AI agents shift from scaling to autonomous discovery

By late 2025, the AI community collectively recognized that the era of scaling large pre-trained models—characterized by ever-increasing data, parameters, and compute from 2020 to 2025—was reaching its natural limits. Ilya Sutskever articulated this shift, noting that simply multiplying scale by 100x would not yield transformative breakthroughs, as data availability is finite and the returns on scale are diminishing. This realization has propelled a return to the 'age of research,' where innovation focuses on fundamentally new training paradigms beyond brute-force scaling, such as reinforcement learning and alternative compute utilization strategies.

This transition also marks a conceptual evolution from viewing AI as a static, fully trained artifact toward systems that learn continually and adapt dynamically, mirroring human cognitive processes. The community is increasingly emphasizing continual learning and gradual deployment as safer and more effective pathways to superintelligence, where AI models incrementally acquire skills on the job rather than arriving as finished products. As one analysis put it, superintelligence is now envisioned as a learning algorithm capable of mastering all human tasks over time, fundamentally redefining the AGI paradigm away from monolithic pretraining.

Inspired by neuroscience, researchers are pioneering architectures like liquid neural networks that overcome the core flaw of epistemic stasis—where models are frozen post-training and unable to adapt without costly retraining. These dynamic systems adjust internal states in real time, enabling continuous learning and generalization without catastrophic forgetting. This shift aims to create AI that evolves through interaction with its environment, exhibiting biological-like adaptability and fluidity, thereby fostering a symbiotic intelligence ecosystem that augments rather than replicates human cognition.

By early 2026, leading figures like Demis Hassabis highlighted the community’s pivot from viewing large language models as endpoints to considering them foundational components upon which more sophisticated systems—such as interactive world models—will be built. While breakthroughs beyond scaling laws remain uncertain, the focus is clearly on research innovation that integrates grounding, spatial understanding, and planning capabilities. This strategic redirection is expected to catalyze a new golden age of scientific discovery and medicine, underscoring a broader shift from engineering scaling challenges to high-risk, high-reward research that demands deep domain knowledge and refined intuition.

By late 2025, autonomous AI agents such as GPT-5 and platforms like FutureHouse Kosmos and AI Scientist-v2 began transforming scientific workflows by not only analyzing data but actively generating hypotheses, designing experiments, and producing verified scientific results across diverse domains including mathematics, biology, and materials science. These systems automate vast portions of research, exemplified by Kosmos processing 1,500 papers in 12 hours and generating tens of thousands of lines of code that link claims directly to evidence, marking a shift from passive AI tools to active upstream collaborators that explore large hypothesis spaces while humans provide framing and validation.

The early months of 2026 witnessed landmark achievements where autonomous AI agents like GPT-5.2 Pro, Harmonic's Aristotle, and Google's Althia independently solved longstanding Erdős mathematical problems—some unsolved for decades—with formal verification by experts including Terence Tao. While these breakthroughs tackled what Tao called the 'lowest hanging fruit,' they signaled a new era of AI-driven discovery, demonstrating AI’s capacity to generate novel proofs, refute conjectures, and even author full research papers without human intervention, thereby expanding the frontier of autonomous scientific research.

In biomedical and materials science, autonomous AI agents have dramatically accelerated discovery timelines by generating complete analytical pipelines and experimental designs, enabling even junior researchers to perform complex analyses in minutes rather than months. Systems like UCSF’s LLMs reduced biomedical research from two years to six months, while DeepMind’s GNoME predicted over two million novel stable crystal structures, vastly outpacing prior databases. These advances are coupled with iterative AI-human-verifier loops—where AI proposes hypotheses, independent systems verify results, and humans curate priorities—creating fast, scalable, and cost-effective autonomous discovery workflows.

By mid-2026, multi-agent AI systems such as Google DeepMind’s Co-Scientist and FutureHouse’s Robin exemplify a new paradigm where specialized autonomous agents collaboratively generate, critique, and prioritize hypotheses, design experiments, and interpret results across complex interdisciplinary challenges like drug discovery and climate modeling. While these agents significantly speed ideation and experimental design, challenges remain in reliable hypothesis validation and managing AI hallucinations, underscoring the necessity of human oversight. Thought leaders like James Zou envision virtual labs of AI agents augmenting scientists throughout the entire research lifecycle, heralding a transformative shift in scientific discovery.

By late 2025, the fusion of AI reasoning with robotic automation marked a transformative shift in scientific research, enabling collaborative workflows where AI systems and human scientists jointly plan and execute experiments. While fully autonomous, closed-loop 'self-driving' labs remain a longer-term aspiration due to uneven progress in foundational AI capabilities like reasoning and robot learning, early deployments are already reshaping domains such as life sciences, pharma, and materials science, driven largely by market demand and established research infrastructures.

Entering 2026, pioneers like Periodic Labs and DeepMind advanced the integration of AI models with autonomous laboratories to create continuous, scalable scientific workflows grounded in real-world experimentation rather than solely textual data. This approach treats nature itself as a reinforcement learning environment, where AI scientists iteratively test hypotheses through simulations and physical experiments, mirroring a graduate student's methodical derisking process. OpenAI's demonstration of GPT-5 autonomously proposing, executing, and refining experimental protocols without human intervention underscored the feasibility and accelerating pace of these closed-loop systems.

By mid-2026, national labs like Argonne and Oak Ridge exemplified the transformative potential of autonomous labs, with Argonne’s RAPID facilities conducting hundreds of simultaneous experiments across chemistry, materials science, and bioactive factors, achieving throughput unattainable by human researchers alone. Ian Foster likened this revolution to the advent of early computing, highlighting how AI and robotics enable massively parallel experimentation. Meanwhile, Oak Ridge’s real-time learning systems replaced the traditional linear scientific method with fluid, closed-loop processes that continuously adjust hypotheses and experiments based on incoming data.

