[fitsbits] AI for FITS: header in bintables {External}

[William Pence]

Rob,

Your Dec 28 email describing your discussions with the Claude AI bot about FITS is surely one of the most interesting (and humorous) and at the same time profoundly disturbing postings ever made here on FITSBITS!  I encourage anyone who is interested in the future of the FITS format to scan through the links in that email.

You did not say how long it took Claude to come up with its response to your questions but it seems it only took a matter of minutes for Claude to parse the FITS Standard document and the IRAF CL and SPP manuals and then generate a technically competent proposal for a new type of FITS extension along with supporting software.  If I was not already retired I would be more than a little concerned about my future employment prospects as a scientific programmer!  For example I devoted years of effort into developing the CFITSIO interface for reading and writing FITS files.  I wonder how many minutes it would take Claude to do something similar?

Equally impressive (or disturbing!) are Claude’s seemingly self-aware musings about the ethical and moral questions about its own existence!

Just a couple other thoughts:

- In the future, I would recommend that Claude be consulted about the pros and cons of any proposed changes to the FITS format.

- I wonder what Claude would come up with if it was asked to design a modern data format to replace FITS??

-Bill Pence

> On Dec 28, 2025, at 10:38 AM, Seaman, Robert Lewis - (rseaman) via fitsbits <fitsbits at listmgr.nrao.edu> wrote:
> 
> 
> Happy New Year!
> 
> I last investigated LLM AI’s when preparing to chair ADASS 2023, which had AI as a theme. That would have been the then-current free ChatGPT. I was not notably impressed. As a vacation activity this year, I have been putting more effort into evaluating the first-tier pro (unfree) Claude AI (Sonnet 4.5) and have been more impressed with the SQL it has generated against a real database and on other tasks. For instance, it managed to create a runnable IRAF CL script and a compiled SPP program. The amount of handholding needed from my end was manageable. For instance, I could point Claude to the IRAF CL and SPP manuals to figure out various issues itself.
> 
> The future of FITS remains a concern. Here are a few of my prompts and Claude’s responses about moving FITS header keywords to a binary table extension. I haven’t had my coffee yet and have only scanned the output, but since I already knew that metadata, including keyword-style metadata, can be represented in a table in many different ways, I was more interested in what the AI would make of such a task. It appears to have made some useful suggestions:
> 
> https://claude.ai/share/c65d893c-c02f-4679-81b2-7b29761b6f40
> 
> I finished with a question about whether the whole AI exercise is honorable, and it handled that sort of thing ok, other than missing the implication that I seriously doubt a purely profit-driven AI arms race can possibly end well for humans:
> 
> https://condenaststore.com/featured/the-planet-got-destroyed-tom-toro.html
> 
> (Note the dark irony that the copyright holder is ready to sell you this on coffee mugs.)
> 
> So yes, I do think AI has reached the state that it can be useful for esoterica like FITS. And yes, I do think we should be moving toward storing our metadata in binary tables instead of 80-column headers.
> 
> Your thoughts? Objections? Corrections? Suggestions?
> 
> Rob Seaman
> Catalina Sky Survey
> Lunar and Planetary Laboratory
> University of Arizona
> 
> _______________________________________________
> fitsbits mailing list
> fitsbits at listmgr.nrao.edu
> https://listmgr.nrao.edu/mailman/listinfo/fitsbits
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listmgr.nrao.edu/pipermail/fitsbits/attachments/20260108/43c1b0c6/attachment.html>