A new line of tools aims to let artificial intelligence carry out the full research cycle with little help from people. The project, often called an AI scientist, reads papers, forms a clear question, runs code-based tests, and drafts a paper with figures and references. It even applies basic peer review on its own output to spot weak claims and remove made-up details. The first public demos focus on machine learning research, not wet labs, and follow a strict, step-by-step script to lower errors.
The workflow is simple to understand. The system searches the literature, lists possible ideas, and scores them on interest, novelty, and feasibility. It then checks that the best idea is truly new, edits a starter codebase, runs the experiments, keeps notes, and writes the paper. A review module grades the paper against standard venue criteria and asks for fixes when needed. This process can loop so the system learns from each round.
Limits remain. The models can still “hallucinate,” which means stating facts that are not supported by sources. They may also show popularity bias, leaning toward crowded topics. Early tests suggest its judgments can look like those of a young graduate student, with smart ideas mixed with clumsy readings of the results. In short, the tool is promising, but it is not ready to replace expert review and careful replication.
The best results appear when people and machines work together. AI can search a huge space of options fast and without human habits getting in the way, which can unlock fresh routes to test. In physics, for example, teams have used AI to design unusual quantum optics experiments that humans had not thought of, and later showed that they work. In medicine, AI has already helped scientists find new antibiotic candidates, a sign that this kind of help can speed discovery when paired with lab checks.
The take-home message is clear. These systems can help plan, test, and write, but they still need human goals, ethics, and domain sense. Used well, they can save time and widen the range of ideas worth trying. Used alone, they can recycle trends or miss the real meaning of a result. Human guidance stays at the center.
The AI Scientist – Towards Fully Automated Open-Ended Scientific Discovery – 2024
Technical report that defines the end-to-end pipeline, from idea generation to automated review, applied to machine learning topics like diffusion models and transformers. It describes novelty checks, experiment execution, paper drafting, and a reviewer that scores papers near human levels.
Sakana AI blog – The AI Scientist – 2024
Project announcement with a plain-language overview of the system, its phases, examples of generated papers, stated limits, and ethical concerns, plus links to code and report.
WIRED – An ‘AI Scientist’ Is Inventing and Running Its Own Experiments – 2024
Reporting on the UBC, Oxford, and Sakana AI effort, noting that current outputs are incremental, that open-ended exploration is the goal, and that expert voices see both promise and risk.
IEEE Spectrum – Will the “AI Scientist” Bring Anything to Science? – 2024
Critical view that the system often behaves like an early PhD student, mixing creative ideas with weak interpretations, and that careful human oversight is still required.
KJZZ News – UBC’s AI Scientist automates scientific discovery – 2024
News brief confirming the collaboration between UBC, Oxford, and Sakana AI, and describing the goal of automating the full discovery pipeline.
Nature – Discovery of a structural class of antibiotics with explainable deep learning – 2023
Peer-reviewed study showing that deep learning can surface a new structural class active against MRSA, with an approach that explains which chemical parts drive activity.
MIT News – Using AI, MIT researchers identify a new class of antibiotic candidates – 2023
Plain-language explainer of the Nature paper above, highlighting how explainable deep learning guided screens to find candidates that worked in lab and animal tests.
MIT News – Artificial intelligence yields new antibiotic (halicin) – 2020
Report on the earlier machine-learning discovery of halicin, a broad-spectrum antibiotic candidate, marking a landmark for AI-guided drug discovery.
Scientific American: AI designs quantum physics experiments beyond what any human has conceived – 2021
Feature on AI systems that propose non-intuitive quantum optics experiments which are later validated, illustrating how AI can bypass human bias and enable new tests.
Sakana AI – The AI Scientist generates its first peer-reviewed publication – 2025
Update claiming an AI-generated paper accepted at a workshop, suggesting rising quality of the pipeline, though still within narrow research scopes and venues.
MIT News – Using generative AI, researchers design compounds that kill drug-resistant bacteria – 2025
Follow-on report that generative models designed antibiotic candidates effective in preclinical tests against hard infections, underscoring rapid progress when AI and lab work are combined.
0 Comments