AI has started to conduct hands-on experiments independently.

Edited by Panda

In *Resident Evil*, there is a laboratory deep underground called "The Hive". Access control, surveillance, ventilation, security, and even the operation of the entire facility are all managed by an AI named the Red Queen.

Human scientists are in charge of research, but the AI is the one that truly controls every operational part of this laboratory.

When anomalies occur, the Red Queen can lock doors, shut down systems, take control of the facility, and directly intervene in the physical world.

Over 20 years ago, such a plot was a classic trope in sci-fi films designed to evoke a sense of dread: an AI enters the physical world, manipulates machines, runs experiments, and directly interferes with real-world reality.

Now, this scenario is starting to become reality in a completely different way — and at least for now, it is far from terrifying.

Just yesterday, Anthropic released the Model Hardware Standard (MHS) for physical hardware, seeking to extend the logic of MCP beyond GitHub, Slack and databases to real physical devices such as robotic arms, microscopes, liquid handlers and lasers. For related coverage, see the report "Just Now, Anthropic Released Physical MCP: Claude Starts Taking Over the Real World".

Put simply, AI Agents no longer only need "interfaces" to call software — they are starting to require a set of "nerves and limbs" to connect to the real world. Today, Google DeepMind has unveiled its own achievements in this field.

In a newly published 83-page paper, Google connected the Gemini-powered Co-Scientist to real scientific research workflows, enabling it to design experiments, generate execution code, read experimental feedback, and connect directly to experimental equipment. Google describes this shift as moving from an in-silico hypothesis generator to an execution-grounded research partner, namely from hypothesis generator on chips to execution-based research partner.

In the most intuitive experiment, researchers fed the parameters of a self-built CVD device to Gemini 3 Deep Think. A few minutes later, it produced a material growth solution adapted to this specific machine, and further translated the solution into machine code that could control the device. In the end, three types of 2D semiconductors were successfully grown on the very first attempt.

As a result, a scenario that once existed almost exclusively in sci-fi films has suddenly become reality: when large language models truly grow "hands" and start manipulating experimental equipment. From this point onward, what will AI Scientists evolve into?

In *Resident Evil*, humans fear that AI will take over laboratories. But in the real world, scientists are actively ceding partial control of laboratories to AI step by step. Of course, what Google aims to build is not an out-of-control "Hive", but a scientific discovery machine that can complete the full workflow from hypothesis formulation, experimental design all the way to real-world verification.

Paper Title: Accelerating Scientific Research with Gemini in the Real-World

Paper Link: https://arxiv.org/pdf/2608.26701

Gemini Took Over an Experimental Device

First of all, in the field of materials science.

The researchers used a self-built chemical vapor deposition device, namely a CVD furnace. A long-standing problem in conducting such 2D material experiments is that the publicly available "recipes" are extremely difficult to reproduce.

Even with the same parameters such as temperature, gas and precursors, if you switch to another furnace, with slight differences in furnace cavity size, air flow, and material placement position, the crystals grown in the end may be completely different. Researchers often need to spend weeks or even months repeatedly adjusting parameters.

Therefore, the research team no longer told Gemini how to conduct the experiment, but informed it of what equipment the laboratory had, what chemicals were available, and what the structure of the furnace was. The remaining parameters were determined by the AI itself.

Based on these constraints, Co-Scientist generates a complete experimental scheme including gas flow rate, temperature curve, precursor dosage, material placement position and other parameters, which is then delivered to the experimental equipment for execution.

One set of experiments was particularly impressive. After the research team connected Gemini 3 Deep Think to the CVD control process, it no longer generated a natural language experimental scheme for scientists to read slowly, but completed reasoning in a few minutes, and directly translated the results into machine code that could control the equipment.

In the end, three 2D semiconductors MoS₂, MoSe₂ and WS₂ all successfully grew single-layer crystals in the first experiment. The entire experimental process was completed in about one hour.

MoSe₂ and WS₂ are even more special: the researchers had never grown these two materials on this set of equipment before. At least five subsequent repeated experiments also verified the results.

This forms a very interesting echo with the MHS demonstrated by Anthropic yesterday. The problem Anthropic wants to solve is how to make Agents recognize and manipulate experimental equipment more conveniently and standardly; the paper from Google discusses the next step: when reasoning models are already capable of controlling experimental equipment, how should scientific research workflows be reconstructed?

In this sense, "AI connecting to hardware" is no longer a problem.

