This piece was originally published at Second Best.
The incredible success of Large Language Models like ChatGPT is both a scientific breakthrough and a boon for future scientific discovery. As a recent editorial in Nature explains,
…large language and vision models that can digest the literature will be used to identify gaps in knowledge, help summarize and understand unfamiliar topics, and find the most relevant references, protocols, data and experts. They will also generate and explain complex graphs and schematics, and help write and edit routine computer code as well as scientific papers, reviews, grant applications, curriculum vitae and all sorts of reports. Producing content without assistance from machine-learning applications may soon be as rare as writing snail mail.
Saving researchers’ time to focus on what matters will speed-up the scientific process and help to extract new connections from the voluminous research that already exists. Longer term, however, we may even find ourselves asking the AIs to probe deep scientific questions for us. As Nat Friedman put it, paraphrasing OpenAI’s David Dohan, imagine one day prompting a next-generation model with “A well known formula for a room temperature superconductor is…” and fully expecting it to output an answer.
