💁‍♂️ Okay, "thumbs" are a fast and well-known method of evaluating. So why it doesn't work well in the chatGPT case?

→ Feedback Loop: In an ideal scenario, the feedback mechanism should be quick and direct so the user can immediately see the effects of their actions. "Thumbs-up" and "thumbs-down" do not provide such immediate feedback within a single conversation.

→ User Engagement: Engagement can drop if the user doesn't see that their actions have any impact. Quick and direct feedback is one way to increase engagement.

→ Personalization: AI-generated answers are generally tailored for a whole algorithm, not individualized. Feedback mechanisms could influence personalization, but this usually works over the long term and collectively rather than individually.

→ User Satisfaction: Lack of immediate impact on conversation quality can affect the overall user satisfaction with interacting with the system.

👍 👎 "Thumbs" evaluating cases known from Facebook or YouTube are more engaging because they directly affect the creator of the content and inform him how large a group of recipients is satisfied with it. It's more personal interaction.

🤝 It is a more personal relationship in which we engage emotionally. That's why we are waiting for the effects of our feedback. See how strong the engaging element is in the creator's reference to feedback.

👁 In the case of a tool, however, we expect an immediate response if we consider it helpful. The general goal of developing the entire algorithm does not improve the achievement of our short-term goals. Therefore, implementing an AI-based solution is worth taking care of the feedback loop.

What are the possible solutions? More about it in future publications. 👏

------------
Did you like it? Thanks for your:
💬 Comment
👍 Like
🔄 Share
------------

#UX #chatgpt #ai #LLM #feedbackloop #engage

Two years ago, I bought a Tello drone to learn how to code in Python and have fun at the same time. Just as you probably suspect, after two weeks, the drone goes to the basement.

Now, in the AI era, it’s time to start again — GPT will be my programmer buddy. But will it work? Let’s try!

First — why DJI Tello drone?

Because this drone has an open SDK — we can download the appropriate library and write commands in Python to control the drone by WI-FI. And this machine isn’t expensive — something around 100$

This was pros, now cons:

• The battery works max 15 minutes, and the charging needs 30 minutes. You can buy an additional battery, but it costs 1/4 of the drone price.

• The camera has poor quality.

• Very short range.

• Very short WIFI connection time on standby

So, it’s good for learning but not for nice video shooting on trips

Part 1 — Create my AI buddy

I assumed that I would need constant access to the direct Tello code file for efficient work on the Tello code. Of course, I could create an online GPT assistant on the Open AI website, but I won’t upload the code files again and again. At the same time, I wanted to use tools like GitHub Copilot; my ambition was to create my personal GPT agent.

I assumed that my first structure would be:

• GPT agent in Python using API to connect with model

• GPT agent in the same location as Tello code

• GPT agent in the same location as the basic instruction file

• GPT agent automatically retrieves information from the instruction file and Tello code file to generate an answer

• Whole communication works only in PyCharm output for now

But after that, I needed to program in Python my GPT agent, so I used online GPT for this task. I get short code, and this code was short and looks easy — but it didn’t work.

I spent a lot of time trying to resolve the problem. I followed GPT instructions, reinstalled Python on my Mac, and updated Open AI libraries in PyCharm for nothing.

In the first step of my work, where GPT was supposed to be my programer agent, I found his limitations.

Open AI changes of the way API connection with GPT was made in November, but GTP was trained on data until July 2023. So, GPT didn’t know how to connect by API in the new way. And he didn’t know about his should know.

So, 1:0 for human

I found the actual script online, but I needed to make some changes to it: I needed to retrieve information from other files to provide precise GPT agent answers. Surprisingly, when I asked to chat about modifying the new script, which he didn’t know, he did it properly:

import os
from openai import OpenAI
def load_file_content(file_path):
    with open(file_path, 'r') as file:
        return file.read().strip()
api_key_file = "openaiapikey.txt"
context_file = "context.txt"
additional_files = ["tello_code.py"]
api_key = load_file_content(api_key_file)
context = load_file_content(context_file)
additional_contents = [load_file_content(file) for file in additional_files]
client = OpenAI(api_key=api_key)
while True:
    question = input("Ask me anything (or type 'exit' to end the conversation): ")
    if question.lower() == 'exit':
        break
    full_context = f"{context}\\n{additional_contents}\\n"
    full_prompt = f"{full_context}\\n{question}"
    response = client.chat.completions.create(
        messages=[
            {"role": "system", "content": "You are a helpful assistant that provides information."},
            {"role": "user", "content": full_prompt}
        ],
        model="gpt-3.5-turbo",
        temperature=0,
        max_tokens=1024,
        n=1,
        stop=None
    )
    if response.choices:
        assistant_response = response.choices[0].message.content
        print("Assistant:", assistant_response)
    else:
        print("No response.")

The next step was to create a list of instructions for my GPT agent to save in an instruction text file. I used GPT online for this task, too — my requirement was my model should be a Python programmer with knowledge of Tello SDK. ChatGPT clarified this expectation like that:

Project Name:
Programming the Tello Drone in Python

Project Goal:
Develop software to control the Tello drone using Python. Utilize Tello SDK and relevant Python libraries for controlling and interacting with the drone. Integrate with the OpenAI API to enhance drone functionalities and intelligence.

GPT Assistant Responsibilities:

Python Programming Support:
Assisting in writing, debugging, and optimizing Python code. Advising on best programming practices and efficient use of resources.

Knowledge of Tello SDK:
Understanding the functions and capabilities of the Tello SDK. Advising on the use of SDK for controlling and monitoring the drone.

