Dasheng takes on real-world science as China's system-level AI partner

Meet "Dasheng": a system-level AI agent for real lab work

On March 1, the Shanghai Academy of AI for Science and Fudan University unveiled an upgraded "NovaInspire: Scientist-Centered AI Open Platform" and introduced its core module: Dasheng.

Dasheng-nicknamed after the Monkey King-is a system-level scientific AI agent that brings together multimodal foundation models, long-term multi-threaded collective memory, expert-level scientific "skills," self-driving laboratories, and a secure, trustworthy framework.

What sets Dasheng apart

• Multimodal reasoning: Works across text, data tables, code, spectra, and images. Useful for unifying papers, ELN notes, figures, and raw instrument output in one loop.
• Long-horizon memory: Keeps threaded context across projects, experiments, and collaborators. Ideal for multi-week campaigns and iterative protocols.
• Expert "skills": Encodes domain tasks-literature triage, hypothesis drafting, method adaptation, parameter suggestion, code generation, and result critique.
• Self-driving labs: Orchestrates instruments, runs closed-loop experiments, and updates models with measured results. See a foundational overview of self-driving laboratories in Nature (2020).
• Secure-by-design: Emphasizes safety, auditability, and trust. Clear logs, permissions, and isolation to protect people, IP, and data.

Why this matters for your workflow

• From ideas to instruments: Turn hypotheses into executable plans, schedule runs, collect data, and refine settings-without handoffs breaking context.
• Tighter feedback loops: Use Bayesian or heuristic search to explore parameter spaces and converge faster on useful results.
• Less glue work: Reduce time spent stitching together ELNs, LIMS, scripts, and spreadsheets. Keep provenance traceable end-to-end.
• Institutional memory: What was learned last quarter persists-methods, edge cases, calibrations-so teams don't repeat dead ends.

Getting your lab ready

• Map the stack: Instruments, drivers/APIs, data formats, ELN/LIMS, storage, and compute. Close gaps that block automation.
• Standardize data: Use schemas, units, and metadata that support reuse and reanalysis. The FAIR principles are a practical baseline.
• Define guardrails: Permissions, sandbox runs, kill switches, and safety interlocks. Require human sign-off for risky steps.
• Start small: Pilot on a stable, well-instrumented task (e.g., reaction temperature tuning, solvent screening, or calibration routines) before scaling up.
• Measure impact: Track cycle time, yield/accuracy improvements, reagent use, and error rates. Keep a clear before/after.

Risks to plan for

• Hallucinations and overconfidence: Require citations, simulation checks, and sanity bounds on proposed parameters.
• Instrument drift and brittleness: Schedule reference runs, validation plates, and periodic recalibration.
• Safety and compliance: Enforce SOPs, chemical/biological safety limits, and audit logs that satisfy internal and external reviews.
• Data leakage/IP exposure: Keep data local where needed, use fine-grained access control, and redact sensitive fields.
• Objective misspecification: Optimize for the right metric-include constraints for cost, safety, and generalizability.

Early use cases worth testing

• Materials and chemistry: Catalyst or electrolyte screening, thin-film deposition settings, reaction optimization under resource limits.
• Bio and health: Media optimization, assay parameter tuning, purification gradients, and automated QC checks.
• Physics and engineering: Controller gains, optical alignments with feedback, and repeatable environmental scans.
• Compute-only loops: Literature synthesis, method transfer suggestions, dataset curation, and error analysis.

What to watch next

• Deeper toolchains: Tighter links between ELN/LIMS, schedulers, robotics, and model-based planners.
• Standards for trust: Shared formats for provenance, evaluations, and safety attestations across institutions.
• Multi-agent setups: Specialized agents for planning, execution, and critique coordinating as one system.

Upskill your team

If you're preparing to adopt agents and self-driving workflows, this practical route map helps with tooling, data, and lab integration: AI Learning Path for Research Scientists.

Dasheng signals a clear direction: research that keeps momentum from idea to instrument, with safety and provenance built in. The labs that prepare their data, guardrails, and teams now will feel the compounding benefits first.