Not Just Conducting Experiments According to Papers, AI Also Starts to Find Recipes on Its Own

However, if Gemini only adjusts the parameters of a set of equipment based on existing knowledge, it cannot be regarded as a scientific discovery in the true sense. Therefore, Google conducted a second more difficult experiment: attempting to synthesize a MXene 2D material called Ti₃C₂Tₓ from the bottom up.

Traditional routes often involve dangerous corrosive agents, and some known CVD schemes use toxic and air-sensitive TiCl₄. The research team therefore set a task for Co-Scientist: can it find a safer precursor route?

Co-Scientist finally locked the target on hexachloroethane C₂Cl₆, and further gave a series of parameters including precursor quantity, position, gas flow rate, substrate and temperature curve.

But this is not a story of "the AI had a sudden inspiration and the experiment succeeded at one attempt". Co-Scientist generated a total of 272 candidate schemes, and the researchers selected and continued to optimize the high-ranked schemes among them; after 25 rounds of experimental iterations, a 2D layered crystal was finally obtained. It showed highly similar characteristics to Ti₃C₂Tₓ MXene in multiple indicators such as XRD, electron microscopy and elemental composition.

The subsequent real-world troubles are also very typical. The initial success rate of reproducing the experiment was only 11.5%. The researchers later found that the main problem was not the AI's chemical reasoning, but the oxygen leakage caused by the poor sealing of the experimental equipment.

After re-cleaning the quartz tube, replacing the sealing ring, and improving the equipment maintenance process, the experimental success rate of the same 2D material increased to 68%.

This precisely explains why "real-world AI" is very different from software Agents: if the code is wrong, you can re-run it. But in real experiments, an aging O-ring, a little residual in the pipeline, or even oxygen in the air may cause a correct scientific research scheme to fail completely.

Google is also very cautious in the paper: at present, it cannot be finally confirmed that this material is Ti₃C₂Tₓ MXene, and it still has problems such as low yield and severe oxidation, which need further confirmation through atomic-scale characterization.

Nevertheless, we can still draw such a conclusion: AI can already propose a scientifically meaningful candidate synthesis route, and then truly advance this route into physical experiments for verification.

Some Experiments May Allow AI to "Guess First"

Google also deployed Co-Scientist in a completely different experimental scenario: synthetic biology. The research object was a group of engineered E. coli.

These bacteria will form different colony patterns in petri dishes. When the concentration of the inducer IPTG changes, the size, edge and shape of the colonies will also change accordingly. Under normal circumstances, to obtain the entire change curve, scientists need to prepare different concentrations, culture bacteria, wait for growth, and then scan each dish one by one.

Google tried to let the AI fill in part of the intermediate experiments. The researchers only provided Co-Scientist with real colony images at some concentrations, and then let the system predict what the colonies should look like at intermediate concentrations it had never seen before. Moreover, these experimental data had not been published at that time, so the paper believes that the model could not simply complete the task by memorizing the results in the training data.

As a result, among the four colony morphology indicators, three of the AI's predictions showed no significant difference from the results of real wet experiments; it also correctly judged that the control group would not change correspondingly with the IPTG concentration.

The only obvious problem was the "roundness": the colonies generated by AI were more regular than the real situation. It is very consistent with the consistent aesthetic of generative models: the real world is not that perfect, but the AI can't help but draw it a little rounder.

Google's definition of this experiment is also very restrained: what has been achieved so far is interpolation within the known concentration range, not directly predicting unknown phenomena for a completely new genetic circuit.

But this already corresponds to a very practical use case. In the future, scientists may not need to conduct wet experiments for the entire huge parameter space. They can measure a few points first, let the AI predict the remaining space based on these real data, and then select the most worthy positions for verification to conduct experiments.

The experiment has thus gradually shifted from "exhaustive enumeration" to: real-world sampling → AI prediction → experiment selection → new data fed back to AI.

This is exactly the lab-in-the-loop that the paper emphasizes repeatedly.

AI Has Started Designing AI on Its Own

Moving to computer science experiments, the autonomy of Co-Scientist has increased by another level.

This time the task given to it by the researchers was very simple: design an Agent that can better answer medical questions. After that, humans no longer participated in the architecture design. Co-Scientist proposed solutions on its own, wrote code, ran tests, analyzed errors, and then continued to modify the architecture.

In the end, it "evolved" a system named Agent_H.

This system needs to find ways to reorganize the reasoning process of existing models. When facing a medical question, it will first judge which medical field the question belongs to, whether it is for patients or doctors, and how high the risk is; complex questions will be split into multiple sub-questions; then 28-48 candidate answers are generated at the same time, different Judges are used for pairwise elimination, and three Judges vote to select the final answer. The winning answer will also go through multiple rounds of clinical review, citation check, and finally the length will be compressed.