IBM and NASA build language models to make scientific knowledge more accessible
IBM and NASA build language models to make scientific knowledge more accessible
At IBM Research, we’re inventing what’s next in AI, quantum computing, and hybrid cloud to shape the world ahead.
Yes please, more scientific papers and non-fiction, and less internet comments.
Cool!
Were the research papers used to train Ibm's language model not public? If the papers are public and there are no restrictions, Chatgpt and Gemini can use those right? If Chatgpt and Gemini used those, what's the benefit of using Ibm's language model?
Odd that heliophysics (sun physics) was specified. It's under astrophysics.
Let me mansplain this.
There are 2 main ways to alter the way llms like chatgpt respond: 1) training & fine-tuning and 2) prompting.
Prompting means feeding the LLM information in the form of text, so it has actual stuff in what you could call its "short-term memory". That can be done interactively and only allows you to feed it a limited number of words.
Training and tuning are much slower processes and are used to build up and refine the "long-term" memory and to give importance and detail to specific topics, or to reinforce some kind of trait, like tone or type of answers it can produce.
I'm pretty sure what ibm and nasa did was a round of fine-tuning with material in that area of knowledge, not just prompting.
Meh. I'll wait for the papers.
The good news is you could use the RAG model to get the papers related to your search.
So, just like a model to generate pics gives me characters with 11 fingers on their hands, and a model to generate something gives me POV-Ray code to draw a cylinder when asked for a Moebius stripe, this thing is going to tell us of many beautiful worlds with Pi ranging between 2 and 33?
In a non-Euclidean space you can have values of pi that aren't even constant. The first that came to mind was the surface of a sphere. This is actually an interesting problem.
Yes, bad choice of words, the point is that the model is going to generate loads of garbage imitating scientific text, as if real humans making errors (or sometimes intentionally putting out dubious works, fabricating experimental data etc) were not enough.
Humm interesting..