AGI Made Clear › Anthropic's biology lab
When AI leaves the screen: Anthropic is giving Claude a biology lab
Anthropic has confirmed it is operating a physical biology laboratory. Claude is already accelerating biomolecular research, and the longer-term ambition includes AI working with laboratory equipment. This is a significant step in AI's transition from answering questions to acting in the real world.
For most of the generative AI era, artificial intelligence has lived behind a screen.
We ask a question. The model generates an answer. Perhaps it searches the internet, analyses a document, writes some code or creates an image.
Even increasingly capable AI agents generally operate within the digital world.
Anthropic is now pushing that boundary.
The company behind Claude has confirmed that it is operating a wet laboratory in the San Francisco Bay Area where physical biological experiments are being conducted.
That distinction matters.
A wet lab isn't another AI benchmark or computer simulation. It is a laboratory where scientists work with actual biological materials and equipment.
Anthropic's head of life sciences, Eric Kauderer-Abrams, explained the reasoning simply: In biology, ultimately, theories have to be tested in the real world.
And Anthropic says it is already doing that.
The longer-term direction is even more interesting. The company is exploring how Claude, scientists and automated laboratory equipment could work together, potentially allowing AI systems to help direct experiments carried out by laboratory robots.
AI is beginning to leave the screen.
Why does an AI company need a biology lab?
AI has already become extraordinarily useful in biological research. Models can analyse scientific literature, work with genomic information, predict molecular structures, generate computer code and help scientists reason across enormous datasets.
But biology has an unavoidable limitation. Eventually, something has to be tested.
A computer model can predict that a molecule should behave in a particular way. It cannot establish that the prediction is correct simply by producing another prediction.
Researchers ultimately need physical evidence. Cells have to behave as expected. Proteins have to bind. Experiments have to reproduce. Results have to survive contact with reality.
That creates an important boundary for AI-driven science. An AI may be able to generate thousands of hypotheses extremely quickly. But if humans remain responsible for manually testing every one of them, physical experimentation becomes a bottleneck.
Anthropic's laboratory is an attempt to start connecting those two worlds. AI reasoning on one side. Physical experimentation on the other.
Claude is already being used for biology
The wet lab isn't Anthropic's first move into scientific research. The company has been building a much larger life-sciences operation around Claude.
Earlier this year it introduced Claude Science, a specialised AI workbench designed for scientists. Instead of simply chatting with a general-purpose model, researchers can connect Claude to scientific databases, computational tools and specialist models.
Claude Science can work across areas including genomics, proteomics, single-cell biology and cheminformatics. It can also coordinate multiple specialist tools during a research task.
This matters because modern biological research rarely happens inside one piece of software. Scientists may move between databases, programming environments, scientific papers, specialist models and computing infrastructure. AI agents can increasingly coordinate that process.
But Anthropic's latest research suggests Claude can do more than organise existing tools. It can improve them.
Claude optimised more than 30 biological AI models
Anthropic recently gave Claude access to a collection of open-source biomolecular models. These are specialised AI systems scientists use for tasks such as predicting protein structures and designing biological molecules.
Claude's task was essentially to make them run better.
According to Anthropic, Claude optimised more than 30 models in just under four weeks. On average, the models ran approximately four times faster after optimisation.
Claude also developed techniques that reduced the amount of computer memory required for some biological modelling tasks. That allowed researchers to model much larger biological structures using existing hardware.
This is an important distinction. Claude wasn't merely using scientific software. It was improving the software scientists use to conduct research.
That begins to create an interesting feedback loop. Scientists build AI. AI helps scientists improve scientific tools. Those improved tools enable more experiments. The results of those experiments can then inform future models and research.
Add automated physical experimentation to that loop and something important starts to change.
From AI assistant to AI scientist?
We should be careful with that phrase. Claude isn't independently operating as a scientist in the human sense. Human researchers still define objectives, build infrastructure, interpret results and oversee the work. Anthropic itself emphasises human oversight.
But the division of labour is changing.
Imagine a simplified scientific process: A researcher identifies a problem. They review existing research. They develop a hypothesis. They design an experiment. They run it. They analyse the results. They modify the hypothesis. Then they run another experiment.
Traditionally, humans have performed almost every stage. AI is gradually becoming capable of assisting with more of them.
It can search the literature. Analyse data. Generate hypotheses. Write experimental code. Operate specialist scientific software. Design molecules. Interpret results. Suggest the next experiment.
The remaining boundary is increasingly physical. Someone, or something, still has to carry out the experiment.
That is where laboratory automation enters the picture.
What happens when Claude can control laboratory equipment?
Modern laboratories already contain substantial automation. Robotic systems can move samples, operate liquid handlers, prepare plates and perform repetitive experimental procedures.
The interesting possibility isn't therefore that Claude somehow develops robotic hands. It is that an AI agent could interact with machines that already exist.
A future workflow might look something like this: Claude analyses previous experimental results. It identifies a promising hypothesis. It proposes the next experiment. A human scientist reviews or approves the plan. Laboratory robots perform the physical work. Sensors and instruments record the results. Claude analyses those results. It proposes the next experiment. Then the cycle repeats.
The important change is speed. Scientific research normally contains delays between reasoning, experimentation and analysis. If parts of that loop become automated, researchers may be able to run far more experimental cycles in the same amount of time.
That could have enormous benefits. It could also create new questions about oversight.
The promise is extraordinary
Biology is one of the areas where advanced AI could potentially produce some of its greatest benefits.
Drug development is slow and expensive. Rare diseases often attract limited commercial research because relatively small patient populations make traditional drug-development economics difficult. Scientific literature is growing faster than any individual researcher can realistically absorb. Biological systems are extraordinarily complex.
