Google DeepMind's Co-Scientist Extension Brings AI into Real-World Labs

Google DeepMind published "Accelerating Scientific Research with Gemini in the Real-World," upgrading its Gemini-based Co-Scientist from an in-silico hypothesis generator into an execution-based research partner that designs experiments, writes code, and interacts with real lab hardware across the full research pipeline, from ideation to paper writing. The work, a collaboration among Columbia University, Duke University, Google, and Texas A&M, was validated in real-world physical and clinical settings and is framed as an early step toward fully automated AI laboratories, assisting rather than replacing researchers (@burkov).

Confirmed

• The system was deployed with tiered autonomy across domains: human-executed in materials science, expert-AI collaborative in biology, and autonomous in computer science, plus a study of 150 AI-generated papers (@SRSchmidgall).
• Materials science: Co-Scientist interacted with a semi-automated CVD instrument and designed a safer precursor, C2Cl6, for bottom-up MXene synthesis; after extensive human-led lab iterations, the team grew a 2D material structurally similar to Ti3C2Tx (@SRSchmidgall, @iScienceLuvr). @philschmid noted it produced 3 atomically thin semiconductors on first attempt; @markk added these include monolayer semiconductors such as MoS₂, with the system generating synthesis recipes based on lab hardware and inventing new precursor routes. The stated vision includes breaking sub-3nm scaling limits of silicon chips.
• Biology: from sparse imaging data alone, the system quantitatively predicted collective swarming patterns of engineered E. coli across IPTG concentrations, matching wet-lab measurements it had never seen (@DrSingularity); @philschmid reported it correctly predicted 3 of 4 growth metrics.
• Computer science: the system autonomously discovered a test-time scaling architecture that beat six frontier models in blinded evaluations including blind physician review, with results on HealthBench (@omarsar0, @DrSingularity).
• In the study of 150 AI-generated papers, validated through 450 double-blind expert reviews, its tools reduced hallucination and plagiarism; @philschmid cited a hallucination rate of 4%.

Unconfirmed

• Stronger claims such as "AI-controlled chip manufacturing" appear only in second-hand reposts; the primary material covers CVD-based MXene synthesis and the sub-3nm scaling vision, so actual deployment scale awaits official detail.

Why it matters

• This is a landmark case of AI scientists moving from simulation to real-world experimentation, closing the loop across hypothesis, experiment design, hardware execution, and feedback.
• The materials results show practical value in safer chemical routes and first-attempt synthesis of new materials like monolayer MoS₂; the biology predictions quantitatively matched wet-lab data; the discovered reasoning architecture shows transferability to AI research itself; the 4% hallucination rate offers a reference for trust-sensitive uses like paper review.
• Google stresses this is early-stage work far from fully autonomous science, but it provides a reusable deployment paradigm for cross-disciplinary human-AI research collaboration.

2026-08-28 ~ 2026-08-30 · 30 related posts

• Google's Co-Scientist autonomously discovers molecules and algorithms — iScienceLuvr · 2026-08-28
• Google DeepMind brings AI Co-Scientist from simulation to real-world experiments — omarsar0 · 2026-08-28
• [source] Google releases Co-Scientist to accelerate real-world scientific discovery — SRSchmidgall · 2026-08-29
• [source] Gemini Designs Precursor for MXene Synthesis in New Materials Science Experiment — SRSchmidgall · 2026-08-29
• [source] Co-Scientist deployed across materials science, biology, and CS with varying autonomy — SRSchmidgall · 2026-08-29
• Google Extends Co-Scientist to Accelerate Real-World Discovery — vivnat · 2026-08-29
• Google shows Co-Scientist extension to accelerate scientific discovery — ChengleiSi · 2026-08-29
• Google upgrades Co-Scientist with Gemini to design experiments, write code, and control lab hardware — SRSchmidgall · 2026-08-29
• New Work: Accelerating Scientific Research with Gemini in the Real-World — SRSchmidgall · 2026-08-29
• Google's Co-Scientist connects to lab hardware for end-to-end automated research — SRSchmidgall · 2026-08-29
• Google shares progress on Gemini-powered real-world scientific discovery — SRSchmidgall · 2026-08-29
• DeepMind extends Co-Scientist, using Gemini to accelerate cross-disciplinary research — SRSchmidgall · 2026-08-29
• Google Gemini Co-Scientist Generates Hypotheses and Discovers Medical AI Architecture — Dr_Singularity · 2026-08-29
• Gemini Co-Scientist predicts living-cell behavior accurately — Dr_Singularity · 2026-08-29
• Google Gemini Co-Scientist discovers new medical AI architecture — Dr_Singularity · 2026-08-29
• Google AI connects Gemini to lab hardware to advance materials science automation — SRSchmidgall · 2026-08-29
• Google's Co-Scientist evolves from in silico ideation to lab-based execution partner — SRSchmidgall · 2026-08-29
• Google's Co-Scientist accelerates research and achieves AI-controlled chip fab — SRSchmidgall · 2026-08-29
• Google DeepMind expands Co-Scientist for execution-grounded research with Gemini — yawnxyz · 2026-08-29
• Gemini Co-Scientist controls CVD systems to accelerate 2D materials fabrication — SRSchmidgall · 2026-08-29
• Gemini-powered multi-agent system conducts closed-loop scientific research — burkov · 2026-08-29
• Google DeepMind's Co-Scientist Achieves Full Loop in Physical Labs — AI寒武纪 · 2026-08-29
• Gemini Co-Scientist cuts AI hallucinations to 4% in lab trials — _philschmid · 2026-08-29
• Google's Co-Scientist Accelerates Discovery in Materials Science and Biology — SRSchmidgall · 2026-08-30
• Google's Co-Scientist autonomously runs experiments, synthesizing materials on first try — mark_k · 2026-08-30

5 near-duplicate retellings: SRSchmidgall · Dr_Singularity · SRSchmidgall · SRSchmidgall · burny_tech