Are you enjoying all of the ChatGPT and OpenAI hype???
What about what actually goes on behind the scenes?
I have worked through 23 proposals and sold 2 ChatGPT apps for companies you have heard of in the past two months (And the list is getting longer)
Thread about how Corporations are ACTUALLY using ChatGPT.
Spoiler, ChatGPT is replaceable -and I’ll tell you why.
First, ChatGPT is one of the many Language Models that exist. There is OpenAI, PaLM, LaMDA, Galactica, OPT, Megatron-Turing, Jurassic, and the list goes on and on.
But what’s noteworthy is that they are all built basically the same way. All of these collective models are pre-trained by humans, self-supervised, and circular (Yes, it teaches itself).
ChatGPT is ultimately a tool. That is all it is right now - and it is not really all that smart when it comes down to it.
Still, every fortune 500 company should be building an application that uses LLMs to replace customer service agents and chatbots - and they are hiring NineTwoThree Studio to do Prompt Engineering to control the results so that the business can maintain their IP and sell more products.
This might mean NOT using ChatGPT at all.
But how do they learn the correct way to build a chatbot or AI content writer? Let's dig in.
It also is very good at sucking in the knowledge of the internet. Anything that is posted online in public blogs, newsletters, or websites is going to be crawled and used by OpenAI to train the LLM.
This poses the first problem businesses have with an LLM they do not own. Why would a business share company data with a third-party bot? Companies will always protect their IP and ChatGPT has no ability to protect company data in a separate instance.
It is literally a tool that learns from the knowledge it's given. If you are the only person in the world that knows how Turtle Tails work - then the second you tell ChatGPT about it - your knowledge is now every user's knowledge for the rest of eternity - something of an issue when it comes to intellectual property.
The second thing that businesses think about when considering ChatGPT is “how do I control the answers so that ChatGPT can recommend my products or solutions? Especially if I am paying to produce those answers?”
What ends up happening is that building ChatGPT the way you read online is an impossibility for a business. Neither of those concerns above is true - or will ever really be true in the future.
Every proposal I've seen ends up navigating the same path. After it is explained that ChatGPT cannot be used as intended, we begin to talk about Silos and that’s when things get really interesting.
The first silo that needs to be created is a super secret product database. This database will consist of products that the company is interested in selling or a service it offers.
For this type of project, our data scientists will parse this database into chunks to determine similarities of the products and begin testing which semantics lead to which product solutions.
Let’s say for example that you were a dog company selling toys for canines. You want to know if a person wants a dog bone or a squeaking toy. Both are chunks that can be easily determined with a few simple questions.
Our goal is to sell a product that solves the customer's wants. Why are they here, and what do they want?
We proposed a chatbot for a beer company that you have probably sipped on before. What was their number one request?
No matter what, at no time should this chatbot say anything negative about alcohol. It should not demonstrate the negative effects or that it causes sleep loss etc. It would be horrible if said beer company paid millions of dollars in tokens to find out that all their users were being told the harmful effects of their product.
Do you know what else should be in your negative word database? Competitors. At no time should the ChatBot be paid for by Company A recommend Company B.
Now that the stage has been set it is time to understand how to execute a proper ChatGPT build.
In short: you need to be incredibly good at prompting ChatGPT from the questions the customers are asking you.
Prompts should ALWAYS consist of three things:
1 . A persona (Respond like a Harvard professor, respond like a newspaper reporter, etc.)
2. A task or question (Copy and paste the customer question or concern here.)
3. Modifiers and Clarifiers (How can you iterate?)
While 1 and 2 are generally understood, let’s dig into the third option. Modifiers and Clarifiers.
With every prompt, there should be a laundry list of modifiers and clarifiers. This list should inform the ChatBot of the types of answers it SHOULD provide and the type of answers it should NEVER provide.
With this in mind, each prompt can navigate the seas of predictability without fear of misstepping. ChatBots should be predictable and standardized for businesses - not random. Random could be bad for business.
With the controlling modifiers or clarifies our engineers and scientists then go into “Supervised Training” mode which puts the ChatBot to the test.
We basically bring in professionals to have a conversation with the bot and rate the results. Were they able to buy the product they wanted? Did they get the answer they desired? Were they treated nicely along the way?
All of this matters because Chatbot should have the same values as an employee. If you want to treat your customers with respect and kindness - you want the Chatbot to do that too. Therefore, we can even inject your values directly into the Chatbot's persona.
Overall, the result is that all of the information is parsed from the “Positive Feedback Product Seller” database and only provided to ChatGPT through prompts. Therefore, ChatGPT or OpenAI has no ability to collect company IP.
We are two months into building products using ChatGPT. If this thread interests you or you want to build a ChatBot for your business please reach out to NineTwoThree Studio.
We have been building Machine Learning applications since 2016 - but this is different. ChatGPT changes everything. We are already working with clients building applications using OpenAI. Experience matters - and we have the team to help you.