Things to consider about using Chat-GPT in business
Aren't you getting a bit tired of the hype around Chat-GPT? Let's cool it off a little, and take a more stringent, rigorous look at this shiny, talkative AI tool that is bookmarked by half of the business world by now.
You know, like an engineer would do.
Before we dive into it, we want to give you some context:
The founding members of Lexunit decided to dedicate their professional lives to the AI revolution, 5 years ago. Natural Language Processing has been with us for long years now, but Chat-GPT is the first tool that really caught the attention of mainstream media, mainly because it has a smooth User Interface: you can chat with it. And it's free to use by anyone.
At this point, it's getting hard to join this conversation without regurgitating the basics about Chat-GPT, again and again. But because we've helped many companies and startups to utilize automation, neural networks, and similar AI solutions in the past few years, we feel obligated to share our own points of view about these new developments.
Hopefully, we will be able to fire up some guiding lights you can follow to explore the areas of improvement, whatever your field of expertise might be.
We are also motivated to highlight some aspects of Chat-GPT, and upcoming AI technologies we feel are missing from most articles on the subject you can find on popular websites.
We chose a casual, conversational format. Chat-GPT was not involved in creating this dialogue, in any way.
What is Chat GPT?
Chat-GPT is a Language Model that has been trained by Reinforcement Learning from Human Feedback. (RLHF)
Okay, what is a Language Model then?
Language Models are commonplace by now: they are required for providing translations, to create subtitles for videos in realtime, and they make your phone or any digital device understand what you say when you instruct them using only your voice.
Language Models enable Gmail to offer pretty good suggestions for the next words you might want to type.
They are needed every time you want a digital device to interact with human language. This is the realm of Natural Language Processing, which has been an existing scientific field since the 50s. Obviously, it gained some momentum in the last two decades, thanks to advances in using Neural Networks.
This is why, for example, Google Search exists. And this is why it's gotten ridiculously smart, or at least we can argue that it is more and more capable of discovering true user intention behind every search query.
Who built Chat GPT?
Open AI LP is an 'AI research and deployment company', which is owned by the non-profit OpenAI Inc. This is not a garage project: the founders include PayPal founder Peter Thiel, and Elon Musk (it really feels redundant to introduce Musk to anyone who can read and has internet access in 2023).
It's headed by the former president of Y-Combinator, Sam Altman. (You can safely add Jeff Bezos to the list as well since AWS is also involved.)
The project started out with a cool 1 billion dollars pledged by the founders - which has been doubled by Microsoft, and as recently as January 2023 - probably not unrelated to the intense hype around the tool - the Redmond software giant showed up with a new, 10 billion USD investment round, reportedly, extending their partnership with Open AI in multiple ways.
What's so special about Chat-GPT? Is it different because of RLHF?
We'll get there in a sec. Let's take a quick look at the evolution of Chat-GPT first.
Chat-GPT had three or four major predecessors:
GPT or Generative Pre-trained Transformer was the result of the method where a Language Model was built with generative pre-training. It has been published on June 11, 2018, less than 5 years ago. Deep Generative Models are formed through the combination of generative models and deep neural networks.
GPT-2 and GPT-3 have been marked improvements, the latter containing 175 billion parameters - now that's some learning that goes really deep. (GPT-1 knew a meager 117 million parameters)
GPT-3 introduced generative tools for the public, namely DALL-E, a Transformer that creates a new image based on your instructions.
The next step was InstructGPT, which was trained with 'humans in the loop', to make it follow English instructions better. Beyond that, the people working on the project are also labeling the outputs from the models, and use this as fine-tuning. So basically they reward answers they think are appropriate, and reject the ones they are not satisfied with. The model learns this and provides more appropriate outcomes afterward.
So yeah, Chat-GPT is basically better because it's larger, and before we get to play around with it, its creators fine-tuned it for us. This process is not finished - this is why you or anyone using Chat-GPT can send feedback about anything the program says.
Is Chat-GPT a groundbreaking innovation?
Chat-GPT is not innovative in terms of the methods or the technology involved in its creation. It's certainly not unique: you can be sure that Meta, Alphabet, and other tech companies have very capable Language Models at their disposal.
Chat-GPT is certainly innovative and exciting in the sense that it's a real, working tool, and people can freely play around with it. It finally brought AI right into the forefront of everyday conversations, where, in our opinion, it belonged for a long while now. It was long overdue.
The arrival and cultural relevance of Chat-GPT probably makes our job a bit easier too, because when we talk about business leaders about how they could utilize AI, we can now skip some early steps of these conversations.
Chat-GPT does a lot of the explanatory work sheerly by existing. Here's an apt quote from a McKinsey article:
'For us and many executives we’ve spoken to recently, entering one prompt into ChatGPT, developed by OpenAI, was all it took to see the power of generative AI.'
