I don't hold that view exactly. But something related...
I once tried to replicate a bioinformatics result based on published data (for a class). I found that although the process did indeed yield plots A and B, as the authors claimed, they were typeset wrong in the PDF so plot A had B's caption and plot B had A's caption.
It would be an easy thing to provide assurances against, if you wanted to. You could repeatably build the pdf so that such a mistake was in plain view, as a bug in the pipeline, rather than something you had to do offline calculations to support or reject.
The situation as it is is not ideal. Instead of anything that would verify either side, it's my word against the author's until a third party bothers to repeat the analysis. That's the best we can do for scientific claims, but there are friendlier ways to make the computational claims verifiable.
The Claude science video showed a little "provenance" button and talked about exactly this. Life sciences have their hands full with the actual science. They're not immature, but they are not in a great position to be proving the validity of the computational connective tissue that underlies their results. That's a whole thing on its own, independent of the underlying scientific reasoning being presented (though I wouldn't call it data science).
Plus, its exactly the sort of thing we need AI to get better at: sourcing evidence that proves its claims and stitching it together so the proof is easily verifiable.
I don't hold that view exactly. But something related...
I once tried to replicate a bioinformatics result based on published data (for a class). I found that although the process did indeed yield plots A and B, as the authors claimed, they were typeset wrong in the PDF so plot A had B's caption and plot B had A's caption.
It would be an easy thing to provide assurances against, if you wanted to. You could repeatably build the pdf so that such a mistake was in plain view, as a bug in the pipeline, rather than something you had to do offline calculations to support or reject.
The situation as it is is not ideal. Instead of anything that would verify either side, it's my word against the author's until a third party bothers to repeat the analysis. That's the best we can do for scientific claims, but there are friendlier ways to make the computational claims verifiable.
The Claude science video showed a little "provenance" button and talked about exactly this. Life sciences have their hands full with the actual science. They're not immature, but they are not in a great position to be proving the validity of the computational connective tissue that underlies their results. That's a whole thing on its own, independent of the underlying scientific reasoning being presented (though I wouldn't call it data science).
Plus, its exactly the sort of thing we need AI to get better at: sourcing evidence that proves its claims and stitching it together so the proof is easily verifiable.
I too am excited.