"Scientific Progress in Artificial Intelligence: History, Status, and Futures" is a review chapter by Eric Horvitz (Microsoft, formerly President of AAAI) and Tom M. Mitchell (Carnegie Mellon University, AAAI past president), published in 2024 as Chapter 8 of Kathleen Hall Jamieson, Anne-Marie Mazza and William Kearney (Eds.), Realizing the Promise and Minimizing the Perils of AI for Science and the Scientific Community (University of Pennsylvania Press). The authors state that the chapter was commissioned to provide a February 2024 snapshot of the state of AI in support of a series of meetings on AI and the sciences organized by the National Academy of Sciences and the Annenberg Trust. It surveys the history, current status, and expected near-future trajectories of AI as a scientific field.
Summary
The chapter anchors its account in the Association for the Advancement of AI (AAAI) definition of the field as the pursuit of "the scientific understanding of the mechanisms underlying thought and intelligent behavior and their embodiment in machines," and lists its core capabilities as learning, reasoning, problem-solving, planning, language understanding, and visual perception. The authors state that over the last twenty years AI research transitioned from a niche scientific endeavor to an impactful set of technologies.
Their narrative has four parts: the birth of the field and its shifts among representation and reasoning paradigms, culminating in a mid-1980s turn from logic to probabilistic methods; machine learning as the foundation of current AI; two "inflection points," the rise of deep learning and then the advent of generative AI; and a forward look at research directions, engineering trends after ChatGPT, application areas, and sociotechnical issues. The conclusion states that "we have little understanding of large generative AI models" despite their rising capabilities, which it describes as "sprinkled with both systematic and poorly understood weaknesses."
Birth and evolution of the field
The authors trace the framing of thinking as computation to Alan Turing's work on computability, efforts by John von Neumann, Turing and others to build general-purpose computers, McCulloch and Pitts's computational abstractions of neurons, and 1940s cybernetics. They date the modern discipline to the 1956 Dartmouth summer workshop proposal by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, which contained the first use of the phrase "artificial intelligence" and set 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 field then developed overlapping subdisciplines (natural language understanding, problem-solving, planning, vision, robotics, machine learning) and communities organized around logical reasoning, reasoning under uncertainty, and neural-network ("connectionist") approaches, and stimulated cognitive science as a joint subdiscipline with cognitive psychology.
Representations and reasoning
Early work used both perceptrons for visual pattern recognition and symbolic logic, including general problem-solving architectures. Symbolic representations dominated the first several decades, with rule-based expert systems (if-then rules chained in production systems) at the center of attention in the 1970s and 1980s. The authors describe a mid-1980s shift from logic-based methods to statistical approaches for handling real-world uncertainty, for example in medical diagnosis, using probability and decision theory, Bayesian networks and probabilistic graphical models, expected-value-of-information computations to guide which tests to collect, and operations-research techniques such as Markov decision processes. They stress continuity: today's large neural networks extend "in a recognizable line" from 1960s perceptrons, and symbolic reasoning research continues, including work on integrating it with neural models.
Machine learning as the foundation
The chapter defines machine learning as algorithms that improve at a task through experience, whether from large labeled datasets (for example, classifying credit-card transactions as fraudulent) or from active experimentation such as self-play in games. It attributes AI's breakthroughs of the preceding fifteen years largely to machine learning, and traces the field's roots to early game-playing systems in checkers and chess and to Oliver Selfridge's Pandemonium system. Machine learning accelerated in the late 1990s, driven by rising computing power and storage and the growth of online data from the mid-1990s. Enabling advances included late-1980s and early-1990s methods for learning probabilistic graphical models from data and for neural networks, and, underpinning deep learning, backpropagation and convolution.
The chapter distinguishes discriminative models, which output labels (spam filters; hospital-record models predicting readmission, sepsis and infection onset), from generative models, which learn the data-generating distribution and produce new instances (GANs, variational autoencoders, diffusion models and transformers; DALL-E and Midjourney imagery; protein-sequence design; scientific simulation). It likewise distinguishes supervised, unsupervised and self-supervised learning, describing self-supervision (hiding tokens and predicting them, or predicting the next word or video frame) as a shift away from reliance on human-labeled data.
First inflection: deep learning
The authors describe an acceleration in the rate of innovation attributable to deep neural networks (DNNs). They cite a 2009 reduction in word error rates on conversational speech recognition, including the Switchboard benchmark, on which progress "had essentially plateaued for over a decade," followed by AlexNet's performance on ImageNet. DNNs have since reached expert-level performance in dermatological diagnosis from skin images and in reading radiological films (the chapter reproduces a CheXNet visualization highlighting a finding in a chest radiograph). Their Figure 8.2 marks human parity reached on seven challenge benchmarks between 2016 and 2020, covering object recognition, speech recognition, reading comprehension, machine translation, speech synthesis, English language understanding, and image captioning. They cite the AI Index, the annual Stanford study of AI trends, for detailed tracking of this progress (see Stanford HAI AI Index Report 2026 for a later edition), and benchmarks such as GLUE.
