider has been seen as a pivotal feature for understanding the context and nuances in language, in distinction to earlier approaches for learning about sequences, of only looking adjacently for the context of generation. The power of transformers in various applications, including language translation, text generation, and image processing tasks has led to their broad adoption.

The second pivotal development was combining the Transformer architecture with self-supervised training from a diverse, web-scale dataset. This approach was first demonstrated with the construction of the BERT foundation model.³⁸ BERT learned language by predicting parts of text that were hidden from it, gaining a broader and more contextual understanding of language via broader learning about where to attend. These innovations laid the groundwork for the development of follow-on Transformer models like the GPT series, LLAMA, and others, each building upon and extending the transformative capabilities introduced by their predecessors.

Alignment with Human Intent and Interaction

A third innovation for enabling modern generative AI is a mechanism for shaping models to follow natural language instructions and to sustain a conversation, versus simply generating tokens that are most likely to follow the input prompt. This process of learning to respond to the intentions of people involves fine-tuning the model on a new dataset composed of various tasks, each linked with explicit instructions and rating the output. The instructions are designed to mimic the way humans would typically instruct each other to perform tasks. The dataset is typically initially generated or refined by human annotators who craft the instructions and provide example outputs or correct responses. To scale instruction tuning, a method referred to as reinforcement learning from human feedback (RLHF) is used to expand the instruction dataset and provide measures of the quality of generated outputs. This method involves training and then using an automated approach to scale up the shaping of the model’s behavior to ensure a wide coverage of task types and linguistic variations.

Scaling Laws and Emergent Capabilities

A remarkable property of large language models based on the transformer architecture is the existence of a strong empirical relationship between the accuracy of the trained language model and the size of the model (the number of parameters optimized during training), the amount of data on which it is trained, and the amount of computation used during training. This relationship, known as “scaling laws,” has been empirically validated multiple times.³⁹ These scaling laws are important because they predict how larger models trained on larger datasets using greater computing resources yield increased accuracy; if they continue to hold as models are further scaled up, then one can expect even greater accuracy. Figure 8.3 displays a measure of the ability of a learned model to predict next tokens (the “Test Loss”), given a sequence of words at focus of attention, as a function of increases in powers of ten in the compute time, training data, and number of parameters of models.

Scaling laws have provided a reliable framework up to now for predicting basic performance metrics, such as error rates in next word prediction. However, they fall short in anticipating the competencies of models on challenging tasks, including benchmarks in natural language and problem-solving. Training large-scale neural models from broad datasets can be viewed as a form of multitask learning with new tasks being learned with increasing amounts of computation for training and with the size or capacity of models. Task-centric jumps, which have been referred to as the emergence of new capabilities, have been observed in neural language models on diverse tasks at different thresholds of model parameters, compute power, and training corpus size. Emergent behaviors include the relatively rapid increase in performance on benchmarks after reaching particular threshold levels of investments in computation for training, as captured in Figure 8.4.⁴⁰ Emergent capabilities include jumps in performance on nuanced language understanding benchmarks and with the acquisition of higher-level abilities, such as “theory of mind”—the ability of AI systems to solve challenges with interpreting and predicting the intentions, desires, and beliefs of people.⁴¹ To date, we have a poor understanding of the basis for such jumps in capabilities as a function of model size, extent computation, and training data, and links between the accuracy of next word prediction and performance on the more sophisticated benchmarks.

To date, the exact mechanisms and thresholds that trigger emergent capabilities remain largely unpredictable and are an important research direction. This unpredictability underscores a significant frontier in AI research, where the confluence of parameters, computing resources, and training data size creates a complex landscape, within which unexpected and sophisticated AI capabilities can spontaneously manifest.

Surprising Powers of Abstraction, Generalization, and Composition

The original set of ChatGPT systems, using GPT-3.5, GPT-4, and related models have surprised the world with their generalist powers to perform abstraction, generalization, and numerous forms of composition. The models also show broad “polymathic” capabilities, demonstrating the ability to weave together concepts and content drawn from multiple disciplines. The scientific community does not yet have a good understanding of the emergence of their abilities to perform various kinds of summarization, text generation, problem-solving, code program generation, and conversational dialogue. Multiple projects are underway to probe the powers and failings of these models.