AI can potentially help researchers navigate that complexity.
Anthropic has talked about using AI to accelerate the development of medicines and treatments, particularly in areas that conventional commercial models may underserve.
If AI can make scientific models faster, help design biological molecules and reduce the time between hypothesis and experiment, the consequences could be enormous.
Research that currently takes months might eventually take weeks. Processes taking weeks might take days. Far more hypotheses could be tested. Failed approaches could be abandoned sooner. Promising ones could be identified faster.
That is the optimistic case. And it deserves to be taken seriously.
But biology is different from writing code
There is also an obvious reason for caution.
If an AI writes bad software, the software can often be deleted, rolled back or isolated. Biology interacts with the physical world. Experiments involve real materials.
That doesn't automatically make AI-driven biological research dangerous. Scientists already operate under extensive laboratory procedures, safety rules and regulatory requirements.
But increasing autonomy changes the nature of oversight.
If an AI system can propose experiments, operate tools and respond to results, researchers need to know exactly what it is allowed to do.
Which equipment can it control? Which materials can it access? Which experiments require human approval? Can it change an experimental plan midway through a process? How are unexpected actions detected? Who is responsible when something goes wrong?
These questions sound remarkably similar to the ones businesses are beginning to ask about AI agents. The consequences are simply much greater.
Anthropic clearly recognises the tension
This is where Anthropic's broader approach to biology becomes important.
The company isn't simply making its most capable biological models freely available without restrictions. It has introduced a Life Sciences Verification Program giving vetted researchers access to more capable models with safeguards adjusted for legitimate biological research.
Anthropic also explicitly acknowledges that advanced AI capability in biology could potentially be used for harmful purposes.
That creates a difficult balance. The same capability that helps a researcher design a useful biological molecule might potentially help someone design something harmful. The same AI that makes legitimate scientific experimentation faster may also make dangerous experimentation easier.
So the question isn't whether AI should be allowed to understand biology. It already does.
The challenge is deciding how increasingly capable systems should be given access to the tools required to act on that understanding.
This connects directly to the wider AI safety debate
Over the last few weeks, several apparently separate AI stories have emerged.
Frontier AI leaders have debated whether development should slow down. OpenAI has disclosed examples of models behaving in unexpected ways. Researchers have demonstrated increasingly powerful AI-assisted cybersecurity. And now Anthropic is connecting Claude to physical biological research.
These stories are different. They shouldn't be collapsed into one dramatic narrative about AI becoming uncontrollable.
But they do share something. Capability is moving outward.
First, AI generated text. Then it generated code. Then it began using software. Then agents began carrying out multi-step tasks. Now frontier models are interacting with increasingly consequential digital systems. And in scientific research, they are beginning to connect with the physical world.
Each step increases what AI can accomplish. Each step also increases the importance of understanding what happens when something goes wrong.
The physical world changes the stakes
There is a useful concept in AI called a sandbox. A sandbox is an isolated environment where software can operate without affecting important external systems.
For much of AI's development, the physical world has acted as a kind of natural sandbox boundary. A language model could suggest something. A human still had to act.
Agents have already weakened that boundary in the digital world. An AI can now potentially send an email, edit software, access a database, make a purchase or operate a browser.
Laboratory automation extends the same idea into physical science. An AI doesn't need a humanoid robot body to act in the physical world. It simply needs access to machines that can.
That makes permissions, monitoring and human approval increasingly important.
Is this AGI?
No.
Operating scientific software, improving biological models or helping direct laboratory experiments doesn't establish that Claude is artificial general intelligence. There is no universally agreed definition of AGI anyway.
But developments like this help explain why the AGI discussion is changing.
The important question isn't simply whether a model can answer questions across lots of subjects. It is increasingly about agency.
Can the system pursue objectives? Can it use tools? Can it plan? Can it learn from results? Can it operate across different environments? Can it carry out increasingly long chains of work with less human intervention?
And perhaps most importantly: Can it affect the world outside the computer running it?
That is why Anthropic's biology lab matters beyond biology.
Scientific automation could create a powerful loop
There is another possibility worth watching.
AI development is frequently discussed in terms of recursive self-improvement, usually in a much stronger sense than anything running today. That usually refers to AI helping researchers build better AI.
But science could develop its own version of an accelerating loop.
AI proposes an experiment. Automation performs it. AI analyses the results. AI proposes the next experiment. Automation performs that. Repeat.
Humans would still define objectives and supervise the process. But if the time required for each iteration falls dramatically, the pace of discovery could increase.
This is one of the central promises of AI-driven science. It is also why governance becomes important before these systems become substantially more autonomous.
The time to decide which decisions require human approval isn't after the human has disappeared from the loop.
We may be watching the next transition in AI
It is easy to think of AI progress purely in terms of model releases. GPT gets smarter. Claude gets smarter. Gemini gets smarter. Benchmarks go up.
But perhaps the more important transition is what those models are being connected to.
A highly intelligent model trapped inside a chat window can only do so much. Connect it to software and it becomes an agent. Connect it to business systems and it becomes a worker. Connect it to scientific databases and it becomes a research assistant. Connect it to laboratory automation and something else begins to emerge.
Not an autonomous scientist. Not AGI. But an AI system capable of participating in the cycle of scientific discovery rather than merely talking about it.
That is a meaningful threshold.
Because the future impact of AI won't be determined solely by how intelligent the models become. It will also be determined by what we allow them to touch.
Anthropic's biology lab gives us an early glimpse of what happens when the answer begins to include the physical world.