Abstract Neural networks usually answer queries with a single forward pass—an analog of Kahneman's System 1. In unfamiliar or high‐stakes regimes, this can be brittle and overconfident. The Scientist AI framework augments a fast Predictor trained to approximate the Bayesian posterior with System 2 inference: structured probabilistic reasoning over intermediate natural‐language statements and causal hypotheses. The units that are generated and related via probabilistic dependencies are not the words or tokens but these Boolean statements about arbitrary properties of the world. Because inference‐time compute is spent only on the query at hand, marginalizing over latent statements can reduce computational uncertainty, produce better calibrated conditional probabilities and confidence intervals, and generate explanations as explicit argument graphs. With data transformed into these statements processed to make a syntactic difference between communication acts and factual hypotheses and the Predictor trained to be consequence‐invariant —insensitive to downstream effects of its outputs (Fornasiere et al., 2026)—this epistemic caution supports a safety case: the system is not imitating humans, has less incentive to acquire implicit goals, and is better suited as a guardrail for untrusted agents.
Abstract Agentic LLM agents can sometimes seem like magic, granting our (almost) every wish. But like the magic genies in stories like the Sorcerer's Apprentice, sometimes you have to be careful what you wish for. The difficulty of precisely specifying what it is we want and unintended consequences, sometimes derail this magic. As with genies, they grant requests in a top‐down manner. The agent autonomously breaks the wish into subgoals, until it reaches concrete actions that can be done by Agentic facilities. If left unsupervised by the user, a misunderstanding can result in disaster. But it's also possible to adopt a bottom‐up approach, starting with concrete examples, demonstrating procedures on these examples, and only then generalizing to a wider set of cases. It is often easier to control and debug, and easier for beginners and nontechnical people. Ultimately, we need flexibility to go in either direction. The contribution of this paper is to bring a bottom‐up methodology to “vibe coding” with Agentic Agents like Cowork and ChatGPT Agent. It draws on the technique of Programming by Example (also called Programming by Demonstration) for conventional programming. The advent of Agentic AI may unblock a 50‐year obstacle that has prevented PBE from more widespread acceptance.
Abstract There remains much more to Artificial Intelligence than understanding how to learn with neural networks. Inspired in part by how babies, and humans in general, tend to choose for themselves what (and when ) to learn in addition to how to learn, this article discusses similar opportunities, in the areas of multiagent systems and robotics. Grounded in past research on the topic, it categorizes opportunities for choosing what to learn as coming from task decomposition or simplification for difficult problems and/or large environments; unsupervised pre‐training of representations and priors; multiagent learning; and learning from human interaction. The article is based closely on a keynote talk with the same title, delivered at the AAAI‐2026 conference in Singapore on January 22nd, 2026.
Abstract Modern artificial intelligence (AI) has achieved remarkable progress by scaling data, model size, and computation. Yet scaling alone does not resolve one of the most fundamental gaps between machine and human intelligence: the ability to learn, adapt, and act from limited experience. This limitation is especially visible when labeled examples, expert feedback, user interactions, or real‐world trials are scarce, costly, or dynamically changing. Humans can often grasp a new concept, adapt to a new goal, or personalize behavior after only a few examples or interactions, but many AI systems remain fragile under the same conditions. My research studies data‐efficient generalization: how learning systems can acquire useful priors, adapt to new tasks, and reuse limited experience under bounded supervision. In this article, I trace a continuous research path from few‐shot learning to meta‐learning, in‐context learning, and data‐efficient agentic learning. I first revisit few‐shot learning as a foundational problem and explain how data‐, model‐, and algorithm‐level priors make learning possible when empirical risk minimization is unreliable. I then describe our work on meta‐learning, where task‐level information and parameter‐efficient adaptation enable fast generalization. Next, I discuss in‐context learning in large language models (LLMs) as an implicit form of learning to learn, and show how its connection to classical meta‐learning provides both theoretical insight and practical guidance. Finally, I turn to large language model (LLM)‐based agents, where data efficiency expands beyond labels to include demonstrations, reasoning traces, tool‐use records, preferences, memory, and feedback. Across AI for science, biomedical knowledge discovery, recommendation, search, and personalization, the common thread is simple: more capable AI should extract more learning signal from less experience.