Since the release of GPT-4, and related large-scale models such as Claude and Gemini, numerous studies and associated papers have probed potential uses and have provided an array of evaluations. An early survey of capabilities was undertaken by Bubeck and colleagues,⁴² spanning a broad set of computing problems, specialist challenges, and the handling of needs and interpretation of events of daily life. The survey highlighted surprising capabilities as well as weaknesses and future directions. Weaknesses include the tendency of the large models to confabulate with the creation of erroneous but persuasive generations and solutions and failures to perform basic arithmetic operations. Studies have also uncovered potentially fundamental challenges with limited abilities to solve complex planning problems that have relied on traditional AI problem-solving on searching through options with backtracking.⁴³ These challenges have been attributed to the sequential generative processes of current models. Exploratory efforts have pursued insights about the root cause of failures, such as weaknesses models can exhibit with accurately solving constraint satisfaction and mathematics problems.⁴⁴

Tapping Specialist Performance via Steering

For years, specialist performance with large language models has been achieved via training with domain-specific datasets, such as with the construction of BioBert⁴⁵ and PubMedBert⁴⁶ or fine-tuning foundation models with domain-specific data to update the parameters of the general models via optimization. In addition to surprising powers of abstraction, generalization, and composition, recent studies have demonstrated that generalist foundation models can be guided through special prompting strategies to perform as top specialists. For example, prompting methods can guide GPT-4 to act as a top medical specialist, with record performance on the MedQA benchmark of medical challenge problems.⁴⁷ Innovation with prompting shows that generalist models can be steered to perform as experts on competency exams in other areas, including electrical engineering, machine learning, philosophy, accounting, nursing, and psychology.

Research Directions on Generative AI

A great many questions have been framed by the successes and failures of generative AI models. The current questions and curiosity frame a set of research directions and underscore the critical importance of furthering the scientific study of the methods and models.

Representation and reasoning. There is evidence that pushing Transformers via intensive optimization to become increasingly better at predicting the next tokens in their generations, under bounded computing and representational resources, induces the models to induce rich world models as an ideal form of compression. Although several directions have provided insights about the construction of world representations, much remains unknown,⁴⁸ and this is an open and interesting area of research.

In a related direction of research focused more on the microstructure of internal activity within transformers, researchers have begun to study the finer details of the activity of the artificial neurons in neural networks that form large language models, as well as the associations among neurons or “neuronal subcircuits” that are induced during training⁴⁹ and patterns of neuron activation at inference time.⁵⁰ One hypothesis is that a large amount of diverse content forces neural networks to learn generally applicable and special-purpose circuits that can support multiple tasks. Such investigations occur largely in smaller models under controlled learning settings. In such work, small models may be promising as more penetrable, understandable “drosophila,” with results that are generalizable to much larger models, just as smaller animal models are used to do medical research aimed at advancing human biology and health care.

Opportunities for more fundamental research include investigations of how principles and methods of probability and decision theory might be more deeply harnessed in representations and inference methods to guide the allocation of computational effort and the selective gathering of information in learning and reasoning.⁵¹ Another direction is to address challenges noted with the ability of generative models to perform planning of the form solved by methods developed in the AI and Operations Research communities for formulating multistep plans via exploration with search and backtracking.⁵² We also see opportunity to move beyond solving single prompts and problems with relatively fixed models to extended presence and situatedness. Directions include exploration of methods aimed at continual reasoning about streams of problems over time.⁵³ Other opportunities include pursuing understandings and extensions of how the models perform and seeking deeper understandings of challenges and opportunities with the physical embodiment of systems, where grounding of concepts and implications of action are developed with flows of information and learning garnered from immersion in rich, realistic environments.⁵⁴

Memory, learning, and adaptation. Deep neural models do not have the ability to quickly learn and adapt as humans do to real-time experiences and information. Once they are trained, these models are then applied but typically remain fixed, or sometimes they are updated via the traditionally long cycle times of fine-tuning. Long cycles for collecting data and building updated models means that late-breaking scientific advances, news, and information will be unavailable to large language models without the use of special machinery to augment inferences. Efforts to address these challenges include extending large models with methods for search and retrieval of recent information. While these adjuvant techniques are helpful, new methods and machinery that enable faster-paced and near real-time memory and learning would be game changing. Opportunities include developing and integrating methods for ongoing, never-ending learning.⁵⁵ Extending abilities to remember, learn, and adapt would enable models to stay up-to-date and would enable breakthroughs in personalization.

