In spite of the rapid progress in artificial intelligence, the capabilities of AIs are still far from being sufficient to replace humans in the field of scientific research. However, they can still assist in automating certain tedious tasks in scientific experimentation, making the daily routine more efficient. For instance, a few years ago, researchers entrusted an AI with the responsibility of operating automated lab equipment and instructed it to meticulously document all the possible reactions that could take place among a specific group of initial substances.
Although it was helpful, a significant amount of researcher intervention was still necessary to initially train the system. A team at Carnegie Mellon University has recently discovered a method for enabling an AI system to independently learn the field of chemistry. The system needs a trio of AI instances, each with its own specialization for various operations. However, after the initial setup and provision of raw materials, all you need to do is specify the desired reaction, and the system will handle the rest.
An AI trinity
The researchers express their interest in exploring the potential contributions of large language models (LLMs) to the scientific field. All the AI systems utilized in this project are LLMs, primarily GPT-3.5 and GPT-4. However, a few others, such as Claude 1.3 and Falcon-40B-Instruct, were also subjected to testing. GPT-4 and Claude 1.3 were the top performers. However, instead of relying on a single system to manage all aspects of the chemistry, the researchers established separate instances to work together in a division of labor arrangement, which they named “Coscientist.”
They utilized three different systems:
Internet user. There are two primary functions for this. An option is to utilize Google’s search API to locate webpages that could potentially be valuable for gathering information. Another approach is to analyze and extract information from those pages. This is similar to how Chat GPT uses previous parts of a conversation to provide more informed answers later on. The researchers were able to monitor the module’s activity and found that it frequently visited various Wikipedia pages. Among the sites it visited, the top five included the journals published by the American Chemical Society and the Royal Society of Chemistry.
I am looking for documentation. Consider this the RTFM instance. The AI was set to be granted control over a range of lab automation equipment, such as robotic fluid handlers and similar devices, which are typically operated using specific commands or a Python API. This AI instance was granted access to all the manuals for this equipment, enabling it to learn how to operate it.
Organizer. The planner has the ability to give commands to the other two AI instances and handle their responses. The program has access to a Python sandbox, which enables it to execute code and perform calculations. In addition, it has the capability to utilize the automated lab equipment, enabling it to actively conduct and evaluate experiments. Think of the planner as a chemist, gathering knowledge from the literature and using equipment to put that knowledge into action.
The planner is also capable of detecting software errors (either in its Python scripts or in its attempts to control the automated hardware), enabling it to rectify any mistakes.
Putting the System to Use
At first, the system was tasked with synthesizing various chemicals, like acetaminophen and ibuprofen. It successfully demonstrated its ability to find workable synthesis methods by searching through the web and scientific literature. Therefore, the question arises as to whether the system possesses sufficient knowledge of the hardware it has access to in order to effectively utilize its conceptual abilities.
Beginning with a straightforward approach, the researchers utilized a conventional sample plate containing multiple small wells arranged in a rectangular grid. The system successfully filled in squares, diagonal stripes, and other patterns using different colored liquids.
Continuing from that, three solutions of different colors were placed randomly in the grid of wells. The system was then tasked with identifying the colors of the wells. By itself, the coscientist was unsure of how to accomplish this. However, upon receiving a prompt that served as a reminder about the distinct absorption spectra associated with various colors, the entity utilized a spectrograph at its disposal and successfully discerned the differences between the colors.
After ensuring that the basic command and control were working properly, the researchers decided to experiment with some chemistry. A sample plate was given, containing wells filled with various chemicals and catalysts. The task was to perform a specific chemical reaction. The coscientist had a good grasp of the chemistry from the beginning, but unfortunately, its efforts to carry out the synthesis were unsuccessful due to an invalid command being sent to the hardware responsible for heating and stirring the reactions. It was sent back to the documentation module, which enabled it to address the issue and execute reactions.
And it was successful. The reaction mixture contained spectral signatures of the desired products, which were confirmed through chromatography.
Optimization
After successfully implementing basic reactions, the researchers decided to challenge the system to enhance the efficiency of the reaction. They framed this optimization process as a game where the score would increase based on the reaction’s yield.
The system initially made some incorrect guesses during the first round of test reactions, but it swiftly improved and achieved more favorable yields. In addition, the researchers discovered a way to prevent making poor choices in the initial round. They achieved this by equipping Coscientist with data on the yields produced by a few randomly selected starting mixtures. It is evident that Coscientist can seamlessly integrate information into its planning, regardless of its source.
- The researchers have concluded that coscientists possess several remarkable capabilities.
- Utilizing publicly available information to strategize chemical synthesis
- Understanding and working with technical manuals for complex hardware
- Utilizing that knowledge to manipulate various laboratory equipment
- Incorporating these hardware-handling capabilities into a lab workflow
- Examining its own responses and utilizing that data to develop enhanced conditions for reactions.
This seems similar to what a student might go through during their initial year of graduate school. It would be ideal if the graduate student could make progress beyond that. Perhaps GPT-5 will possess the capability to do so too.
On a more serious note, the structure of Coscientist is quite fascinating, as it resembles the way brains function. It relies on the interaction of various specialized systems. Clearly, the brain’s specialized systems have a significantly broader range of activities, and there are many more of them. However, it is possible that this type of structure plays a crucial role in facilitating more complex behavior.
However, the researchers themselves have expressed concerns regarding certain capabilities of coscientists. There are numerous chemicals, such as nerve gases, that we strongly oppose making more accessible to synthesize. Finding a way to communicate with GPT instances about what they should not do has proven to be a persistent and ongoing challenge.