For decades, neuroscience has faced a quiet but stubborn problem: the brain speaks in many dialects at once, and no single instrument can read them all. An electroencephalogram traces the faint electrical weather of the scalp. Calcium imaging lights up individual neurons as they fire. Implanted probes eavesdrop on the chatter of small populations of cells. Each method captures something real, yet each produces data in a form that the others cannot easily understand. Reconciling them has traditionally meant painstaking, hand-built pipelines that rarely transfer from one lab, or even one subject, to the next.
At the 2026 World Artificial Intelligence Conference in Shanghai, a group of Chinese research institutions argued that this fragmentation is exactly the kind of obstacle artificial intelligence is now dismantling. During an "AI for Science" media roundtable convened by Huawei, the Beijing Academy of Artificial Intelligence (BAAI) unveiled Wujie·Brainμ1.0, described as the world's first multimodal foundation model built specifically for neuroscience. Alongside presentations from the Chinese Academy of Sciences, the Shanghai Laboratory of Artificial Intelligence and Tsinghua University, the message was consistent: research cycles that once took years are being compressed toward days, not by replacing scientists but by relieving them of the mechanical work that surrounds discovery.
What Happened
Brainμ's central claim is deceptively simple. Where earlier tools treated EEG readings, calcium imaging and neural-probe recordings as incompatible streams, Brainμ folds all three into a single shared encoding framework. Signals that used to require separate models and separate assumptions can now be aligned and interpreted through one common architecture, making it far easier to compare data across individuals and across experimental setups. The reported value is not a benchmark score but a change in what becomes possible when previously siloed measurements can finally be read side by side.
The model is not a laboratory curiosity waiting for a use case. BAAI says Brainμ has been trained on data drawn from more than 70,000 nights of sleep, and that it has been running continuous, automated analysis across partner laboratories for over a year. Its most striking result arrived in June 2026, when a study built on the model was published in the journal Science. For the first time, researchers showed that reactivating a memory during sleep has opposite effects depending on the memory's emotional tone: positive memories were associated with better sleep quality, while negative ones deepened sleep fragmentation. The finding hints at new treatment routes for the sleep disturbances that so often accompany depression and anxiety.
Lei Bo, a researcher at BAAI, was candid about the field's immaturity. Neuroscience still lacks the kind of standardized, large-scale datasets that powered the rise of language models, and AI in the discipline remains at an early stage. Yet he framed that as a strength rather than a weakness. Unlike the early days of large language models, where capability often arrived in search of an application, here the demand is pulling the technology forward. Real experimental needs are shaping how these models evolve, which tends to produce tools that scientists actually use.
Why It Matters
The significance of Brainμ is easier to see when it is placed next to the other systems presented at the same roundtable, because together they sketch a broader shift in how research itself is organized. The Chinese Academy of Sciences introduced ScienceOne Omni, a model spanning mathematics, physics, materials science and astronomy. Built on a three-layer design that encodes scientific data, aligns it with real-world knowledge and then decodes it into domain-specific tasks, it draws on roughly 170 million scientific papers and more than 8,000 specialized research tools. According to CAS, it has cut some literature reviews from weeks to about twenty minutes and lifted report-writing efficiency several fold, and it is already running across more than a hundred research applications inside the academy.
What links Brainμ and ScienceOne Omni is a rejection of an old trade-off. For years, scientific AI seemed to offer two unappealing choices: broad generalist models that lacked genuine domain depth, or narrow specialist models that performed brilliantly on one task but could not generalize. Xu Nan of the CAS Institute of Automation described ScienceOne Omni not as an incremental update but as a reimagining of what a scientific foundation model can be, one built to reason more like a scientist than a search engine. Brainμ makes the same wager in a single discipline, betting that a unified representation of neural data will unlock questions that no single-modality tool could reach.
This matters beyond any one laboratory because the bottleneck in modern science is increasingly not ideas but throughput. Global research spending keeps climbing, yet the pace of genuine discovery has not kept up. Experiments take too long to run, interdisciplinary collaboration stalls at the seams between fields, and workflows remain fragmented across incompatible tools and formats. A model that can absorb heterogeneous data and reason across it attacks that bottleneck directly, which is why the roundtable's participants kept returning to the same phrase: shrinking the distance between a question and a credible answer.
The Reaction
The framing that drew the most agreement among the researchers was also the most reassuring one: AI here is meant to augment scientists, not to replace them. The stated goal across every presentation was to lift the burden of repetitive labor so that human researchers can spend their attention on the parts of science that still require judgment, intuition and creativity. That is a notably different pitch from the winner-take-all rhetoric that surrounds much of the commercial AI race, and it reflects the particular texture of scientific work, where trust, reproducibility and careful interpretation cannot be automated away.
.jpg){kind=link}
Tsinghua University's Yu Li captured the mood by arguing that AI is shifting from a mere supporting tool into something closer to core research infrastructure, on the level of the microscope or the shared database. That is a meaningful reframing. Infrastructure is not something a scientist chooses to use for a single project; it is the substrate on which entire fields operate. If Brainμ and its peers become that kind of foundation, the interesting question stops being whether a given model beats a benchmark and becomes whether it quietly reshapes the daily rhythm of thousands of laboratories.