Architectural innovation. The Transformer has been a go-to architecture for generative AI. Nonetheless, this architecture and methodology has limitations, such as challenges with handling long-term dependencies in sequences. There are opportunities to innovate with new architectures, including introducing new mechanisms into Transformers.

Reliability, calibration, and trustworthiness. As AI systems become more integrated into daily life, ensuring their reliability and safety is paramount, especially when the methods are applied in high-stakes areas like medicine, criminal justice, education, and industrial process control. Characterizing and communicating potential errors, including erroneous generations and rates of false positives and false negatives in pattern recognition, is critically important in understanding costs of failures. Considerations of types of failures and their rates of occurrence is important in ethical deliberations about uses of AI in specific domains and contexts; AI capabilities and errors frame cost–benefit considerations and decisions that hinge on value considerations.

A weakness of generative AI models is their propensity to generate content that is persuasive yet erroneous. A critical research direction is to develop methods and machinery for assigning well-calibrated confidences to generations and also to deepen understanding of when hallucinating content is expected and desired (e.g., generating fiction) or is a concern (e.g., performing medical diagnoses). Directions include developing internal machinery, fine-tuning, experimenting with new forms of prompting, and calling external tools, such as databases and search engines that perform traditional information retrieval for providing verification and constraints. Recent work has explored careful curation of high-quality datasets, including using large-scale models to generate high-quality data to boost the efficiency of learning and accuracy of inferences.⁵⁶

Some studies have verified good calibration of confidences in specific settings. For example, Figure 8.8 shows good calibration of the confidence of GPT-4 about its answers to multiple choice challenges on competency exams in medicine.

Power of small models. While scaling laws, confirmed by empirical studies and theoretical results,⁵⁷ suggest that large scale is need for top performance, recent work has demonstrated remarkable power with smaller models, some built from high-quality datasets. In recent work, large language models are used to supply training data to build more compact models that show strong performance.⁵⁸ Research is needed to better understand how one can achieve strong capabilities with smaller datasets and computational resources, including questions about whether such model construction depends in some way on the poorly understood special properties of data generated by the larger models.

Grappling with opacity and complexity. Large-scale neural models are difficult to understand, potentially hindering scientific progress dependent on insights about the induction of neural circuits and larger representations. New tools are needed to better understand representations and inference in large-scale models. There is a growing similarity of the “black box” challenges of large-scale neural models and the difficulties of probing the operation of biological nervous systems with fine-grained unit recordings and coarse-grained fMRI and related studies. There is a potential convergence of methods and analytical tools between these fields.

Mechanisms and designs for human–AI collaboration. There are great opportunities to extend prior work on human–AI collaboration.⁵⁹ Although generative AI models are trained to engage in conversational dialogue, there is a large space of possibilities to design interaction strategies that emphasize the primacy of human agency in problem-solving and that introduce new styles of human–AI interaction that enable AI systems to complement human decision-making.⁶⁰ Related goals include developing better ways for generative AI systems to share the rationale for their generations and recommendations.

Engineering Trends with Generative AI

Although it is impossible to predict the future, especially in an area as dynamic as generative AI, it is nevertheless useful to examine emerging trends in the technology that may shape the future of the technology and its applications. In this section we consider several trends that have emerged after the November 2022 introduction of ChatGPT and their potential to change the future of generative AI.

Multimodal generative AI models. Whereas initial LLM models were trained only on text data, one recent trend is toward training models on multimodal data, such as text, image, video, and sound data. For example, Google recently released a model that it calls Gemini, trained “from the bottom up” on such multimodal data. In October 2023, OpenAI made available a version of its GPT-4 system, GPT-4V, which can accept image and text data as input (although its output is still text only). Figure 8.9 shows a typical interaction with GPT-4V, in which it is able to interpret the content of an uploaded image and reason about how to stack the items in a stable fashion.

The significance of this trend toward multimodal models is that such models hold the potential to capture significantly more commonsense knowledge about the physical world—knowledge that cannot be easily captured in text alone. If successful, this trend could lead to significant new applications, for example, systems that observe and guide people step-by-step as they cook a particular recipe for dinner or as you assemble a new piece of furniture. One interesting question is whether successful development of such multimodal models might cause a rapid burst of new progress in robotics, given that much of what limits robotic systems today is their poor ability to interpret and reason about physics of diverse objects and environments.

