>Claude is starting to meaningfully assist chemists with the daily translation, recall, and integration work that complements their judgment, and we plan to keep extending its helpfulness.
This is a real worthwhile goal.
>Understanding what molecule a chemist is working with is critical.
However there are so many tonnes of chemicals where there is not very deep understanding, it would be good if AI could help with that. In the meantime it remains essential to carry on critical operations in environments where there is little to no understanding at all. Without having things go from risky to more than can be handled.
Ask AI how to progress without understanding. Well when you think about it maybe that's what a lot of people say AI can only do.
One thing to think about also is that natural science has its own "corpus" of literature as a subset of all literary publications. There's a particular literary style, but most sciences are not based on language. Not even numbers, as important as they are.
The much larger mass of literature, for those realms that actually are based on language, gets so much coverage by publication, that's enough to be considered a true corpus of accumulated knowledge. By comparison anyway. And a couple of these are some of the most rewarding things like artistic literature whether fiction or nonfiction, as well as computer languages. Along with computer languages comes lots of advanced computer science too, also some of the most brilliant mathematical minds.
With natural science though, the entire world's publications do not even give significant coverage, not even the full 1%.
With a little back-of-the-napkin math, the odds are probably about 80:1 that the same type language model will be as impressive as it is on language-based efforts.
Nice paper though, I've been looking forward to AI help for NMR spectra since the 1970's. Turns out I never needed NMR professionally but there are plenty of other spectra and this approach can leverage the pattern recognition that AI is somewhat based on, along with hopefully more comprehensive comparison to reference spectra that have been published. Using data centers more massive than some chemical plants, things really can be accomplished that were not within reach before.
If you can call that within reach any better than it was, if it's too expensive, oh well.
Ideally the progress would best be pushed forward until the NMR assistance seems about perfect, then "distilled" into a regular open-source desktop program which does the job without further internet contact. You would want the option to have Claude go back online at times in the future if you wanted to search for newly published advances, mainly spectra. The main obstacle since before computerization has always been the licensing of reference spectra, and computerization piled so much more licensing on, otherwise I would say NMR would have a lot wider use outside of research by now.
Alternatively if you look at it the other way, it could be good to conceptually make a particular chemist into a "Claude" of its own, along with making Claude into a more generally useful chemist or assistant.
This is a real worthwhile goal.
>Understanding what molecule a chemist is working with is critical.
However there are so many tonnes of chemicals where there is not very deep understanding, it would be good if AI could help with that. In the meantime it remains essential to carry on critical operations in environments where there is little to no understanding at all. Without having things go from risky to more than can be handled.
Ask AI how to progress without understanding. Well when you think about it maybe that's what a lot of people say AI can only do.
One thing to think about also is that natural science has its own "corpus" of literature as a subset of all literary publications. There's a particular literary style, but most sciences are not based on language. Not even numbers, as important as they are.
The much larger mass of literature, for those realms that actually are based on language, gets so much coverage by publication, that's enough to be considered a true corpus of accumulated knowledge. By comparison anyway. And a couple of these are some of the most rewarding things like artistic literature whether fiction or nonfiction, as well as computer languages. Along with computer languages comes lots of advanced computer science too, also some of the most brilliant mathematical minds.
With natural science though, the entire world's publications do not even give significant coverage, not even the full 1%.
With a little back-of-the-napkin math, the odds are probably about 80:1 that the same type language model will be as impressive as it is on language-based efforts.
Nice paper though, I've been looking forward to AI help for NMR spectra since the 1970's. Turns out I never needed NMR professionally but there are plenty of other spectra and this approach can leverage the pattern recognition that AI is somewhat based on, along with hopefully more comprehensive comparison to reference spectra that have been published. Using data centers more massive than some chemical plants, things really can be accomplished that were not within reach before.
If you can call that within reach any better than it was, if it's too expensive, oh well.
Ideally the progress would best be pushed forward until the NMR assistance seems about perfect, then "distilled" into a regular open-source desktop program which does the job without further internet contact. You would want the option to have Claude go back online at times in the future if you wanted to search for newly published advances, mainly spectra. The main obstacle since before computerization has always been the licensing of reference spectra, and computerization piled so much more licensing on, otherwise I would say NMR would have a lot wider use outside of research by now.
Alternatively if you look at it the other way, it could be good to conceptually make a particular chemist into a "Claude" of its own, along with making Claude into a more generally useful chemist or assistant.