AI scientists in Nature test hypotheses and round-the-corner LiDAR imaging

Overview

The podcast examines how AI agents assist scientists in generating testable hypotheses, testing them in a lab loop, and accelerating discovery. It highlights two AI systems, Robin from Future House and Co Scientist from Google DeepMind, and contrasts their approaches to ranking and refining ideas. The episode also introduces a tangible demonstration of non line of sight imaging using LiDAR on consumer grade hardware and discusses broader implications for science and society.

Key insights

• AI aided hypothesis generation leads to exploring thousands of biological mechanisms and drug interactions beyond what a single researcher could contemplate.
• Lab in the loop vs human guided AI approaches differ in how hypotheses are refined and whether experimental protocols are produced directly by the system.
• Drug repurposing and mechanism based reasoning show potential to identify therapies more quickly by connecting existing drugs to new targets.
• Round-the-corner imaging via LiDAR demonstrates the possibility of tracking, shaping, and localizing objects around obstacles, with realistic limitations for consumer devices.

Introduction to AI scientists in science

The podcast discusses a shift in scientific workflows where artificial intelligence tools are designed to assist researchers in formulating testable hypotheses, performing data driven analyses, and aiding decision making. These tools are presented as complements to human expertise, with the aim of expanding the range of ideas scientists can consider and accelerating the pace of discovery. The conversation frames the advancement as a potential democratization of scientific capacity, while also raising questions about model reliability, reproducibility, and the role of humans in guiding research priorities.

Robin from Future House: accelerating hypothesis generation

Robin is described as an AI scientist designed to identify novel hypotheses by scanning vast literature and datasets, including research articles, patents, and clinical trial data. The approach emphasizes exploring thousands of potential biological mechanisms for a disease and generating many possible drugs for testing. Robin leverages a suite of AI tools—Crow and Falcon—to scour the literature, propose mechanisms, and test drug hypotheses. The process is structured as a tournament where hypotheses are ranked by the strength of their scientific rationale and supporting evidence. A key advantage highlighted is the combinatorial explosion of possibilities that human researchers cannot realistically consider alone.

In a practical case, Robin focuses on dry age related macular degeneration. The system identifies phagocytosis as a potentially beneficial immune process and connects it to brookinase inhibitors, a drug class that can stimulate phagocytosis. After identifying ripasudil as a candidate drug, the human team confirms its potential through cellular and animal experiments. Robin’s strength lies in refining hypotheses through its own data analyses and proposing logical follow up experiments.

Co Scientist: a tournament based autonomous ideation system

Co Scientist, built by Google DeepMind, functions similarly by mining the literature but adds a tournament style mechanism and uses its past work on game playing to guide hypothesis selection. The system is designed to evolve ideas iteratively, with autonomous agents critiquing and refining proposals to ensure novelty remains grounded in science. It emphasizes ranking ideas that are both new and realistic, and avoiding suggestions that are purely speculative or unverifiable. Human scientists then test the selected ideas in the lab or in silico, creating a loop between AI generated hypotheses and empirical validation.

Both systems aim to produce hypotheses that are testable and actionable. They differ in whether they produce experimental protocols directly and in how they incorporate human feedback, but both acknowledge the risk of hallucinations and misinterpretations rooted in the source literature. The podcast highlights the importance of validation and ethical considerations in the dissemination of AI assisted scientific results.

Addressing hallucinations and reproducibility

The hosts and researchers acknowledge that large language models can generate plausible but false information. Robin restricts its answers to information it actually possesses, attempting to minimize hallucinations, whereas Co Scientist uses a team of agents to critique and verify ideas. The speakers note that even with safeguards, AI systems can propagate erroneous findings if the underlying literature is flawed or not reproducible. This underscores the need for human oversight, cross validation, and careful interpretation when leveraging AI in scientific workflows.

Real world demonstrations and ethical considerations

The discussion contextualizes laboratory in the loop discovery within broader ethical debates about how AI tools should be used in science. There is emphasis on reserving critical decision making for human researchers, ensuring transparent reporting, and maintaining rigorous editorial standards for AI assisted content. The conversation also considers how human researchers should allocate attention and resources to the most impactful questions, suggesting a continuing, collaborative role for humans rather than a replacement by machines.

The speakers acknowledge that the technology is early and not yet mature. Future iterations could address issues like model bias, data quality, and reproducibility while expanding the range of use cases in medicine and biology. The podcast closes with reflections on how AI powered tools could democratize scientific work and invite broader participation from fields like robotics and UI/UX design to explore novel applications of AI generated hypotheses.

Non line of sight LiDAR imaging: round the corner sensing on consumer hardware

The podcast then shifts to a study published in Nature about using LiDAR chemistry to see around corners. The approach repurposes consumer grade LiDAR sensors to perform non line of sight imaging, building a picture of hidden objects by analyzing light that reflects off a nearby wall. This technique relies on a motion model that captures how light returns from hidden surfaces and how the camera motion can reveal hidden geometry. The researchers describe four capabilities demonstrated in the paper: tracking objects around corners, reconstructing shapes using camera motion, tracking multiple objects, and deducing the camera position using a hidden landmark. They emphasize that the resolution is not equivalent to a smartphone camera, but the method yields useful information such as whether something is present, whether it is moving, its rough size, rigidity, and approximate shape within sensor limits.

The discussion with Siddharth Somosandoram, a co author from MIT, covers experimental steps and key insights. Early experiments used highly reflective patches to detect a signal in the data and verify the presence of motion at certain frequencies. The team then extended to scenarios that could benefit robotics, including self driving cars or delivery robots, where rapid sensing around occluded regions could improve safety and navigation. The researchers articulate an aperture sampling model for efficient motion representation and discuss challenges such as low light return, noise, and the need for algorithms that can robustly interpret very noisy signals. They also address ethical questions around privacy and potential misuse, noting that the sensor quality does not produce high fidelity images. The potential applications include accident avoidance, search and rescue, and improved localization in robotics, with an optimistic view about making such capabilities accessible to researchers and hobbyists through affordable hardware and publicly available code.

Limitations, ethics, and the human role in AI enabled science

Ethical considerations are a recurring theme. The speakers stress the importance of responsible uses of AI technology and awareness of potential misuse. They also discuss the current limits of the technology, including the fact that visible artifacts may not be interpretable as accurate photographs or direct representations of reality. They emphasize that AI systems should be guided by human expertise to maximize societal benefit and avoid misinterpretation or overclaiming. The editorial perspective from Nature is cited, arguing for scientists to use AI tools to advance humanity while preserving critical human judgment and domain knowledge.

Practical takeaways and access for the public

Listeners are informed that the technology is not yet plug and play for everyday smartphones, but sensor components can be assembled affordably and with open source code. The episode closes with the sense that this is a pivotal moment for AI powered discovery and round the corner sensing, inviting engineers, roboticists, and designers to explore novel uses and contribute to the field.