Power of synthetic data. Generative models and more traditional simulation methods are being used to generate large quantities of training data that are being used successfully to build and extend neural models. Datasets being generated and harnessed includes visual datasets and focused, high-quality distillations of specific types of output, such as reasoning strategies⁶¹ and domain-specific data.⁶²

Incorporating software plugins. LLM’s like GPT-4 exhibit many impressive abilities, they also have many limitations and shortcomings. For example, today’s LLM’s cannot reliably perform arithmetic with large numbers (e.g., multiply 483 times 9,328) and can hallucinate incorrect answers to factual questions. Model plugins consist of traditional software (e.g., a calculator, a database of factual information) that can be called as subroutines by LLMs. Providing LLMs with plugins allows them to overcome numerous limitations and to take advantage of the vast store of software developed by many groups over multiple decades of effort. For example, as of November 2023, ChatGPT had access to approximately 1,000 plugins—from calculators, to web search engines, to restaurant reservation apps—which significantly extend its capabilities beyond those provided by its trained neural network (Figure 8.10).

The model decides whether and when to invoke any given plugin, depending on the prompt it is responding to, but at present most generative AI models limit the number of plugins to be considered in any given conversation. For example, ChatGPT requires users to preselect at most a handful of its available plugins for any given conversation. It remains to be seen how large a set of plugins a model will be able to automatically consider invoking. However, giving models access to the vast store of software developed across the computer industry will be a goal for future systems. One question raised by the rise of plugins is whether generative AI models will become user interfaces of choice to many software packages that currently have their own idiosyncratic interfaces. Will future users prefer to interact in natural language conversation instead of learning the specialized interface for each software application? Plugins are extensions that allow LLMs to impact the world beyond conversations, such as researching.

Beyond such tasks as arithmetic calculation and information retrieval, plugins can enable LLMs to perform myriad functions, including executing actions in the open world, such as making purchases, sending messages, and controlling physical systems. While such integration with broader software and systems can provide new functionalities and services, the new powers also pose risks to safety and security and must be handled with care.

Multifunctional interactive workspaces. In a direction of innovation related to plugins, we see the rise of integrated interactive experiences that promote human–AI collaboration by enabling users and AI components to work together in a step-by-step manner on problem-solving, where multiple tools, data analysis abilities, and code creation are made available in a collaborative approach to solving problems and subproblems. A portion of a sample session of such a multifunctional interactive workspace provided by OpenAI named Advanced Data Analysis is displayed in Figure 8.11.

Software development environments for programming with AI models. In contrast to using tools that enable generative AI models to call other software as subroutines, this trend supports the development of new software systems that instead call generative AI as subroutines. Frameworks such as the open-source LangChain, Microsoft’s Semantic Kernel, and AutoGen have emerged to support software developers in building systems that call multiple instances of a generative model.⁶³ These frameworks make it easier to build software systems that capture the benefits of LLMs (e.g., to interact in natural language, and to perform certain types of commonsense reasoning) while also incorporating standard programming and capabilities missing from generative AI, such as long-term memory and database access. One aspect of generative AI that makes this especially interesting is the ease with which one can “program” or “instruct” an instance of a generative AI model on how to behave. For example, Figure 8.12 shows the text used to instruct, or program, an instance of GPT-4 to perform the role of an artificial student, to be taught by a human teacher within an online education application. The programming of the LLM is done here using only natural language instruction rather than a programming language.

One interesting question about the future which is raised by this trend is whether we are beginning to see the emergence of a new paradigm for software development which, unlike previous paradigms that relied exclusively on formal programming languages to instruct the machine, will in the future seamlessly blend natural language instructions with formal languages (Figure 8.13).

