!Discover over 1,000 fresh articles every day

Get all the latest

نحن لا نرسل البريد العشوائي! اقرأ سياسة الخصوصية الخاصة بنا لمزيد من المعلومات.

Everyone is Replaceable – Gathering Several Chat GPT Models to Learn Chemistry

Even with the rapid advancements in artificial intelligence, AI is not close to being ready to replace humans in the field of science. However, this does not mean it cannot help alleviate some of the monotony from the daily routine of scientific experiments. For instance, researchers a few years ago placed an AI system in control of automated laboratory equipment and taught it to classify all the interactions that could occur between a set of raw materials.

Three AI Systems

The researchers noted that they were interested in understanding the capabilities that large language models (LLMs) could offer to assist in scientific work. Therefore, all the AI systems used in this work are LLMs, specifically GPT-3.5 and GPT-4, although some other models like Claude 1.3 and Falcon-40B-Instruct were also tested. Instead of using a single system to handle all aspects of chemistry, the researchers set up separate models to collaborate within a division of labor and called them “Coscientist.”

Three Artificial Intelligence Systems

The researchers used three different systems, which are:

Web Search Engine

This system features two main capabilities. The first is using the Google Search API to locate pages that may be worthy of absorbing the information contained within. The second is digesting those pages and extracting information from them – think of it as similar to the context of previous parts of a conversation that Chat GPT can retain to inform its answers later. Researchers could track the sites where this unit spent time, and half of the places it visited were Wikipedia pages. The top five sites it visited included journals published by the American Chemical Society and the Royal Society of Chemistry.

Document Search Engine

Think of this as an RTFM model. The AI was supposed to be granted control over various operational equipment in the lab, such as robotic liquid handlers and the like, which are often controlled via specialized commands or something like a Python API. This model was given access to all the manuals for this equipment, allowing it to learn how to control them.

The Planner

The planner can issue commands to both of the other AI models and process their responses. It has access to a Python development environment to execute the code, allowing it to perform calculations. It also has access to the automated laboratory equipment, enabling it to conduct and analyze experiments physically. You can think of the planner as part of a system that should act like a chemist, learning from the literature and attempting to use the equipment to carry out what it has learned.

The planner can also identify programming errors (whether in its Python scripts or in its attempts to control the automated devices), allowing it to correct its mistakes.

System Deployment

Initially, the system was tasked with synthesizing a number of chemical compounds such as acetaminophen and ibuprofen, confirming that it could generally identify a feasible synthesis after researching the web and the scientific literature. Thus, the question was whether the system was capable of identifying the equipment it had well enough to put its conceptual ability to work.

To start with something simple, the researchers used a standard sample plate, which contains a set of small wells arranged in a rectangular grid. The system was asked to fill in the squares with the diagonal lines or other patterns using different colored liquids and was able to do so effectively.

Next, they placed three different colored solutions at random locations in the grid of wells; the system was asked to identify which wells corresponded to each color. On its own, Coscientist did not know how to do this. However, when reminded that different colors would show different absorption spectra, it used the spectrometer to which it had access and was able to identify the different colors.

With…

The presence of commands and basic control appears to be working, researchers decided to test some chemical interactions. They provided a sample plate with wells filled with simple chemicals and catalysts, and asked it to perform a specific chemical reaction. The Coscientist successfully achieved the chemical reaction from the start, but its attempts to run the synthesis failed because it sent an invalid command to the devices that heat and stir the reactions. This returned the system to the documentation unit, allowing it to correct the issue and run the reactions.

The spectral signatures of the desired products were successfully present in the reaction mixture, and their existence was confirmed by chromatography.

System Improvement

After the basic reactions started working, researchers asked the system to optimize the reaction efficiency – they presented the optimization process as a game where points increase with higher chemical yields from the reaction.

The system made some poor calculations in the first round of test reactions but quickly focused on better yields. Researchers also found that they could avoid poor choices in the first round by providing the Coscientist with information about the yields produced by a set of random primitive mixtures. This means it doesn’t matter where the Coscientist gets its information – whether from the reactions it performs or from an external information source – it is capable of integrating the information into its planning.

Researchers concluded that the Coscientist has several notable capabilities:

  • Chemical synthesis planning using general information
  • Browsing and processing technical manuals for complex devices
  • Using that knowledge to control a range of laboratory equipment
  • Integrating equipment handling capabilities into the laboratory workflow
  • Analyzing its own reactions and using that information to design improved reaction conditions.

In many ways, this seems similar to an experience a first-year graduate student might go through. At best, the student will progress academically from there. But perhaps GPT-5 will be capable of that as well.

More seriously, the structure of the Coscientist, which relies on the interaction of a number of specialized systems, is similar to how brains work. It is clear that the specialized systems in the brain are capable of a wider range of activities, and there are more of them. But this type of structure may be crucial for enabling more complex behavior.

However, researchers themselves are concerned about some of the Coscientist’s capabilities. There are many chemicals (think of things like nerve gases) that we do not want to see become easier to synthesize. And it has become a constant challenge to figure out how to tell GPT models not to do something.

Source: Nature, 2023. DOI: 10.1038/s41586-023-06792-0 (about DOIs).

About the Author

John Timmer is the science editor at Ars Technica. He holds a Bachelor’s degree in biochemistry from Columbia University and a Ph.D. in molecular and cellular biology from the University of California, Berkeley. When he steps away from the keyboard, he tends to seek out a bicycle or a natural spot to reconnect with his walking shoes.

Source: https://arstechnica.com/science/2023/12/large-language-models-can-figure-out-how-to-do-chemistry/


Comments

Leave a Reply

Your email address will not be published. Required fields are marked *