Adventures in Using LLMs

This past week I’ve been using the three major LLMs available to the general public: ChatGPT, Microsoft Co-Pilot, and Google Gemini. I’m not writing this to compare the three, but to report out how a few things I’ve learned and how I’ve adopted these tools to help me in one of my hobbies.

Image Generation

The DALL-E model with ChatGPT 4 was used to create images of characters based upon a novel I’m writing. The results are interesting. When modifications are requested, it produces a completely new image, it doesn’t really go back to the one it just created for you.

The depiction of people in this model is good, much better than the earlier model I’d played with maybe two years ago. With some more work, I think I could generate the novel’s book cover!

Spreadsheet Analysis

In speaking to colleagues this week about using LLMs at work, I wanted to see how easy it would be to analyze a spreadsheet of data. I fed the LLM ChatGPT 4 an Excel file of my music collection. It included the album titles, the performers, the genres, and the bit rate of the digital files. I created it with Roon.

It was easy to ask questions, like, “Which artist is most popular in my collection?” to which it said “Ton Koopman.” I have a lot of Ton Koopman CDs, but I can say with authority the answer was not right.

The reason the AI mis-reported it was because the data wasn’t as clean as it could be. “Ton Koopman” by itself and “Amsterdam Baroque Orchestra / Koopman” isn’t the same to the AI. So my success would be limited.

We had an issue when I asked for the “top 5 genres.” It correctly reported this, at least in its misunderstanding of the genres and how they’re listed. One CD may have multiple genres applied: Opera, Baroque. That’s two applied to one recording. After we sorted that out, it returned the right answer.

It stunk at writing a macro

Then I asked it to create a button in Excel and apply to it a Macro that would report my “top 5 genres.” It took so much work to get it to write this that my access to the GPT 4 model was curtailed; despite paying $20 to use it, I had to wait several hours, I was told.

Upon coming back to the problem, the model returned, finally, an answer. But Excel told me my top 5 genres were the one listed first in the spreadsheet. But ChatGPT knew I meant the five most frequently-appearing. I told it knows this and to apply that logic to the Excel macro.

It finally worked. But it took a lot of time. I had previous experience with ChatGPT 3 trying to write Google Sheet “macros.” I finally gave up because it didn’t have the right information about Google’s new frameworks since it’s knowledge base was too old.

Writing and Language Is Its Forte

Separate from this website, I’ve been publishing reviews of classical, baroque, and jazz CDs since the late 1990s. I loved reading reviews when I was in college in magazines like Gramophone or Early Music. It gave me something to listen for if I eventually procured those recordings, or at least checked them out from the music library’s collection.

I have good taste and have tried to apply that filter to my reviews for some time. I rarely take notes when listening, but I sometimes do. Writing the reviews can take a lot of time. I wondered what might happen if I employed the help of the AI?

I took bulleted notes of my listening sessions. I provided some background information, but a lot of my notes were details about what I heard. I was opinionated, I said what I liked and what I did not. I included the details I usually speak to, including sound quality, a little history, and my own connection to the music.

I wrote complex prompts telling the AI it was a CD reviewer. I included comments about what tone to write in. And I provided my notes. I had fun asking for a review to be re-written as if by David Sedaris. It picked up on some of his stylistic aspects without making terribly good sense, but it warned me that his style for a CD review might not be what I really want. (Ha!)

Here is one of the reviews that the AI wrote (Chat GPT 4). I also provided links to Wikipedia pages about two of the works and about the performer, which it did not appear to use, despite me asking specifically for it to do so. (The AI’s ability to access webpages is not always consistent, or sometimes it reports it cannot, despite it being a feature of GPT 4.)

Here were my notes:

My Notes

  • Album of Mozart piano works performed by Jérôme Hantaï on a period piano. Hantaï also plays the viola da gamba and is brother to harpsichordist Pierre Hantaï and flutist Marc Hantaï.
  • Album provides 73 minutes of music, including Mozart’s 7th and 13th piano sonatas. The booklet notes give a good historical perspective of Mozart’s life and the potential impact on the music recorded here.
  • According to Hantaï, The fortepiano used for this recording is an anonymous and undated instrument, but its careful workmanship is very clearly in keeping with this emerging tradition in southern Germany. It has benefited from an exemplary restoration in Christopher Clarke’s workshop.
  • The overall sound of this instrument is “hard” and not at all delicate, as seems often the case with recordings of pianoforte music. While Hantaï can play softer and more delicately, where needed, his overall approach and dynamic is on the louder side of dynamics. When he hits his loudest notes the piano responds with an almost singing tone which is one of its nice qualities.
  • The overall sound of the album is well-captured. The producer and recording engineer is Jiri Heger and the recording was made at Théâtre élisabéthain d’Hardelot in February 2020.
  • Hantaï’s touch on this instrument is firm and it comes across in the opening Allegro of the Sonata K. 333. Mozart’s intricate melody is rendered clearly, the rhythm of so many notes perfectly articulated which should impress listeners with the technical requirements needed to let the piano sing so well. I admired Hantaï’s second movement which didn’t linger in tempo for too long; I think his tempo and the use of rubato effects do well in the middle slow movement. Renditions on a modern grand my adopt a softer approach, but here the more limited range of dynamics available on the fortepiano are a contrast from what you may be used to hearing.
  • The same sonata’s third movement with its recurring theme is filled with tongue-in-cheek moments that Hantaï manages not to over indulge us with, which in a recording is probably the right approach.
  • The Rondo in A minor, K. 511 may be the most well-known work on the recording, which is its last track. I felt the melody (along with the left hand) are a masterclass in how to articulate on an early piano. Longwinded phrasing is left behind, instead to only slur with shorter groups of notes; in the contrasting major section his evenness in articulation with a longer string of notes provides an interesting contrast. Later in the Rondo I’m reminded of a groundswell of emotion that emerges as I wonder if Mozart’s piano was capable of fully realizing his beautifully crafted melodies and the colorful places he takes us harmonically. The intensity I think Hantaï does well to capture, pushing this instrument to its extreme in the few places demand it.
  • The opening sonata in C K. 309 is also well-known, I am guessing, for fans of Mozart’s piano sonatas. Hantaï doesn’t shy from emphasizing the hammer action of the piano. For me, his interpretation is very fresh. I shouldn’t keep picturing scenes of Tom Hulce from Amadeus but it is dearly impossible to not hear a gutsy, comedic, and brilliant young pianist behind the genesis of this sonata’s opening movement. Hantaï does an excellent job of treating us to dynamic contrasts in the development section, where Mozart presents the theme in a minor mode.
  • The theme in the third movement of the sonata K. 309 almost has too many notes in the theme; on repeating the theme Mozart adds embellishments in the form of ornamental turns which give the extra, repeated note in the melody its reason for being there. It’s all a ploy, perhaps, for us to recognize Mozart’s genius. Later with a repeated figure in the right hand, Hantaï makes the instrument ring-out in spectacular fashion.
  • Conclusion: Hantaï is a brilliant interpreter of Mozart, going by what’s presented in this album. His ability to bring out significant dynamic contrasts from the instrument to me speak to the novelty of playing on an instrument that, well, offers dynamic contrast. It was put to use by Mozart to delight us and in this recording, Hantaï has obliged the composer and delighted me in turn. While I do not enjoy Mozart’s keyboard music to the same degree as say Bach, Rameau, or Handel from the Baroque era, this performance in particular helped me hear well all the small details that give Mozart’s music charm beyond his endearing melodies. This album I am hoping will be of special interest to any students of Mozart and what things are possible, even if they do not have access to the extraordinary instrument used on this recording.

Background Information

LLMs can be a value-add for creating content

I don’t enjoy listening to music with a laptop on my lap, but I felt I was able to publish a review quicker than using my typical regimen. In this case, the AI did not understand all my notes which required me to make small edits to the copy.

However I did save time and I am not unhappy with the results. My nxt step will be to create a custom GPT that is trained to write more like me. (In this case, I did not provide my own writing samples, but asked it to be philosophical in style, whatever it thought that meant.) And, yes, I applied my 4.5 star rating, I did not ask the AI to guess the rating.

For kicks, I asked for a rating. Now I’m impressed.

Given your detailed and generally positive notes about Jérôme Hantaï’s performance on the period piano, it seems that you appreciate the album’s historical authenticity, sound quality, and the technical prowess displayed by Hantaï. You highlighted the effective use of the fortepiano’s dynamic capabilities and the fresh interpretations of well-known compositions, as well as your admiration for the nuanced articulation in pieces like the Rondo in A minor, K. 511.

However, you also noted the “hard” sound of the fortepiano and the departure from the softer sound profiles typically expected in such recordings, which might be a slight drawback depending on listener preferences. Yet, there’s no indication that this significantly detracted from your overall enjoyment of the album.

Considering these factors, and the lack of significant negative comments, it sounds like you’d likely rate this album quite highly on a 5-point scale. A reasonable estimate might be around 4.5, indicating it’s nearly a “must-buy,” especially for those interested in historical performances or the nuances of Mozart played on period instruments. This rating reflects your appreciation of the album’s strengths and its few minor caveats.

Copy from Chat GPT 4