The emergence of the '4th+ paradigm,' as detailed by Hao Li’s team at Tohoku University in mid-2026, integrates large materials databases, machine learning interatomic potentials, large language models, intelligent AI agents, and automated laboratory workflows into a unified closed-loop framework. This system achieves near-atomic-level accuracy in predicting material properties and dramatically accelerates the discovery of clean energy materials critical for batteries, hydrogen storage, and fuel cells. Despite challenges like standardizing databases and improving AI reliability, the vision is to develop fully automated research platforms that seamlessly blend AI-driven predictions with experimental validation, heralding a new era of data- and autonomous decision-making-driven science.

Verifiability has emerged as the cornerstone of AI-driven scientific discovery, defining the boundary of what tasks AI can effectively automate. As Andrej Karpathy analogizes, verifiability in Software 2.0 mirrors specifiability in traditional software, requiring tasks to be resettable, repeatable, and automatically scorable to be optimizable by AI methods like gradient descent. This principle explains the 'jagged frontier' of AI progress, where domains with immediate feedback loops such as mathematics and protein folding see rapid advances, while complex, ambiguous tasks like strategic planning lag due to verification asymmetry.

Despite AI's accelerating role in proposing scientific hypotheses and generating research outputs, rigorous human oversight remains the critical bottleneck in validation and reproducibility. Studies from 2025 to 2026, including Donald Knuth’s meticulous manual verification of AI-generated mathematical proofs and Google Research’s Science One framework, underscore that AI models excel as proposers but cannot replace independent verifiers. The Science One framework notably combats citation hallucination—previously as high as 21%—by enforcing a 'Chain-of-Evidence' architecture that grounds every claim in authentic sources and logged experimental data, highlighting the indispensable role of interpretability and transparency in trustworthy AI-driven science.

The evolving human-AI discovery loop is characterized by a dynamic division of labor: AI models rapidly generate hypotheses and explore vast problem spaces, while humans retain responsibility for framing questions, imposing constraints, and critically verifying results. This collaborative cycle, described by Terence Tao and exemplified across disciplines from mathematics to materials science, ensures that AI’s creative speed is balanced by human judgment and accountability. However, current AI systems suffer from 'epistemic stasis'—being static and unable to adapt without costly retraining—prompting research into liquid neural networks that enable continuous learning and real-time interpretability, essential for deepening human-AI symbiosis in scientific workflows.

As AI agents increasingly automate scientific workflows—from hypothesis generation to experiment design and data analysis—the challenges of interpretability, validation, and safety become more acute. Frontline researchers like James Zou envision virtual labs of AI agents accelerating research, yet studies reveal that these agents often produce plausible but unverified outputs, struggle with creative adaptation, and require rigorous human debugging and oversight to prevent 'confident nonsense at scale.' IBM’s emphasis on ethical AI training and layered access controls further illustrates the necessity of treating AI agents as insider threats, balancing autonomy with strict safety measures. This complex interplay demands new social and technical infrastructures, including enhanced peer review, transparent reasoning, and frameworks like Science One, to maintain scientific rigor in the AI era.

The evolution of AI training underscores the paramount importance of data quality and human judgment in shaping model success, particularly in specialized domains like coding. As articulated in late 2025 analyses, the 'taste and sophistication' of researchers influence data selection and post-training adjustments, which are as much an art as a science, affecting not only performance but also stylistic nuances such as coding aesthetics or poetry quality. This human-centric curation ensures that models are optimized for real-world effectiveness rather than merely excelling at academic benchmarks, reflecting a nuanced balance between marketing-driven metrics and practical utility.

By early 2026, the frontier of AI training expanded beyond sheer data volume to incorporate sophisticated reinforcement learning (RL) techniques and experimental data from autonomous labs. Companies like Periodic Labs and DeepMind are pioneering physical labs that serve as real-world RL environments, enabling AI to learn from actual experiments rather than just textual data—a shift predicted to outperform models trained solely on literature. This closed-loop system, mirroring a graduate student's iterative scientific workflow, leverages high-fidelity simulations to derisk experiments and generate experimentally verified reward signals, thereby enhancing sample efficiency and grounding AI reasoning in tangible outcomes.

Recent years have witnessed a dramatic scaling of reinforcement learning datasets—from tens of thousands to nearly a million examples in models like Kimmy and Deepseek—unlocking capabilities such as iterative answer refinement and deeper reasoning beyond pre-training alone. OpenAI’s Foundations team, with early access to vast compute resources, has been dissecting the science and scaling laws of RL, revealing that while pre-training largely compresses existing knowledge, combining it with RL promises orders of magnitude improvements in compute and data efficiency. However, challenges remain, including RL’s susceptibility to reward hacking and false belief attractors, underscoring the ongoing need for richer learning paradigms beyond traditional RL.

Despite the massive scale of training data—far exceeding human lifetime experience—and the burgeoning data industry generating billions in revenue for expert labeling and RL environment curation, true sample efficiency gains remain elusive. Progress has primarily stemmed from expanding and improving data distributions rather than breakthroughs in training efficiency, with open-source models rapidly closing the gap to closed-source counterparts by leveraging accessible data. Yet, as Jose Martinez warns, the accelerated adoption of AI also amplifies risks of misinformation propagation, since AI’s logic-driven outputs are only as reliable as the curated, normalized, and high-quality data they ingest, making vigilant data curation indispensable for trustworthy AI advancement.

Physical labs and scaled RL datasets are giving AI agents real-world learning environments, but challenges like reward hacking and false beliefs highlight the limits of current approaches.