What can Chat-GPT help me with?
When you find yourself creating text following a certain pattern, it's highly probable that Chat-GPT can help you with that.
Cover letters, contracts, company Facebook posts, generic blog posts, most of any written form of Customer Support communication, job descriptions, and almost anything related to HR and recruiting falls into this category.
And yes, some teachers say, that most of the homework essays, up to the college level, can be done by Chat-GPT in a better-than-acceptable way. They need to grade it as an A. (Is this a danger? We'll get back to that soon.)
The question is: does the body of text you want to create have any sort of identifiable structural characteristics?
If yes, then most probably Chat-GPT has already read a lot of it, so it can create one for you, sort of 'averaging' all of those similar texts that had been created before.
This is important because if you understand this, you will realize, that Chat-GPT has no idea what it's doing (because there is no 'one' there, who could have an 'idea' about anything. You are talking to an abstract mathematical formula. We leave it at this now, but we will dip our toes in the philosophical depths a bit further down the article.).
That's neat, but I want some real use cases.
Google it! Seriously, playing around with the tool is important (We honestly feel it's kind of like turning on a radio or a TV set for the very first time, you should take part in this experience if you hadn't before!). But also feel free to check out the compilations people have already built up, don't settle for the few examples journalists could come up with while reading what other journalists came up with.
This one could be a good start, there are probably plenty of similar others.
Well, that list is… formidable. Is my career over?
No, your career is changing.
Just like it's already been changing. Depending on your mileage, you have seen some major shifts, we're sure of that.
And although it might seem tempting to ask Chat-GPT for a training regimen or a week's menu, we think you should still consult professionals about things that really matter.
So, how does all this affect your business?
Imagine that Chat-GPT or a similar tool enables you to cut through the red tape. If a significant chunk of your work revolves around sorting, editing, reading, and creating text, these tools can help you do it faster, leaving more time for the deeper, more complex tasks.
If you imagine, for example, how the recruiting process looks right now, you can see that most of it could be done by advanced chatbots. Human experts could sit on the top of a pyramid, overseeing tasks performed by robots. Instead of doing all the repetitive tasks involved with this field, these people could dedicate a lot more time to do interviews, to make sure newcomers fit better into the organization, and the overall mental health and professional fulfillment of their coworkers at the company.
Here's another quote that sums it up well:
Many of the tasks that have too high costs right now, could become easily doable in the near future because professionals have more time to focus on them, thanks to trivial tasks being outsourced to AI tools.
What are the dangers? What should I be aware of?
First, let's go back to why Chat-GPT is the first of its kind.
Even Open AI was cautious about it: the previous models, like GPT-2, have not been released to the public like this.
This is because AI experts are aware of how generating content can cause trouble.
Just imagine an army of bots, spewing out generated, and totally made-up content that is largely indistinguishable from content created by humans.
Twitter, e-mail, Facebook, virtually any and all digital communication channels could be flooded. We could go blind and deaf in seconds if we don't have a way to separate signal from noise.
We are already in deep trouble because of misinformation, which is often empowered by bots.
This is why Chat-GPT has been heavily re-trained, again and again, by the experts at Open AI, until it came up with answers they felt are safe to show to the public.
With the caveat, of course, that Chat-GPT can be factually incorrect. (Of course, it can: it learned from the Internet!)
You need to remember what Chat-GPT is: it's a slot machine, which gives you a result that has the highest probability. It arranges words, sentences, and expressions in a particular order that is highly probable to be a fitting answer to your prompt because it is most similar to the ones in the corpus it was trained upon. Without any sort of concern about the meaning of those sentences and words whatsoever.
Some critics point out that Chat-GPT can be terrible at math, it doesn't understand basic principles - which is true, but Chat-GPT doesn't understand anything at all.
You can have a chat with it, but it's still just a machine, right? Like, a calculator.
Exactly, and you see, here's what's problematic about how people often talk about AI. The media likes to anthropomorphize these complex mathematical tools.
And we get it. Being human is being lonely.
For thousands of years, we've been eager to find someone to talk to, who can also experience existence, but who is not one of us, so it has a distinct point of view, and we can compare notes.
This is why the question of alien life in the universe is so exciting. This loneliness is beneath the anxiety that is resolved by religion. And this is why you can read news about a scientist who helped a Language Model to hire a lawyer to get free.
I've heard about some kind of parrot…
Yes, the stochastic parrot. We don't like calling Language Models 'parrots', because parrots are at least living-breathing creatures, with some cognitive capabilities.
Think of talking to Chat-GPT more like shouting into a chasm with highly complex geometry. The sounds you will hear from it will sound interesting and original, but it's still your own voice, distorted, manipulated by the weights on the neural network.
