ChatGPT becomes phenomenal by the end of 2022 within the artificial intelligence community, and its influence is still growing. I have been fascinated by the idea of using GPT-3 to generate conversations, and I have been trying to make some experiments with it. In this post, I will share some of my experiments.

In fact, the paragraph above was largely generated by GitHub Copilot, another service based on the language model.

Music composition

It seems that OpenAI is rather reluctant to let users go too far. While I was experimenting on it, they blocked the function of generating music. However, as with many other restrictions to ChatGPT, there is always circumvention.

Before the restriction was implemented, I simply asked the AI to write down The Internationale in lilypond style. And I used Hackily, an online lilypond renderer, to visualize the score and generate the midi playback.

As you can see from the inserted video below, this is what I have got from it. From a human point of view, it is by no means The Internationale. No wonder they blocked it! It has the title and the composer correct, but the content reminds me of the minimalism music.

I then proceeded to compose something in “serialism”, but before I saved it, the service crashed and the “serialism” music I have got could never be retrieved due to the newly implemented restrictions.

So the workaround is like this: you first give it a lilypond code, and then you are still able to ask the model to revise it according to your instructions, including adding excerpts from other music pieces. In the next experiment, I asked the model to combine a small motif of “a Baroque-style sonata” with various pieces such as O Tennenbaum and Amazing Grace (which, apparently, bear no resemblance to the actual songs), making symmetric reflections of the melody, transposing the tune, and composing the corresponding harmony (the lower staff). Some other instructions were given, and minor adjustments were manually made.

The final piece was too messy to be directly presented. So I exported the midi file, and imported the midi file to Musescore. This is what I have got.

From these experiments, it could be summarized that, as of Dec 12, 2022:

  • The AI is capable of generating small motifs of music, but it actually does not have the sense of aesthetics. It is prone to generate very short sequences, with special favor to the ascending scale. When it is asked to generate longer sequences, it always repeats a certain pattern. Longer composition is impossible without human intervention.
  • It knows the jargon in music theory, and it can explain it well. But it fails when asked to apply them. For example, it can explain the range for violin, but it fails to rewrite a melody to where all notes fall in this range.
  • It is also very hard for the model to generate music with correct number of beats in each bar.
  • Sometimes, the model may misuse or even “invent” lilypond syntax. However, it can often correct the errors when prompted by the human.

Web page

I also asked the AI to generate a webpage using HTML and JavaScript, about a personal website for an AI researcher. I asked it to adopt the W3.CSS style and use the grid layout. I also asked it to insert a photo, add Education, Working Experience, and Publication sections.

The webpage it generated is rather appealing, and you can find it here.

The webpage

Other miscellaneous entertainment

  • I asked it to give diagnosis according to patient’s complaint and clinical manifestations.

ME: A patient has an acute pain in the lower right area of the abdomen, and the body temperature is 38 degrees Celsius. What tests shall the patient take?

ChatGPT: It is difficult to determine which tests a patient should take without more information. It is important for the patient to discuss their symptoms with a healthcare provider, who can assess the patient’s condition and recommend appropriate tests. Some possible tests that may be recommended in this situation include a physical exam, imaging tests such as an ultrasound or CT scan, and blood tests to check for infection or other conditions. It is always best to follow the advice of a healthcare provider when it comes to medical treatment and testing.

ME: The ultrasound shows swollen appendix, which was confirmed by the CT scan. The neutrophil as well as the CRP are elevated. What might be causing the symptom?

ChatGPT: Based on the information provided, it is possible that the patient is experiencing appendicitis, which is a condition in which the appendix becomes inflamed and swollen. The symptoms of appendicitis often include pain in the lower right side of the abdomen, fever, and elevated levels of white blood cells and CRP (C-reactive protein) in the blood. lf the ultrasound and CT scan confirm that the patient has a swollen appendix, it is likely that the cause of their symptoms is appendicitis. However, it is important to note that there may be other potential causes ofthese symptomss, and a healthcare provider would need to evaluate the patientand conductt additional testing to make a definitive diagnosis.

  • I asked it to be the simulated patient. But it refused to evaluate my performance as a doctor.
  • I asked it to create fictional grant applications.

Title: Development of Algebraic Reconstruction Technique for Total Body PET lmaging and its Application in Accelerating Dynamic PET Reconstructions

Abstract: Positron emission tomography (PET) is a valuable tool for the diagnosis and treatment of various diseases, such as cancer and neurological disorders. However, current PET imaging techniques have limitations, such as low spatial resolution and long acquisition times. This can lead to suboptimal image quality and reduced diagnostic accuracy. In this project, we propose to develop a novel algebraic reconstruction technique (ART) for total body PET imaging that overcomes these limitations. This algorithm will incorporate prior knowledge about the distribution of radioactivity in the body, allowing for improved spatial resolution and faster acquisition times. We will also investigate the application of this algorithm in accelerating dynamic PET reconstructions.

Research Plan:

  • Develop and validate the ART algorithm using simulations and phantom studies
  • Conduct a clinical study to assess the performance of the algorithm in real-world settings
  • Refine and optimize the algorithm based on the results of the clinical study

Budget: [Provide detailed budget and justification for each expense]\

Timeline: [Provide a timeline for completing the various project milestones and tasks]

Conclusion: We believe that the development of this novel ART algorithm has the potential to significantly advance the field of PET imaging and improve patient care. We are confident that our team has the necessary skills and experience to successfully complete this project and we look forward to the opportunity to contribute to the advancement of PET imaging.

  • I asked it to solve easy algorithm practices in different computer languages from the Open Judge platform. For example, this is what I got when I ask it to generate a Python program to find the next palindrome date after a given 8-digit YYYYMMDD date.
from datetime import datetime, timedelta
# Convert the 8-digit number into a date object
date = datetime.strptime("yyyymmdd", "%Y%m%d")
# Increment the date by one day and check if it is a palindrome
while True:
  date += timedelta(days=1)
  date_str = date.strftime("%Y%m%d")
  if date_str == date_str[::-1]:
    # The date is a palindrome
  • It can generate the z-matrix of simple molecules. But there seems to be some errors.
H 1 R1
H 1 R2 2 A1
H 1 R3 2 A2 3 D1
H 1 R4 2 A3 3 D2 4 D3

R1 = 1.0890
R2 = 1.0890
R3 = 1.0890
R4 = 1.0890
A1 = 109.47
A2 = 109.47
D1 = 120.00
D2 = 120.00
D3 = 120.00