Japan's AI-for-science workshop is trying to pick the next five years

THE SIGNAL IN ONE SENTENCE

There is a wonderfully unfashionable question hiding inside a five-day workshop at Nagoya University. What should scientists actually work on? Not which model has the longest context window. Not which vendor has the loudest launch. Not which laboratory can squeeze the word agent into the most grant proposals. Which scientific problems are ready for help from artificial intelligence, and which AI methods might genuinely help solve them? That is the premise of FAIRS Japan 2026, short for Future of Artificial Intelligence for Science in Japan. The workshop began October 5 at Nagoya University's Kobayashi-Maskawa Institute and runs through October 9. Its stated goal is to help define AI and machine-learning research directions in Japan across several scientific domains for the next five years. The official program starts with the science. The first three days cover challenges and opportunities in astrophysics and cosmology, neutrino physics, collider physics and accelerator physics. The final two days turn toward common methods that could bridge those fields and create opportunities for joint research. That order matters. AI-for-science projects can go sideways when a team begins with a fashionable tool and then searches for a respectable problem to attach it to. A model becomes the answer before anyone has agreed on the question. FAIRS Japan is structured to reverse that reflex. Domain researchers describe the hard parts of their work. Groups discuss where AI or machine learning may fit. Later sessions look for methods, datasets and infrastructure that can travel across disciplines. This is not yet a national research plan. It is a workshop designed to help make one. The distinction is important because the event had begun when this article was verified, but the organizers had not published final priorities, funding commitments, project owners or a shared infrastructure plan. Friday's timetable includes a group discussion, group presentation and a white-paper planning session. Those are intended outputs, not completed ones. The most interesting part of the schedule is its use of an unconference format. Instead of filling every hour with fixed lectures, the agenda gives participants repeated discussion sessions organized first by domain and later by cross-disciplinary groups. Summaries follow the discussions. The format creates room for researchers to surface problems that did not fit neatly into the original program. That can be useful in science because the real bottleneck is often not the one that looks best on a slide. An astrophysics team may have an enormous observation archive but weak labels. A collider group may have excellent simulation data and a painful mismatch between simulation and the real detector. A neutrino experiment may need rare-event detection with calibrated uncertainty. An accelerator team may care less about a flashy prediction than about a control system that stays reliable under changing operating conditions. Those examples describe common classes of scientific work, not conclusions already announced by FAIRS Japan. The workshop exists precisely because each community needs to say which problems matter most and what evidence would count as progress. The official timetable makes that breadth visible. It includes sessions on scientific challenges, current AI research in Japan, neutrino datasets, machine learning in particle physics, agentic AI in high-performance computing and Japanese grant opportunities. Confirmed speakers come from universities and research institutions including Nagoya University, the University of Tokyo, Kyoto University, Tohoku University, Chiba University, Osaka Metropolitan University, KEK and RIKEN, with participation from US national laboratories also listed. The event page showed 54 participants when checked on October 5, below its stated cap of 70. That number can change as registration records change. It is useful as a snapshot, not as a final attendance count. The workshop's five-year horizon is ambitious enough to force choices. Five years is long enough to build shared datasets, evaluation suites, computing services and interdisciplinary teams. It is also short enough that a priority should name an observable result rather than float forever as a research aspiration. The cleanest agenda would separate four layers. First comes the scientific question. What phenomenon, measurement or operational problem needs to be understood? Why is the current method insufficient? Which people, facilities or decisions depend on a better answer? Second comes the evidence. What dataset, simulation, instrument or experiment can test an approach? Which baseline should a new method beat? How will uncertainty, reproducibility and failure cases be reported? Third comes the shared machinery. Which computing systems, data formats, software tools and access rules could serve more than one field? A common tool is valuable only if researchers can actually use it, inspect it and reproduce its results. Fourth comes ownership. Who maintains the dataset? Who pays for compute after a pilot ends? Who validates results outside the original team? Who publishes progress when the method fails as well as when it works? That last layer is where workshop energy often evaporates. People leave with a bright diagram and a crowded group photo. Six months later, nobody is sure who was supposed to clean the data, write the benchmark or apply for the joint grant. FAIRS Japan can avoid that fate if its white paper is more than a list of promising nouns. For each priority, the paper could name the scientific challenge, current baseline, available evidence, missing infrastructure, participating institutions, first responsible owner, twelve-month milestone and conditions for stopping or redesigning the work. It should also distinguish between shared methods and field-specific validity. An anomaly detector may be reusable across several instruments. Its meaning still depends on the instrument, calibration process and physical theory around it. A workflow agent may help submit jobs to a computing cluster. That does not make its scientific interpretation trustworthy. A model may summarize a paper or suggest code while remaining a poor judge of whether an experimental result survives systematic uncertainty. The shared layer should reduce duplicated engineering. It should not flatten the science. There is a practical reason to do this work in Japan now. The schedule links university researchers with national research organizations, large scientific facilities and computing expertise. Those communities already operate instruments and collaborations whose useful lives stretch across years. If they can agree on common data practices, evaluation methods and computing access, a successful tool has somewhere to live after the demonstration. That is the difference between AI as a visiting performer and AI as research infrastructure. The human consequences are less theatrical but just as important. A shared agenda can help early-career researchers see which skills and projects have durable institutional support. Common tools can spare small teams from rebuilding the same data pipeline. Clear evaluation practices can make negative results useful. Better access rules can keep AI-for-science from becoming a club for laboratories with the largest computing budgets. The opposite outcomes are possible too. A five-year plan can concentrate resources around fashionable methods, reward groups already inside the room and turn a provisional benchmark into a gatekeeper. Shared infrastructure can become mandatory infrastructure. A white paper can sound national while reflecting only the people with time, travel support and confidence to attend. The workshop page says domestic travel support is limited by budget, and registration closes when participation reaches 70. That does not invalidate the meeting. It does mean the eventual agenda should disclose how input was gathered, which communities were missing and how people outside the room can challenge or extend the priorities. The public should also be able to see what happens next. A useful progress page could track every priority from proposal to dataset, benchmark, pilot, publication and shared service. It could record compute used, institutions involved, reproducibility status and the most important unresolved risk. It should include projects that stopped, because a well-documented dead end can save another laboratory years of effort. None of that has been promised in the reviewed event materials. It is the standard the workshop's ambition invites. The plain signal is that choosing the right scientific problems may be more important than choosing the newest AI model. Nagoya's workshop has put the questions first. The next test is whether five days of questions become a five-year ledger that researchers can inspect, fund and revise.