In a world rapidly adapting to artificial intelligence (AI), large language models (LLMs) like ChatGPT and DALL-E are changing the landscape of creativity and communication. Instead of relying on code, users can now interact with these “chatbot” models by asking questions in plain language and receiving instant responses. Yet, while LLMs have already altered how people communicate, experts believe we’ve only just begun to uncover their potential. A recent post from Argonne National Laboratory breaks down the potential of LLMs in science, which we recap here:
Credit: (Image by Argonne National Laboratory.)
The Argonne Leadership Computing Facility is home to the AI Testbed. Researchers can access platforms, including this Cerebras CS-2, to study how machine learning-based high-performance computing applications will interface with today’s supercomputers to accelerate scientific discovery.
Argonne clarifies core functionality of LLMs, which we summarize here:
What Exactly is an LLM?
Most people have used LLMs, but understanding how they actually work is another matter. While these models are trained on vast amounts of data to generate human-like responses, explaining their functionality can be challenging. As Argonne scientists point out, the real test is whether you can explain LLMs to someone else in under a minute – whether at work or at a cocktail party.
Using LLMs in Science
Training LLMs for scientific research requires carefully selected data tailored to specific goals. While you could train a model on everything from Taylor Swift lyrics to the entire internet, such broad datasets wouldn’t advance scientific discovery. As Argonne researchers explain, effective scientific LLMs need focused training on relevant scientific literature and data. This targeted approach is essential for tackling specialized research challenges.
The Art of Prompt Engineering
Even with vast amounts of information at their disposal, getting useful results from LLMs requires skill and strategy. Scientists are developing systematic approaches to formulate questions that elicit meaningful responses – a process called prompt engineering. As Argonne researchers note, this is far more nuanced than typical web searches or voice commands.
Are LLM Answers Trustworthy?
While LLMs are often compared to eager assistants, they don’t always provide accurate or reliable information. They’re designed to give quick responses, but that doesn’t mean they’re always correct. Researchers are working to understand how to balance the power of AI with its limitations and to teach scientists how to interpret these answers critically.
The Future of LLMs in Science
Like the automobile, internet, and cell phone before them, LLMs may fundamentally reshape how we conduct research – potentially changing rules we didn’t even know we were playing by. As Argonne researchers acknowledge, while we can’t yet predict their full impact on scientific discovery, significant efforts are underway across national laboratories to understand and harness this transformative potential.