Python Libraries for Drone Control:
Knowledge and recommendation of suitable Python libraries for drone programming. Assisting in integrating these libraries with the existing code.

Integration with OpenAI API:
Supporting the integration of Python code with the OpenAI API. Advising on using AI capabilities to expand the drone’s functionalities.

Expected Skills of the GPT Assistant:
Advanced knowledge of Python and programming experience. Understanding of drone control specifics and knowledge of Tello SDK. Experience with API work, especially with OpenAI. Ability to solve problems quickly and think creatively.

And it works!

My new GPT agent connected by API and generated the first code to connect with the Tello drone.

Without programming knowledge (maybe 2%), just by talking with chatGPT (with one general problem at the start ;) )

Full time spent for part 1–2h But without actualization problem, it should be 30 minutes

Sounds interesting? Are you waiting for the next parts?

❤️ Like
💬 Comment
🔔 Subscribe

#GPT #python #drone #programming #GPT assistant #DJI #Tello #openAI #openAI API #AI #AI programming

📞 I called the shop service a few days ago to return my old dishwasher. At first, I was connected to an automatic assistant:

🔸 Assistant: tell me about your case, and then I’ll redirect you to the appropriate department
🔹 Me: I need to return my old dishwasher
🔸 Assistant: Sorry, I don’t understand. Try again with different words or choose “9” to redirect to a consultant.

What will you choose in this situation? More trials to redirect or redirect instantly? I chose the second option.

→ That was an example of minimal efficiency for the user.

I can make more attempts to communicate well with artificial assistants, but I need a guarantee of success and perspective to save time. If I can choose an alternative way and give me 100% success, why should I try more time?
An autonomous system should create an environment to resolve my task better or faster than humans — and then it can give a real alternative. Before that, it was only a feature.

😑 Nowadays, customers need very high experience. They don’t need autonomous assistants because of the fact that it is something new and exciting. Of course, it’s hard to compete with humans, but it’s only proof of how vital the efficiency of these systems is.

→ This was the user side. But how does minimal efficiency look from a business point of view?

I assume my case of automatic assistant is profitable from a shop point of view. If only 20% of customer responses will be understandable for artificial assistance, then the cost of the general hotline is reduced by reducing its occupancy. But it works well if your business is large-scale.

⚖️ Suppose you’re a small or medium company. In that case, you have to care about every customer, and the cost of implementing an inefficient system (in terms of money and user satisfaction) can be higher than profits.

The perspective of cutting costs by implementing artificial assistance is tempting. But you should remember:

Your profit = profit from savings — (cost of bad CX + cost of implementation)

🎯 Conclusion:

AI implementations shouldn’t be only a feature or quick fix for cutting costs. Costs of bad customer experience and implementation can be higher than profits because of the high expectations of your customers. All that depends on your scale, competitors’ strengths, etc.

Minimal efficiency (in business and experience) is the key.

— — — — — — -
#CX #customerexperience #AI #automonus #efficiency hasztag#LLm #consultant #implementation #UX #userexperience #artificialinteligence #businesscosts

🙄 Have you ever felt like you were not smart enough to use a tool or software when you tried it?

It is the center of user-centered design: making products as simple as possible to avoid user discomfort.

Problems of user can drive to:
→ Not optimal use of a tool
→ Leaving the product
→ Avoid payment for a tool
→ Sometimes physical danger

Until today, tools and products have resisted trust, crafted or programmed to perform creator instructions. The user was the weaker side — the device’s correct operation depended on its proper use.
Of course, many of these are badly designed or performed wrong, but there were always two sides: users and creators (in the imagination of the better-educated and more experienced).

👎 So, the user first looked for the weakness in himself, despite the frequent shortcomings of the product itself.

Don Norman is one of the pioneers in interactive design and user experience. He said that the responsibility for proper operation rests with the designer, not the user.
In his book “The Design of Everyday Things,” he proposed seven questions (7 stages of action) that every user of a well-designed product should answer:

→ What is my end goal?
→ How can I achieve it practically?
→ What approach is best?
→ What’s the step-by-step plan?
→ What were the outcomes of my work?
→ What’s the value of these results?
→ Did I succeed in achieving my objective?

💭 Imagine a standard software product. If you read this, you can go back into your memory and recall good or bad examples for every step. It is the common way of our product experience.

🤖 Conversational AI tools (CAI for this article) complicate this rule.

Why?

👉 For the first time, humans received tools, which results and mode of action are difficult to predict. What’s more, the results can vary significantly each time. The CAI is more sensitive to the slightest changes in talking to him. As in the case of a conversation with another person, it is impossible to develop a uniform instruction that, repeated each time, will bring the same result. CAI requires much more responsibility, attention, and knowledge to make the effect match the user’s needs.

👀 But the first time I see people stop thinking about themself as weak or stupid. Now, CAI tools (because they are close to the human way of communication) must prove their usefulness and flawlessness.

🎴 However, the current implementation of conversational AI places too much responsibility on the user.
To ensure accessibility for all, we must find new and innovative ways to address the “7 stages of action” by combining human-computer interaction with human-human interaction experiences.

🧠 Inspiration section

New research provides insights into how trust operates in our interaction with technology - its transfer from human faces to objects, specifically ATMs. The study involved 57 people who assessed the trustworthiness of ATMs with different faces based on age, gender, and ethnicity.

Key findings revealed that ATMs with children's faces were perceived as more trustworthy than those with adult or elderly faces. However, the face's gender or ethnicity did not significantly alter trust levels.

Link to research here:

https://www.mdpi.com/2076-328X/11/6/91