Let's chat, GPT

As probably also many of you I have been infected by the hype on Generative AI in 2023. I started to play around with text generation and image generation and followed with fascination all news about the next impressive show case and all analyst reports on the impact that GenAI will have on work and society. In order not to be only an excited spectator of the show I also dug myself into the algorithms and their mathematical foundations, take courses and certifications, set myself focus on the business aspects of AI and aim to use AI for my own business and eventually support others in doing so for theirs.

By now I’m using two GPTs, namely ChatGPT and Pi.ai, pretty much every day for different tasks and they became invaluable tools to me. But other than Co-Pilots, ChatGPT cheat sheets or specialized AI tools imply, the huge value in GenAI for me in my business is not to create Excel sheets (I can do that sufficiently well for my purposes), to create with one powerful prompt a competent document acting as whatever persona for whatever purpose (good to start, read, discard and begin the hard work) or to create full presentations just with one prompt (doesn’t work if the topic is more complex or the messages need to be more clear and relevant than can be expressed in few lines of prompt).

For me in my daily work the tremendous value of these tools is the process of exploration and discovery, and the structuring of my own thought process. Diving with conversational GPTs like ChatGPT and Pi.ai into the available knowledge and data from the surface to the deep brings me inspiration and knowledge, and it helps me to challenge my own perspectives or results. That along the way good texts can be generated goes without saying.

So: Let’s chat, GPT!

ChatGPT as the talk-to guy

Inspiration, brainstorming and challenging myself is for me in my daily work the HUGE value of ChatGPT and Pi.ai

I’m using them as the guy at the other desk or in the next room where I can go to say “Hey. Listen, I have this task / problem / question / concern / idea … What do you think about it?”

Discover what you are looking for in a dialog – that’s why it’s called ChatGPT

You probably don’t go to your friend and say “Act as a project manager. You have detected a budget overrun. Write an email to the CFO and ask for more money. Use a friendly but affirmative tone.” Hopefully, your friend would suggest other addressees than the CFO and alternative mitigations. ChatGPT would just create the mail, so you can copy, paste and send it – and then you could use ChatGPT to write application letters for your new job.

So, I let the conversation flow. I start on the surface, process what ChatGPT delivers me, ask back where things are too high-level, unclear or seem suspicious, provide my own thoughts back to ChatGPT for its comments and go with it step by step through a dialog that can be just a few iterations, can take a few hours, or also can take a few days.

Depending on the subject I find during the flow many things that I knew already (but it’s nice to get confirmation that I’m not completely off track) but I also always find many things that I didn’t have on the radar, and where I then dive deeper with further questions or provoking statements.

The creators of LLMs are smart, LLMs are not – remain critical

When working with GPTs it is important to stay alert and cross-check. LLMs have vast amount of data and insights, and are powerful and convincing, but they have - even when they are fine-tuned - the same subject matter expertise as a slice of bread.

I’m suspicious when the conversation gets shallow or in some way contradicts my own knowledge or assumptions, or when it is slightly out of sync with the rest of the conversation. In such cases I see to clarify it with other questions or provocations. But it may also be right and I may learn something new or need to get clearer in my requests.

Anyhow, important information needs to be verified before you make use of it and a clown of yourself.

ChatGPT sees you for the first time – every time

Anytime you start a new thread in ChatGPT it is for ChatGPT the first time it meets you. Means you need to give it context. You can do so in one well engineered prompt or with specific prompting techniques, and it certainly works well to come to quick results on sharp topics.

I usually start by saying “Hi”.

Believe it or not, I feel more comfortable and natural when I’m nice not only to humans but also to GPTs, even though it is not necessary (with GPTs(!) with humans it is(!!!) ). With GPTs it helps me to enter interaction, other than queries in a Google search.

Then, or in the same prompt, I set the scene “I want to consult with you on … . But, before you answer let me tell you …”.

I think that GPTs really enjoy jumping to conclusions. If you tell it “I want to consult with you on my value proposition” it immediately provides you an essay on what a good value proposition should be made of and asks you if there is anything else that it can do for you, before it even knows for what type of business you want to work on a value proposition.

Therefore, it’s good to make them listen before answering. You can spare yourself lengthy, generic stuff. But this lengthy generic stuff may also contain a gem of new insight. It’s up to you to find your own way with the GPTs.

After opening the scene and telling it (politely, see above) to shut-up until being asked, I can give it more or less endless contextual information.

Come prepared to a meeting with ChatGPT

As the dialog progresses it becomes richer. Any question or statement (aka prompt) that I give and any answer that ChatGPT provides adds context information to the chat. The more context there is, the better is the output.

When I’m discussing a multi-level, multi-sided problem with ChatGPT I start at the beginning where not much context is necessary. Then I guide it with statements, questions or provocations until it is where I want it to be. And finally, I jump on my actual topic and take it from there.

I can also switch to other levels or sides of my problem. The thing is that the earlier conversation is part of the context, and the elaboration of other aspects will be more in sync and more relevant for what I want based on this.