Key directions named in the DNN era are representation learning (features learned automatically from raw pixels upward, often more accurate than handcrafted attributes "albeit at the cost of increased data requirements"); robustness and generalization to unseen data; hybrid strategies combining DNNs with simulation, reinforcement learning or symbolic reasoning (AlphaGo is the example); multimodal models; and tools and platforms, including GPUs for parallel matrix operations, optimization methods such as stochastic gradient descent, and the TensorFlow and PyTorch programming environments. The authors describe fine-tuned large models as "platforms," referred to as foundation models, which they characterize as an extension and scaling-up of transfer learning.
Second inflection: generative AI
The authors identify three innovations behind generative AI: the Transformer architecture introduced in "Attention Is All You Need" (Attention Is All You Need), whose attention mechanism lets models learn which parts of the input to weight; self-supervised training on diverse web-scale data, first demonstrated with BERT and followed by the GPT series, LLaMA and others; and instruction tuning, scaled with reinforcement learning from human feedback (RLHF), which shapes models to follow natural-language instructions and sustain conversation.
Scaling laws and emergence
The chapter describes scaling laws, the empirical relationship between model accuracy and parameter count, training data and training compute, as "empirically validated multiple times," reproducing Kaplan et al.'s test-loss curves (Scaling Laws for Neural Language Models). The authors argue that scaling laws predict basic metrics such as next-word error rates but "fall short in anticipating the competencies of models on challenging tasks." They describe task-specific jumps referred to as emergent capabilities at thresholds of parameters, compute and data, reproducing Wei et al.'s figure of jumps on eight benchmarks for five models as a function of training FLOPs, and naming "theory of mind" among emergent abilities. They state that the mechanisms and thresholds that trigger emergence "remain largely unpredictable."
Capabilities and weaknesses
The authors state that the original ChatGPT systems, using GPT-3.5 and GPT-4, surprised observers with powers of abstraction, generalization and composition and "polymathic" capabilities, while "the scientific community does not yet have a good understanding" of how these abilities emerge. Citing Bubeck and colleagues' early GPT-4 survey (Sparks of Artificial General Intelligence: Early experiments with GPT-4), they list weaknesses: confabulation of erroneous but persuasive output, failures at basic arithmetic, and limited ability to solve complex planning problems that traditional AI handled through search with backtracking, which they attribute to the sequential generative process. They also describe steering generalist models to specialist performance by prompting: prompting methods (Medprompt) guided GPT-4 to record performance on the MedQA medical benchmark, and similar prompting produced expert-level results on competency exams in electrical engineering, machine learning, philosophy, accounting, nursing and psychology (their Figure 8.7 compares Medprompt, 5-shot and zero-shot prompting on MMLU subsets and NCLEX nursing exams).
Research directions
The chapter lists open research directions:
- Representation and reasoning. Evidence that intensive next-token optimization under bounded resources induces rich world models "as an ideal form of compression"; study of neurons and "neuronal subcircuits" in small models used as "drosophila" (see Mechanistic Interpretability); deeper use of probability and decision theory to allocate computation; planning with search and backtracking; continual reasoning over streams of problems; and physical embodiment.
- Memory, learning and adaptation. Trained models remain fixed or update over long fine-tuning cycles, so late-breaking information is unavailable without retrieval; the authors call near-real-time memory and "never-ending learning" a potential "game changing" advance.
- Architectural innovation, given the Transformer's difficulties with long-term dependencies.
- Reliability, calibration and trustworthiness in high-stakes domains such as medicine, criminal justice, education and industrial process control, including calibrated confidences for generations and understanding when hallucination is desired (fiction) or a concern (diagnosis) (see Sycophancy and Hallucination). They cite good calibration of GPT-4's confidence on USMLE-derived medical exam questions.
- Power of small models, including compact models trained on data generated by larger ones, and whether this depends on special properties of such data.
- Opacity and complexity, with a possible convergence between tools for probing large neural models and neuroscience methods.
- Human–AI collaboration designs that emphasize "the primacy of human agency" and let systems share the rationale for their outputs.
Engineering trends after ChatGPT
Writing about trends since the November 2022 introduction of ChatGPT, the authors describe:
- Multimodal models, such as Google's Gemini, trained "from the bottom up" on multimodal data, and OpenAI's GPT-4V (October 2023), which accepts images and text; they raise whether such models could produce a burst of progress in robotics.
- Synthetic data from generative models and simulation, including distilled reasoning strategies.
- Software plugins, which let models call calculators, databases and other tools; as of November 2023 ChatGPT had access to approximately 1,000 plugins, though users could preselect only a handful per conversation. The authors note that plugins taking actions in the world, such as purchases, messages and control of physical systems, "pose risks to safety and security."
