Why is ChatGPT such a big deal?
Many of us remember spending our elementary school days chatting with a robot (only years later did we find out it was in fact called a chatbot), not fully understanding how it works. But, after a while, it would become boring because there weren’t countless possibilities. It’d come up with funny answers at first, but would just repeat the phrases as the communication continued.
If we look back, it’s actually fascinating how much this technology has developed over the years. Not only that, but it became so widespread that it’s considered a necessary tool for any business.
But after all this work, is it flawless?
Let’s start by saying that’s the ultimate goal of OpenAI releasing ChatGPT in a test phase, available online for free. But little did they know that the interest will be so huge it’ll regularly crash their website.
As mentioned, creating a dialogue-optimized AI model isn’t new. But training them with human feedback is a prerequisite to improving evaluation and enhancing its application in the real world.
And that’s exactly what they did – they embraced the reinforcement learning from human feedback (RLHF) to its full potential.
That’s where it gets interesting.
And let’s not forget the fact that OpenAI is a tech startup co-founded by Elon Musk. Now, let’s get down to business.
How does ChatGPT work?
As compared to the previous versions of OpenAI GPT-3 speech, they’ve included human-written dialogs in developing ChatGPT as well. In other words, apart from using AI assistants, this San Francisco-based company tested out the alternative where both sides of the dialogue were written by humans in a natural, everyday manner.
After these conversations were recorded, the team used this data to make AI-generated responses with alternative autocompletion and asked users to rate the accuracy and flow from best to worst.
For this purpose, OpenAI used its Proximal Policy Optimization (PPO) algorithm released in 2017 and which has proven its performance when it comes to fine-tuning.
It can go from solving coding tasks to writing unique poems, while still keeping that witty, quirky companion role. But its main function is to help you save time in your everyday tasks – both personal and business – optimize your workflow, and help you dig out that perfect piece of information you’ve been chasing in a never-ending internet forest.
The team admits that this large language model isn’t perfect yet, but promised to keep an eye on users’ feedback and constantly work on improving it. Either way, ChatGPT does make stuff up sometimes, but it admits its mistakes – which is another unique feature.
Even though there have been many tryouts in this field, including Google’s version called LaMDA, users were always a bit skeptical of how accurate these responses are. What ChatGPT is showing, is that we’re getting as close as possible to eliminating this suspiciousness around chatbots.
However, the OpenAI team is aware that the ideal scenario would be if the chatbot asked additional questions when given ambiguous inputs, as opposed to what it does now – which is guessing what the user intended to ask.
This leads to conversations running in circles instead of narrowing it down to a correct answer. Either way, ChatGPT is the most advanced chatbot version we’ve seen so far, even though it’s not ideal yet.
Hopefully, after finishing the test phase, we’ll be able to enjoy the benefits of this new technology – but keep in mind it won’t be free anymore after it’s put into use.
Imperfections of ChatGPT
A lot of thought has been put into ChatGPT; there’s no doubt about it. It’s trained to be especially cautious with hate speech and giving misinformation. Although we all agree that’s the ultimate priority, being overly cautious has one huge disadvantage – avoiding questions.
That wouldn’t be such a big deal if your sole purpose wasn’t exactly that – providing users with answers.
We’ve analyzed users’ feedback all around the web and came up with a list of the most common faulties of ChatGPT:
- The response changes significantly even with the smallest change in inputs.
- It requires users to notice the mistake in inputs if the response is incorrect.
- It sounds confident while giving completely incorrect responses.
- It often gives repetitive, wordy, non-sensical responses.
- Users have a feeling that ChatGPT is over-optimized.
- It doesn’t reply to questions related to current events or people, claiming that it doesn’t have any knowledge outside of training data.
The latter may seem like a serious inconvenience, but OpenAI chose this for a reason. Should we remind you that Deepmind’s chatbot was initially designed to use additional internet research for its responses but the team later decided not to release it for security purposes?
Lastly, if you’ve already given ChatGPT a chance, or are planning to do so, keep in mind you could earn a reward if you provide constructive feedback. More precisely, there’s an open ChatGPT Feedback Contest, where you can earn up to $500.
On one side, this reward system clearly shows that the OpenAI team is sincerely interested in improving their technology according to users’ preferences. On the other side, this reward system may be the exact reason why so many people rushed to try it out, building ridiculous hype around chatbots, for the first time ever, and earning them a million users in less than seven days.