We would advise against asking The Big Questions from Language Models or try solving complex problems with them because they will provide mediocre answers. Rehashed versions of the texts they have been trained by.
They can do incredibly great assistant work, though!
A key aspect to keep in mind is that when you import a neural network into decision-making, you will lose causality. You probably can't have answers to questions like 'why' an AI tool did something specific, beyond a certain level of complexity. If you aren't satisfied with the result, all you can do is go further with the training process.
Just think of what AlphaZero did to chess. It beats the best human chess players in the world. Did people stop playing chess? No. Chess grandmasters have been intrigued because Alpha Zero invented new ways to play the game.
How? Why? No one knows. No one can know.
But now those new tactics are used in high-level chess. Chess changed, forever.
Think of this, when companies start using more and more advanced AI tools, because they achieve higher performance, in ways unknowable to humankind. This is one of the key conclusions in Age of AI, the revealing book by veteran foreign policy expert Henry Kissinger and ex-Google CEO Eric Schmidt.
What if we use this technology for Internet Searches?
Some reports say that Microsoft will integrate this technology into its search engine as soon as March 2023, and Google is probably hard at work to provide a similar solution to the public.
Which begs the question: what exactly do we want from a Search Engine?
Do you want an invisible, digital ghost to give you the best possible answer to your search query?
Or would you prefer to see what other people (sources) think about the subject, in the order of relevance?
When you are looking up simple stuff, quick bites of information, the first one would be a great solution, this is why the Answer Box or Featured Snippet on the Results Page of a Google Search is such a handy tool.
But it won't cut it if you want to research scholarly articles, review opinions of experts, you know, inform yourself.
If you turn this around, it might give us a hint about why is Chat-GPT a problem for education in the short term, and a boon in the long term: maybe we've been testing for the wrong things all along.
Maybe there are better ways to measure how we understand a subject than writing not particularly original essays.
Language Models are disrupting many aspects of education right now, but in the business world, that's something we usually celebrate, don't we? We are positive that the brightest teachers will find out how these tools can help strengthen, and deepen knowledge, sooner than later.
And this is why, overall, Chat-GPT is fantastic: because it offered people from all walks of life a tool to understand where we are with this technology, and where we could be in a few years from now.
We're lucky in the sense that Open AI is kicking the hornet's nest, or letting out the genie from the bottle - even if it's a highly supervised genie with no knowledge about anything that happened since 2021.
How can my business get maximum value from Chat-GPT and similar tech?
We would really like to emphasize, that whatever you do: don't panic. You will not get 'left behind' in the blink of an eye. You don't need to 'implement strategies' immediately. It's important to stay grounded, sit back, and watch. Eagerly, but patiently.
Whatever these new tools are capable of, the ones already in development will be better.
A good tactic would be to research the use cases that already exist, and ask the question:
- What processes are there in my organization which could benefit from this solution?
You should not stop with Chat-GPT though, as there are plenty of interesting free generative tools:
- Dall - E and Midjourney can create images from your descriptions (prompts). (You can ask Chat-GPT to provide detailed descriptions for your Dall-E prompts…)
- There is a tool specifically for generating stock images
- It's practically impossible you haven't yet seen the digital avatars created by Lensa in your social feeds
- Cleanvoice automates podcast editing
- Krisp automatically cancels background noise during virtual meetings
- You don't need to look for royalty-free music for your content: with Beatoven, you can just ask an AI to compose one.
- generative video is already here, check out Synthesia
We could go on, and again, there are regularly updated databases out there, listing these tools.
After a short research, you can start brainstorming solutions, and how these, or similar tools could enhance the efficiency of your business.
It's useful to assign an 'AI hero ' within your organization, someone who is passionate about the subject, and who can manage this process. (We wrote about how companies can use AI successfully here.)
Once you have a list of ideas, you can reach out to ML Ops professionals more confidently, and start discussions about viabilities, development costs, and time frames.
A few examples of our similar projects in the past couple of years:
- we helped to build an AI assistant that gives you feedback about your performance in an online meeting, in realtime, so you can adjust on the fly
- we streamlined and automated the Job Search process
- we've created an automated archive of news articles, scraped from major news portals, with categories according to the client's needs
- we've built an image recognition and claiming process app for an insurance company
Chat-GPT, and the huge array of new AI tools coming out, offer you a chance to see the capabilities of Language Models and other AI solutions in practice.
Through API connections, cloud computing, scraping, and other tools, you can use these examples as inspiration to boost business performance and efficiency!
We hope you've found some interesting angles in this article. Feel free to discuss it in the comments, and contact us if you have more questions or ideas about generative AI, Machine Learning, or anything else!