- Multifunctional interactive workspaces, illustrated by OpenAI's Advanced Data Analysis.
- Software frameworks that call generative models as subroutines, including LangChain, Microsoft's Semantic Kernel and AutoGen, and natural-language "programming" of model instances (the Ruffle&Riley tutoring system). They ask whether this signals a new software paradigm blending natural and formal languages.
- Personalization, including ChatGPT user descriptions, an experimental Google Bard with Gmail and Docs access, LLMs integrated into Microsoft Office, and announced on-device models from Apple and Google.
- Open-source and smaller models. Between 2018 and 2023, frontier models followed a "bigger is better" scaling trend (GPT-1 at 117M parameters in 2018, BERT 345M, GPT-2 1.5B in 2019, GPT-3 175B in 2020, PaLM and Megatron-Turing 540B in 2022), with training costs exceeding $100 million, while smaller 2023 models (Falcon 40B, Vicuna 13B, Alpaca 7B, Phi 1.5 at 1.3B, Phi 1 small at 350M) showed "surprisingly good competence." The authors call this trend possibly "the most consequential" (see Open-Weight Frontier Models).
On the open-source trend, the authors argue that a vibrant open-source movement is likely to speed advances and democratize application development, and note US government concern over the loss of university research as a driver of AI progress because universities lack frontier-scale compute. They describe the proposed National AI Research Resource (NAIRR), with a pilot organized by the National Science Foundation planned for launch in mid-2024. They also argue that if small models become highly capable and easily copied, "it will be very difficult, perhaps impossible, for governments to know which organizations and which individuals have highly capable AI models," making them "very difficult to regulate," whereas a future dominated by very large systems would let governments work with a few large corporations to enforce regulation.
Applications
The chapter surveys application areas: biosciences (AlphaFold and RoseTTAFold; a cross-proteome screen of protein–protein interactions in yeast; RFdiffusion protein design for medications, viral binders and vaccines); health care, which the authors call "a sleeping giant," with diagnosis, treatment planning, remote monitoring and report generation (examples include GPT-4 sessions by Horvitz from January 2023 and LLaVA-Med); physical sciences and materials discovery, citing a Microsoft Research AI4Science and Azure Quantum survey; climate and sustainability, including carbon capture catalysts; education, including GPT-4 deployment at Khan Academy; and engineering design, illustrated with DALL-E 2.
Sociotechnical issues
The authors describe AI as dual use and list reliability and safety, privacy and security, fairness and accountability, data provenance, intellectual property and copyright, military uses and destabilization, the digital divide, labor-market effects, deepfakes and online manipulation. They cite governments of the United States, United Kingdom and European Union calling for study and regulation, and the October 2023 US executive order on Safe, Secure, and Trustworthy AI (Executive Order 14110 (RESCINDED)), quoting its concern that AI could "exacerbate societal harms such as fraud, discrimination, bias, and disinformation; displace and disempower workers; stifle competition; and pose risks to national security."
Relation to companion sources
The chapter pairs with A Rosetta Stone for AI Benchmarks — Ho, Denain, Atanasov, Albanie, Shah (Epoch AI / Google DeepMind, 2025) on the measurement methodology side: Ho et al. provide a capability-stitching framework, while Horvitz & Mitchell supply the historical context. Trends in AI Supercomputers — Pilz, Sanders, Rahman, Heim (Epoch AI / Georgetown / GovAI, 2025) is a companion reference on the compute side. The history-and-status framing offers baseline calibration for the forecasting frameworks in Broad Timelines and AI 2027. Horvitz is profiled at Eric Horvitz and Mitchell at Tom Mitchell. The chapter is an anchor for History of AI.
Relationships
- supports: History of AI, Broad Timelines, AI 2027, Scaling Laws, Emergent capabilities
- depends-on: Attention Is All You Need, Scaling Laws for Neural Language Models, Sparks of Artificial General Intelligence: Early experiments with GPT-4
- related: A Rosetta Stone for AI Benchmarks — Ho, Denain, Atanasov, Albanie, Shah (Epoch AI / Google DeepMind, 2025) (companion methodology), Trends in AI Supercomputers — Pilz, Sanders, Rahman, Heim (Epoch AI / Georgetown / GovAI, 2025) (companion compute-side), Eric Horvitz, Tom Mitchell, NAIRR — National Artificial Intelligence Research Resource, Executive Order 14110 (RESCINDED), Open-Weight Frontier Models, Foundation Models
- instance-of: AI Progress Measurement
Provenance
Single-source summary of the chapter text in Raw Sources/ (pages 147–188 of the volume plus endnotes). Confidence is medium: a sources/ page rests on one document, and the chapter is a February 2024 snapshot whose descriptions of frontier models and trends are fast-decay by the schema's standards. The raw file's frontmatter lists no URL; the chapter's authorship and volume are taken from its first page.