Hi Jim, I personally do not like the acronym, AI, because the “I" part is not consistent with what I would consider intelligence. Nevertheless, it can be extremely helpful for your efforts of improving your software. Maybe it should be called, “AmI”, for amplify my intelligence? In a situation similar to yours, I have provided AI software with my code and asked if it can find ways for me to make the code more efficient. Also, if one discovers a bug or coding inefficiency, AI can be useful for discovering other places in a mountain of code where the same problem could occur. These are helpful uses, should be indirectly helpful for the users of the software, and are not unnecessarily memory-intense uses of AI (I think… at least compared to asking it to perform analyses). But I assume when you asked it to build an app for you, it is going to pick the lowest hanging fruit, at least initially, and it would take a lot of subsequent coaxing to optimize the app to work efficiently. In this case, I question if the juice is worth the squeeze. Asking AI to identify the places in your existing code that could be improved is a good use of a tool that augments your intelligence and would probably make your software better than an AI-generated app could. I could be wrong; maybe I just do not appreciate how great AI is.
Speaking of hallucinations, in the first few years of chatGPT, it would suggest using functions for our software that do not exist. Now any AI software seems to know the software landscape pretty well. But this makes me question the AI limits. It seems to perform established analyses well, but what if one wanted to develop a Procrustes alignment method, for example, that did not require estimating missing landmarks, but using rotation matrices calculated from only matching landmarks between a reference and the specimen? If one asks AI to do this, it will probably find a solution, right? What kind of hallucinations would the solution contain? If the person requesting this is not keen enough to recognize mathematical or statistical issues with the solution, could its haphazard use become published, encouraging others to do something similar? I’m not sure if my example is a good one but hopefully it illustrates my concern. I do not think much good can arise if software developers stop developing software because AI can make an app, on the spot. I think a lot of good can arise if software developers use AI to improve their software. Maybe I’ll see it differently in a few years though. Mike > On Mar 23, 2026, at 11:48 PM, '[email protected]' via Morphmet > <[email protected]> wrote: > > Since I haven’t been able to go out much these past two weeks (we have been > having very strong “Kona storms”), my mind started to wander and I wondered > whether to debug existing code or create new morphometric software. But then > I wondered whether conventional software is still needed in the age of AI. I > described below a couple of the experiments I made. It seems that there are > some alternatives that are worth discussing. > > Google AI Studio > > First, I tried Google AI Studio (https://aistudio.google.com/). I entered the > prompt “I need an app that performs an analysis of relative warps (geometric > morphometrics) and displays the relative warps plot and estimates of implied > shapes at user selected positions in the plot.” I was pleasantly surprised to > find that it knew the terms I used and it created a simple interactive app in > my browser. It already “knew” about geometric morphometrics. Thus, the > computations that were required and how to perform a GPA, a singular-value > decomposition, and compute a thin-plate spline. It could generate a plot of a > thin-plate spline showing the difference between the mean shape and that > corresponding to an arbitrary location in the space of the relative warps. > With a little experimentation I found that I needed to give it additional > details about the specific options I wanted such as the criterion to stop the > GPA iterations and that it should normalize the final configuration to have a > centroid size of 1. It also assumed alpha = 1 when computing relative warps, > so I had to add “using alpha = 0 when computing relative warps” (although I > could have just said perform a PCA” and not mentioned relative warps. I also > had to specify that it should use the same scales for the two relative warp > axes. It automatically called the new app “MorphoWarp” which sounded like a > reasonable name to me. I then had to specify what input file format I wanted > to use (its default JSON format looked pretty tedious to prepare). I > specified a simple text file format with each row corresponding to a specimen > and each containing alternating x and y coordinates. Once the specifications > were made, the app ran reasonably fast, though not as fast as a compiled > program such as tpsRelw. The app can be uploaded to github and shared with > others. I have not done this yet as I would like to make the app more > complete first. > > Claude > > I then wondered about using an interactive AI bot directly. I tried Claude > (https://claude.ai/chat). I gave it the prompt: “Perform an analysis of > relative warps on the provided data file. Display the relative contribution > of each relative warp. Show a scatterplot of relative warp 2 against relative > warp 1 with the axes using the same scale. Label the points numerically.” I > also it with the same simple data file as above but it began with the > following comment about the format: “The lines (specimens) contain > alternating x and y coordinates for 8 landmarks.” It then performed the > computations, but, as before, I found I needed to give more details such as > the GPA iteration stopping criterion, alpha = 0, and to normalize the final > configuration. I could also ask it to produce a thin-plate spline plot of the > difference between the average shape and a selected position in the relative > warps space. The plots it produced were quite clear and could be downloaded > as PNG files. The computations were, however, very slow. No app was produced > this way, so the only software to share is the prompt. > > Broader implications > > An alternative to downloading morphometric software may be to simply download > prompts for AI bots. For serious applications the prompts should, of course, > be more complete than my examples. The prompts should reduce ambiguity about > how computations should be performed. They should also include test data to > make sure the results are correct. Because these systems can “hallucinate,” > validation checks are essential. > > Another consideration is that using conventional software to perform the > computations on your own computer might be better for the environment than > increasing the need for even larger remote data centers that consume very > large amounts of power. This may also be worth discussing. > > It looks like the rain has stopped. I think I will go back outside for a > while. It’s hard to work too hard for too long when on Maui. > > Jim > > > F. James Rohlf > Distinguished Professor, Emeritus and Research Professor > Depts: Anthropology and Ecology & Evolution > Stony Brook University > > -- > You received this message because you are subscribed to the Google Groups > "Morphmet" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected] > <mailto:[email protected]>. > To view this discussion visit > https://groups.google.com/d/msgid/morphmet2/Mailbird-bbe66a46-76ca-4b7b-943c-07c973b058be%40stonybrook.edu > > <https://groups.google.com/d/msgid/morphmet2/Mailbird-bbe66a46-76ca-4b7b-943c-07c973b058be%40stonybrook.edu?utm_medium=email&utm_source=footer>. -- You received this message because you are subscribed to the Google Groups "Morphmet" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion visit https://groups.google.com/d/msgid/morphmet2/024B7059-1BA7-4EEB-96C3-C197B2E8F45F%40gmail.com.
