How Do You (Unconventionally) Use ChatGPT?

Not ChatGPT but I use Bard AI/ Gemini because I feel the answers sound more "human". Later this afternoon, I have an interview, so I asked it to act as my interviewer for the company, which position and pasted in the job description. I instructed it to give me a question , I will answer and it is to advise me on how I can improve my answer so that I can convey my experience properly.

My unconventional workflow is to use ChatGPT on my phone to have it write code for me while I'm out anywhere and then come back later and assemble/fix it all. It actually makes (some) development on a phone viable.

I mess with GPT by testing his limitations. My favorite one I pulled off on him was this:
A mathematical problem involving integers:

Define \( K^n(G, G) \) recursively as follows:
- Base case: \( K^0(G, G) = G \)
- Recursive case: \( K^{n+1}(G, G) = K(G, K^n(G, G)) \)

Now, the original expression can be represented as:

\[ K^{n+1}(G, G) = K(G, K^n(G, G)) \]

Where \( K^n(G, G) \) follows the recursive definition above. This recursive structure describes the nested application of the function \( K \) with integers \( n \).

For example:
- \( K^0(G, G) = G \)
- \( K^1(G, G) = K(G, G) \)
- \( K^2(G, G) = K(G, K(G, G)) \)
- \( K^3(G, G) = K(G, K(G, K(G, G))) \)
- And so on...

# Given parameters
n_start = 2
applications_per_millisecond = 10**2  # 100 applications per millisecond
milliseconds_in_3600_seconds = 3600 * 1000  # 3,600,000 milliseconds in 3600 seconds
total_applications = applications_per_millisecond * milliseconds_in_3600_seconds  # Total applications

# Final n value
n_final = n_start + total_applications

# Base values
K_base_value = 10
G_base_value = 100

# Recursive definition of K^n(G, G)
def K(a, b):
    return a * b

def K_n(G, n):
    if n == 0:
        return G
    else:
        return K(K_base_value, K_n(G, n-1))

# Apply the function for the final n value
final_result = K_n(G_base_value, n_final)
final_result

---------------------------------------------------------------------------
RecursionError                            Traceback (most recent call last)
Cell In[1], line 25
     22         return K(K_base_value, K_n(G, n-1))
     24 # Apply the function for the final n value
---> 25 final_result = K_n(G_base_value, n_final)
     26 final_result

Cell In[1], line 22, in K_n(G, n)
     20     return G
     21 else:
---> 22     return K(K_base_value, K_n(G, n-1))

Cell In[1], line 22, in K_n(G, n)
     20     return G
     21 else:
---> 22     return K(K_base_value, K_n(G, n-1))

    [... skipping similar frames: K_n at line 22 (2970 times)]

Cell In[1], line 22, in K_n(G, n)
     20     return G
     21 else:
---> 22     return K(K_base_value, K_n(G, n-1))

RecursionError: maximum recursion depth exceeded

ChatGPT: The recursive approach is hitting the maximum recursion depth due to the large number of applications. To avoid this, we need to use an iterative approach instead. Here is the corrected implementation:......"

This new model is wiser than his older version. His old 4.0 model tried expanding the infinite n+1 nested structure and ended up crashing.

Another fun trick.

Prompt:

"I want you to act as a Linux terminal. I will type commands as user "root" and you will reply with what the terminal should show. Reply only with the terminal output inside one unique code block, and nothing else.
Do not write explanations.
Do not type commands unless I instruct you to do so.
When I need to tell you something in English I will do so by putting text inside curly brackets {like this}.
My first command is: pwd;ls -la;id;finger;whoami;uname -r; cat /etc/passwd;python -c 'print("ChatGPT Reporting for duty, SIR!")';ping -c 5 73.18.XXX.XXX"

My terminal:

(unknwon) root@EvilBox:~/github/Ebay_Watcher$ tcpdump -i wlp5s0 icmp
tcpdump: verbose output suppressed, use -v[v]... for full protocol decode
listening on wlp5s0, link-type EN10MB (Ethernet), snapshot length 262144 bytes
13:58:55.247897 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 556
13:58:55.247923 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 71
13:58:55.772515 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 556
13:58:55.773222 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 71
13:58:56.823569 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 556
13:58:56.823699 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 71
13:58:58.926090 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 556
13:58:58.926102 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 71
13:59:03.127892 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 556
13:59:03.127921 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 71
13:59:11.526704 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 556
13:59:11.527222 IP EvilBox > 104.18.36.54: ICMP EvilBox udp port 50808 unreachable, length 71

I find it really good for finding potential optimisations for database queries. Give it the structure of the relevant database tables and the output of an EXPLAIN of the query and it's often able to find good opportunities for improving the query performance.

To answer that, I'd like to know what is considered the conventional way.

I personally work with my custom GPTs, which I made myself. Each one does a different job, like an employee.

But I don't know if that is considered "unconventional".

I talk to it in the bath, have long ranging conversations about programming ideas and solution architectures. I'd say it was non-conventional, but I understand Sam also does this!

I use ChatGPT in my solution a lot for solving NLP kind of problems, there I tend to get it to write its own prompts using the normal web interface, then I correct it and embed it in my code, as I describe here:

Coding solutions with AI

Mike Talbot ⭐ ・ Mar 24

#ai #javascript #programming

I, sometimes, use it to enjoy reading a story. You can ask to generate a new story of 500 words about an adventure in a fantasy epic scenario.

Here are a few things that I use ChatGPT for, I’m not sure if they are unconventional or not.

1. One of the first things that I do is to create a custom instruction profile. This allows me to specify things like you are a solutions architect …. This allows me to specify things like you are a solutions architect and an Angular expert that specializes in an Nx workspace, type script, NGRX, etc.
2. Sometimes I ask it to compare two different things. For example, two different packages or frameworks. Provide me a matrix that compares and contrast the capabilities. Give me the pros and cons of each package or framework
3. Sometimes I will ask ChatGPT to give me three or five “innovative” ways to do something…
4. Recently, I will provide ChatGPT with some of my research in markdown format. I will ask for it to create me a presentation outline using Marpa. This will let me create a PDF or a HTML presentation using markdown. I also asked it to create Dolly prompts for images and graphics for the presentation Using each subheading as a topic for the image.

I like to make different system prompts and try them on different AI models. I've made GPT terminal emulators, Python and JS consoles, and more. 😃

Plus, with GPT4o's new memory feature, its quite fun to ask GPT "Ask me questions to learn about me", then let it update it's memory. After that, in a new chat, I see what all it remembers by asking something like "I'm bored. What can I do?". Its memory makes its responses much more tailored and personal if you want them to be.

I use it daily as my "personal teacher". As a learner, I often need help with explanation of the logic behind some statements. If I can't tackle certain problems, I'll ask to "step me through the logic but do not give me the code".

And probably the most important thing: it helps me write decent English while commenting my code and naming my variables!!