This also means three things:

1. If you want to talk with ChatGPT about something else, open a new thread. Otherwise, it may blend weather conditions into your proposal summary.
2. Plan your conversation regarding the topics you want to discuss and go in a reasonable sequence. Similar as you would do also in a natural business call. For ChatGPT the most recent information in the chat is the most relevant one.
3. Don’t get lost. As in real life jumping from left to right, back and forth, makes you and your counterpart dizzy. I usually discuss one part of the problem until I’m happy, then I ask ChatGPT for a summary to be sure we’re on the same page and to refresh its context, and then I move to the next level or aspect.

But, what if

I experienced that when chatting with a GPT in a longer thread, or when creating texts, certain things occur that need to be worked around or that need to be considered.

GPTs have limited attention

When I find that in longer chats that the GPT may forget parts of the earlier discussion, I guide it back to it (“Remind me what we discussed on …”) and if necessary, I point out gaps (“Remember that we also spoke about…”) and we continue with refreshed spirits on the same page.

I also let ChatGPT create summaries of discussions or certain topics of the discussions that I copy and store away in a separate document. This is good for later reference, but also good to give context to a new chat thread or to remind ChatGPT when it got lost.

GPTs want to make you happy

Another trap you can step into is bias. Ok, ok, everybody knows by now that LLMs are trained on data and that if the data is biased so will be the output.

But it is also you in the chat who is creating a bias. GPTs want to make you happy.

Of course they don’t. They don’t want anything. But training data or training feedback may have guided GPTs to “think” that humans want to be happy, and that humans are happy if one confirms them in whatever they say or if one tells them whatever they probably want to hear.

When I give context to the GPT, telling it who I am, what I do, what my business problem is etc. I hear or read things like “impressive”, “well thought through”, “comprehensive and convincing” etc.

And then I’m happy and obviously my job is done and I’m great. Nope, probably not.

Tell GPT not to praise you but to criticize and challenge your views. Let it find the holes in your story, e.g. “Don't praise my view’s comprehensiveness or completeness, instead point out its weaknesses and gaps, and ask the tough questions that also the audience would raise.”

You will be surprised on what you’ll learn about your greatness. As I said, for me the value is inspiration, brainstorming and challenging my views.

Chatting is exhausting for both parties

After a long day’s work ChatGPT or I will get tired. Often it’s both of us.

For myself I can say that working with GPTs effectively is real, intellectual brainwork. I get so much input, need to evaluate and consider so many new points, need to continuously guide the conversation (a good training, by the way) that things at some time begin to blur. It’s like long focussed meetings. If this is happening, it’s time for a break as in normal meetings.

Interestingly, I have the impression that also ChatGPT gets exhausted. Why’s that, I have no idea. Perhaps due to attention limits or OpenAI thinks that I had my share of compute time for the day.

I notice it when the exchange becomes flat, nothing new comes up, the answers become shorter – again very similar to real-life meetings. In this case I ask it for a reminder or summary of relevant parts of our discussion or I provide a refreshment prompt from archived information, and back we are.

ChatGPT as the Ghostwriter and Editor

Of course, I’m using ChatGPT also for translations, emails, document work or texting. And it’s a huge time saver - and it’s a bit humiliating when you figure out that the model not only speaks better English but also better German than yourself. So, it’s my ghostwriter and editor and it’s doing it nicely with a bit of consideration on the instructions.

But: if you use GPTs to generate content that someone shall or will relate to you - you better proofread it and adjust it. The information can be convincing but wrong, the text can actually be high-level and generic while looking competent, or its tonality and style is so far away from yourself that no one will believe you.

Perhaps there are tools that are doing it just perfectly, but fact-checking, mind-reading or personality-reading is probably not part of their skill profile.

Anyhow, the GPT generated text is perhaps 70% ready, but more important to me: It gives inspiration on what I really want to say and what not. And with that I can refine it myself or I work further on it with the GPT.

Goes without saying but can’t be said often enough

Be mindful of IPR, business secrets, personal data, competitive information etc.!!! You can get from GPTs anything you are looking for without sharing critical information. Anonymize your context, talk around things that you can’t talk about, guide with questions rather than sharing.

Apply common sense and your data protection regulation.

Bringing it together

I guess, it became clear that I like to work with GPTs. It is at least for me highly effective. There are the one-shot queries or quick text tasks that can reduce hours of work to minutes. And there is my favourite, the brainstorming and discovery, that takes longer but brings huge value and depth and can reduce the work of days to hours and weeks to days.

I’m really curious what other things will become or already are possible with the tools today, and I keep exploring and trying out.

I mentioned that I’m also using Pi.ai. I’m doing this less often and with a different spin than ChatGPT, but this is another story that shall be told at another time. I encourage you to give it a try, and in general just to try things out.

If you found this story interesting, have comments or questions, or just want to chat with a human - feel free to contact me here on LinkedIn. I'm looking forward to it.

#ChatGPT #AI #ArtificialIntelligence #PiAI #PersonalAI