Here is a screenshot of my interaction with LLM when doing the Industry 4.0 assignment task.

Though our professor gave the documented and oral instructions, it is still unbelievably clear that students felt unclear about HOW and WHY to solve it. The confusion arises in every aspect possible.

Since I am one of them(the class is mixed with CS students and half layman like me), I decided to observe myself regarding the confusion. I decided to interact with GPT, to get a better understanding and finally develop answers to the questions, instead of relying on GPT's response only.

So what I did:

The instruction for the assignment is like this:

Before diving into the details, the starter is a simple refer-to-what human language task based on the text itself. For example,

• These traces -> power trace

• The first two -> current, voltage - used to determine the power consumption

• The digital signal -> used to determine when the device is performing a specific task

Next is to match the text (instruction)-to-code task, where my colleague's feedback is like this:

For me, it's even worse(at the beginning), I cannot understand the code easily. I don't know where my attention should be. There are 2 outputs in [16] and [19], so i randomly take the [16] raw data into GPT,

Oh it reminds me of the things I did in the Preliminary stage.

Then I was confused with time params, so I started asking about delta T.

It turns out I was confused with the relationship between data-sample-rate and delta T, then I can easily calculate the energy consumption.

I knew what I want from GPT, so I prompted it to fill in the blank via (?)

I also quoted GPT response and asked

But interestingly, if we compared the same question given by GPT response and my (another) colleague, the latter is very straightforward.

Takeaway:

• In this practice, I have a "goal"(with many sub-goals) to understand and solve the assignment task.

  • Satisfied🕺🏻: conversational interactions.

  • Back-off🤯: A LOT of responses which add to my attention load.

Observation:

• Prompts are Goal-directed, every prompt contains a signal towards the goal, and the REAL prompt is likely formed after some active interactions. There is unlikely a one-shot problem-solving prompt.

• The subgoal(next prompt) is dynamically defined through interaction, I stared at the things(GPT response) I understood and explored things I didn't AROUND the understandable context. - Here is where the subgoal is defined step-by-step.

• Every interaction, there is a tiny component being picked up, towards the goal.

• It reminds me of a paper*[1]*, the insight there is adding noise to get better signal, where they leverage RL in their model.

• Maybe there is less need to verify the content of LLM FOR the users because we underestimate people's ability to verify the LLM "hallucinations" and overestimate our ability to do this for people. Instead, we might record the interaction and provide the user with a reflection of their learning process - e.g. in TLCA how they draft the input, how they prompt, and what profile they choose.

The paper is interesting in how it uses suboptimal human conversations for training the AI, specifically it employs “offline reinforcement learning to train an interactive conversational agent that can optimize goal-directed objectives over multiple turns.” The insight is a bit like adding noise to get better signal — a method used in audio processing and mastering, and in fact diffusion image AI models.[2]

This also reminds me of the huge gap in the eye of a domain expert and a learner, the linking path to the knowledge is not straightforward and quite personalized.