The infrastructure vs. task mismatch of using ChatGPT for biomedical research is like using NASA's space shuttle for intrastate travel (length: 208 words)
View article
View summary
Sometimes in debates about the usefulness of "AI" technology, people note recent examples in the news of AI being used to do amazing things, like improving cancer detection. However, there's a really important distinction between AI / machine learning tools, even tools including generative AI, and the big LLMs like ChatGPT. You don't need GPT tools to detect cancer.
LLMs are a machine learning framework that can be applied in different contexts. GPT (and friends) are LLMs that are trained on vast amounts of internet social data. An LLM could also be very effective for biomedical uses, but in those circumstances you want to train it on biomedical data, not on social data.
Currently, I think a lot of researchers (I would hope not biomedical researchers, but I have reason to suspect otherwise) are reaching for GPT as a magical tool that can be used to analyze data without needing to be programmed by a trained software developer. Even if this were effective (which it won't reliably be), the infrastructure vs. task mismatch is reminiscent to me of an analogy my uncle once used in another context: it's like using NASA's space shuttle to travel from Dayton to Columbus.
(cross-posted from the Mastodon:
https://social.coop/@dynamic/113781978748003869; boosts welcome)
