I am going to start a research project and figured I’d share at a top level look at my plan of action. To accomplish my goal I plan on leveraging existing AI tools to help me operate faster. Besides I need to learn how to train GPTs and this will be a great starting point.
If you work in content, you can use the existing AI tech to optimize your performance, without actually changing your tone.You’ll need to feed the AI tool you are using some data. Once it is nice and full, it will be able to spew the kind of answers you are looking for (with some coaching).
At first this process will be very manual. You’ll need to do your Googles and start with some data entry. Once you’ve got your initial documentation in place, you’ll be able to start training a tool like ChatGPT
AI tools can read PDFs and data sheets
One of the main blockers with current GPT software is a memory issue. The way they process language uses a system called tokenization. Every word becomes a token and that gets used rather than raw language to generate a response. Every time you type a query, a certain amount of tokens get used.
ChatGPT can remember the last 32000 tokens. So after a while, as your chat progresses, it will start to forget things you’ve already chatted about. You can use tricks like getting ChatGPT to summarize key info, every now and again, to keep that specific chat on point.
Given there’s a limit, a way to work around that is to compile data as a reference point. Instead of copying and pasting the same info each time you get to training a new GPT, you can upload a PDF or CSV with a bunch of common answers/statistics already there.
While I haven’t tested it, you can get the GPT to read the documents you upload, helping fine tune it. An example would be for me to make a spreadsheet, categorizing my social media post performance. You can track things like post copy, impressions, type of content, reaction count, etc. I believe I have 1000 IG posts. Making a file tracking my post performance can let the tool accurately tell me what to do more of and what to do less of.
While they are still learning to read webpages and Google for you, they are known to hallucinate. A hallucination in AI land is jargon for a tool like ChatGPT making stuff up despite the sources it reads.
Keep track of the answers you give ChatGPT and its friends
The best way to work with a GPT is to first ask it a question like, what are you good at doing? Google has Gemini. Microsoft has CoPilot. OpenAI has ChatGPT. There is an app called Perplexity, which is less strong than ChatGPT for a lot of stuff, except searching the internet. Perplexity is a lot better at using Bing then the other ones (is my understanding). I am sure CoPilot will end up pretty good at Binging.
Side note, it’s pretty wild Bing has become a here to stay search. Google misplayed some early AI and now all the GPTs are using a “Search with Bing” feature.
You’ll want to talk to the AI like it’s a child. Ask the generative AI tool what it is good at. Then assume it’s lying and go see what the internet says it’s good at. The art of using AI will be one of conversation.
You may need to use different tools for different things. That being said, you will need to feed these GPTs a lot of information. Let’s say I want to train a chat to help me job hunt, another to help me write social media posts and a third to help me write emails to close deals. Each of these is going to require specific information to help me do the job correctly.
What probably won’t change much is a description of me, my life, my resources and the other basics. As you answer questions for these tools, keep track of every response. Put them into a file. Label these answers clearly with titles. In the future, the collection of all these answers will become a singular PDF you can upload that instantly calibrates a tool to what you require.
If you have hundreds of pages of data about you, mixed with a proper reference for your voice, you can start to automate some boring shit in your life.
The gist of using AI tools
You will need to issue prompts (commands) to the generative AI tool. If you were to copy/paste some well crafted prompts from the internet today, the answers you would get back would be kind of trash. Those prompts were not made for you.
You can tell ChatGPT to write a song in the style of George Carlin pretty easily. To get it to write one in the style of Holden Stephan Roy would require a lot more training. I’d need to give it access to transcripts of my lyrics and poetry. I’d need to feed it a ton of history about me. I’d need to focus it on the cultural references I would employ.
Start with a brief description of the project you need help with, then ask ChatGPT, or whichever one you are using, to list out the information it needs to help you. Then it will list a bunch of things it needs to know and you need to fill out the answers like a form. In a separate ChatGPT chat (since each one has its own memory) you will ask it to summarize all your responses, making them shorter. Once they are as concise as possible, feed those answers to the main chat, and ask it to write a summary for you showing it understands. Then, make corrections to the summary as if you were editing a term paper.
This process will tune that specific chat to remember all the information that matters. Since you may make mistakes, or go past the memory limit, you may need to restart. Only now you have your giant reference list of answers to give to the AI tool, making subsequent efforts take less time.
It will take trial and error, but once the specific chat is ready, you will be able to say things like, “write a song about pickles in the style of Holden Stephan Roy” and rest assured no love will be given to pickles.Eerie in theory but it’s going to be useful to understand how to use this stuff. Over the next couple of weeks I will have an update for you on how it all goes. In the meantime I recommend following the Everyday AI podcast.for daily news on the world of AI.
Live Long and Prosper Everyone
Comentários