There is healthy skepticism embedded in the enthusiasm, too. Lei Bo's acknowledgment that neuroscience lacks standardized data is not a marketing footnote but a real constraint, and it tempers the more sweeping claims. A foundation model is only as trustworthy as the measurements it learns from, and brain data remains noisy, sparse and hard to compare across species and setups. The researchers' willingness to name that limitation openly is, if anything, a reason to take the rest of their claims more seriously.
What Comes Next
The most concrete glimpse of where this is heading came from the Shanghai Laboratory of Artificial Intelligence, working with the Suzhou National Laboratory. Their system, Owl·AuraID, is a multi-agent platform that automates the full experimental process from sample preparation through data analysis. Rather than plugging into instrument programming interfaces, its agents operate the machines the way a human technician would, watching the screen, clicking through menus and reading off the results. The lab reports that the system now works across six categories of precision instruments and has pushed the autonomous completion rate of certain tasks from 33 percent to 80 percent.
If that approach generalizes, the implications are considerable. An instrument that can be operated by a software agent can, in principle, run around the clock, and a model like Brainμ that can interpret the resulting data without weeks of manual cleanup closes the loop between measurement and insight. The vision on display was an end-to-end pipeline: agents that run experiments, foundation models that read the output, and human scientists positioned where they add the most value, framing the questions and judging the answers. The near-term test will be whether these systems hold up outside the demonstrations that introduced them, in the messier conditions of ordinary labs.
For Brainμ specifically, the path forward runs through data. The model's usefulness will grow or stall depending on whether the neuroscience community can assemble the shared, standardized datasets that Lei Bo flagged as missing. That is as much a sociological challenge as a technical one, requiring labs that have historically guarded their recordings to pool them. The sleep-and-memory result published in Science is a proof of concept; the real measure will be how many more findings emerge once the model is fed a broader diet of brain signals.
Closing Thoughts
It is worth pausing on how ordinary all of this is trying to be. The systems presented in Shanghai are not framed as artificial minds or digital scientists. They are framed as plumbing, as the unglamorous infrastructure that lets discovery flow a little faster. A model that reconciles three kinds of brain signal, another that reads 170 million papers so a researcher does not have to, a third that clicks through an instrument's menus at three in the morning, none of these replace the spark of a good hypothesis. They simply clear the underbrush around it.
There is a long tradition of that kind of quiet acceleration in science. The microscope did not think for anyone, but it changed what could be seen and therefore what could be asked. The shared genome database did not generate insight on its own, yet it turned biology into a field where a question could be tested in an afternoon rather than a career. If foundation models like Brainμ settle into the same role, their most important legacy may be almost invisible: not a single dramatic breakthrough, but a steady compression of the time between wondering and knowing. Whether that promise holds will be decided not at a conference podium but in the accumulated work of the labs that adopt these tools, one experiment at a time.
한글 요약
2026년 상하이에서 열린 세계인공지능대회(WAIC)의 'AI for Science' 라운드테이블에서 베이징 인공지능연구원(BAAI)이 신경과학 전용으로 만든 세계 최초의 멀티모달 파운데이션 모델 '우지에·브레인뮤(Brainμ) 1.0'을 공개했다. 이 모델은 그동안 서로 호환되지 않던 뇌파(EEG), 칼슘 이미징, 신경 탐침 신호를 하나의 공통 인코딩 구조로 통합해, 개인과 실험 환경을 넘나드는 신경 데이터 비교를 훨씬 쉽게 만든다. 7만 일이 넘는 수면 데이터로 학습했고, 이를 기반으로 한 연구가 2026년 6월 Science지에 실려 기억의 감정적 성격에 따라 수면에 상반된 영향을 준다는 사실을 처음으로 보였다.
같은 자리에서 중국과학원(CAS)의 '사이언스원 옴니'(1억 7천만 편의 논문·8천여 개 연구 도구, 문헌 검토를 수 주에서 20분으로 단축)와 상하이인공지능연구소(SLAI)의 '아울·오라ID'(정밀 기기 6종을 사람처럼 조작하는 다중 에이전트, 자율 완료율 33%→80%)도 발표됐다. 연구자들은 공통적으로 AI가 과학자를 대체하는 것이 아니라 반복 작업을 덜어 통찰과 판단에 집중하게 돕는 '연구 인프라'로 자리 잡고 있다고 강조했다.
다만 레이 보 BAAI 연구원은 신경과학이 아직 표준화된 대규모 데이터가 부족한 초기 단계임을 솔직히 인정했다. 파운데이션 모델의 신뢰도는 결국 학습한 측정치의 질에 달려 있으며, 브레인뮤의 진짜 가치는 신경과학계가 공유 가능한 표준 데이터셋을 얼마나 모으느냐에 좌우될 전망이다. 현미경과 공유 유전체 데이터베이스가 그러했듯, 이런 모델의 가장 큰 유산은 극적인 한 번의 돌파가 아니라 '질문과 답 사이의 시간'을 꾸준히 줄여가는 조용한 가속일지 모른다.