Personalized generative AI systems. Generative AI models such as OpenAI’s GPT-4 and Google’s Gemini are very costly to develop and are so large (containing hundreds of billions of learned parameters) that they are not downloaded, but only used remotely over the web. As a result, it may seem unlikely that these models could be personalized to each of billions of people on the planet. Nevertheless, we are already beginning to see a trend toward personalized LLMs. For example, ChatGPT allows users to provide a natural language description of themselves and their interests which it can use to modulate conversations with that user (e.g., to customize to their educational background). Furthermore, Google has released an experimental version of its conversational assistant Bard that enables users to give it access to their entire Gmail collection as well as their online Google Docs, then discuss the content of these. For example, Figure 8.14 shows a typical interaction with this experimental version of Bard. Beyond this, Microsoft has also released a new version of their Office software suite in which LLMs are integrated with systems such as Word, PowerPoint, and Excel. Both Apple and Google have announced plans to release versions of LLMs small enough to run on their respective mobile phones, opening the possibility of highly personalized LLM-based agents that preserve privacy by operating solely on personal devices.

The significance of this trend is that it suggests that the future will see an increasing ability of generative AI systems to interface with personal data, and data of corporations, in ways that make them tremendously more useful and knowledgeable about the problems of interest to their users. Customization to specific users, corporations, and problem settings is likely to be supported by a combination of model fine-tuning, providing access to relevant user data, and direct natural language instructions defining roles for the agent.

Open-source models. One trend in generative AI might be summarized as “bigger is better.” Between 2018 and 2023 the sequence of top state-of-the-art generative AI models followed a clear scaling law: models with more parameters, trained on larger datasets, produced significantly improved capabilities (Figure 8.15). This led to models with costs of over $100 million to train and containing so many parameters that they would not fit on most computers. Given this trend, one might expect a future in which only a few dozen well-resourced companies and governments could afford to develop the next generation of models, and where the rest of us would only be able to access those models over the cloud. As mentioned earlier in the discussion of research directions, a number of new models being developed and fielded rely on many fewer parameters—few enough that the models can be downloaded and trained or fine-tuned on much smaller computers. Although these smaller models do not match the competence of the very best models, they exhibit surprisingly good competence, especially when trained for specific domains such as medicine or finance, and when trained using carefully selected training data such as textbooks. These small models make it feasible for researchers and developers across the world to build and work with generative AI, rather than just the employees of a handful of organizations; that is, they make open-source shared development by many cooperating developers possible.

Of all the trends mentioned here, this trend toward smaller, open-source, widely shared models may be the most consequential, as it will strongly influence both the number of researchers and developers who participate in advancing the technology, and it will strongly influence the ability of governments to control and regulate uses of the technology and the “guardrails” placed on it.

Consider first the impact of the open-source trend on the number of technical experts who can work to advance the technology. Because current state-of-the-art models such as OpenAI’s GPT-4 and Google’s Gemini are so large and so expensive to train, they can only be accessed over the cloud, and the next generation of these models can only be developed by organizations such as OpenAI, Google, Microsoft, Amazon, and other organizations who have computational infrastructures that cost hundreds of millions of dollars. Such organizations may have many thousands of employees, but this number is dwarfed by the number of researchers and developers outside such large organizations (e.g., university faculty and students in computer science, and employees at small startup companies). Because the rate of research progress is often strongly dependent on the number of researchers working on a problem, a successful and vibrant open-source movement is likely to result in more rapid advances and in the democratization of application development. One concern of the US government as it seeks policies that enable the United States to lead in this technology is the potential loss of university research as a key driver of AI advances. For many decades, US universities drove the key advances in AI. However, in recent years the greatest AI breakthroughs have instead come from industry, because universities lack the high-cost computational resources necessary to train and experiment with the largest, most advanced AI foundational models. One proposal under consideration is to fund a National AI Research Resource (NAIRR) to provide computational resources to keep US universities a vital part of research at the frontier of AI. A pilot NAIRR effort is being organized by the National Science Foundation and is planned for launch in mid-2024.

The success or failure of smaller models and therefore of the related open-source effort in generative AI will also have a strong impact on whether and how governments can track and regulate AI technology. Large corporations that work in this area are already cooperating with various governments to create frameworks, best practices, and regulations to minimize the risk of AI being used for nefarious purposes, as well as risks of adverse unintended consequences. If only very large AI systems dominate in the future, then the open-source movement is likely to be small or nonexistent, and governments can continue to work with large corporations and can effectively enforce any government regulations. However, if small AI models and the corresponding open-source movement succeed, then it will be very difficult, perhaps impossible, for governments to know which organizations and which individuals have highly capable AI models and what they are using them for. In short, if small AI models become highly capable and easily copied and ported, then they will become very difficult to regulate.