ChatGPT Review — What You Can Do and Can’t Do with this OpenAI App
I know I’m super late to leave this review. I’ve been reading and hearing praises of ChatGPT and also negative reviews, but mostly positive reviews from all around the world for quite a while now. You must have had similar experiences yourself. While I have been enjoying playing with the application, I must say I did learn a lot.
The Goal of ChatGPT:
Know that ChatGPT is a language model that is meant to engage the user in text-based conversations of different kinds, such as having empathetic conversations or answering questions asked by the user or fetching examples or explaining theories or even writing code.
Writing code and writing papers are some of the superpowers of ChatGPT.
ChatGPT is a language model trained using the GPT-3 architecture, which is designed to process sequential data like natural language. A huge volume of data, specifically, a dataset called Common Crawl was used.
Speed:
The speed is nothing less than impressive! While typing out long answers for you, occasionally, it pauses for a short while, but that’s barely a few seconds. There really is not much to complain about.
On the average (especially when we are having conversations only in English), within milliseconds, you can get appropriate responses.
Handling Multiple Languages:
ChatGPT is trained to handle conversations in several languages. How many languages can this AI handle? According to Open AI, ChatGPT has been trained in languages such as English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean and more. And just like you can guess, like most other conversational AIs, ChatGPT can also deal with many but not all Indian languages.
The Indian languages that ChatGPT can handle as of now include Hindi, Tamil, Telugu, Marathi, Bengali, Punjabi and Urdu. Well, as you can see, some other Indian languages are missing.
Grammar and Grammatical Mistakes:
The grammar is almost flawless but not entirely flawless. In rare cases, where you’d expect it the least, grammatical mistakes happen. For example, in one case, while talking about a mathematical method, ChatGPT unnecessarily used an article “the”, but that is understandable; since such topics do not commonly come up in conversations, I’d not expect trainers to devote enough time and attention to this.
That is our experience with English, but what about language generation other languages? When big companies, especially those created in well-developed countries boast of being able to handle several languages, evidently, they mean languages that are commonly used in those countries. That is why I was curious to test ChatGPT’s language generation in a minority language, and I chose a language I am familiar with: Tamil.
While I wasn’t surprised that the understanding was pretty good for simple, commonly asked questions, that the AI caught exactly what my question meant, the output or response wasn’t up to my expectations. Also, some other prompts were clearly misunderstood. The first time, ChatGPT understood this prompt (“What does artificial intelligence mean?”), but the second time, this was misunderstood, and the response was along the lines of “machine learning”, not “artificial intelligence”. There were many grammatical mistakes, too, for whatever reason, the output took more time, and answers were often not complete.
Sentiment Analysis:
Sentiment analysis is almost perfect, too. When you share how you feel, the AI gets it, and gives you an appropriate response. But this is a skill that has already been mastered by many models, so I’m not surprised to see this working well in ChatGPT.
You Can Automate Boring Tasks:
Yes, the very purpose of creating and inventing new applications, technologies. and machines is to make life easier for human beings, and that is the goal of AI, too. You can use ChatGPT to automate boring tasks. You can give a prompt, and if your prompt is clear, chances are high you’d get a reasonably good response. (There could be some errors, and once again, just paste the error and ask for help, and ChatGPT would try to solve the problem for you.)
Not Trained to Answer Subjective, Opinion-based Questions:
Humans seek acceptance and also want to be challenged in a positive way, which is why we connect better with people who are “like-minded” or share “similar interests” or have “similar values” or people we consider either “equals” or “inspiring” and we avoid people we consider “inferior” or “worth less”. This is human nature across the world. This is how we have been for long. All this put together makes us opinionated and subjective. (Learning to be objective and not judge with our feelings is not something we can naturally do; that’s something we have to consciously learn to do.)
But with a conversational AI, which side should it take? Should it be conservative or liberal or communist? With anything technical, you get what you train it to be. Technical products don’t acquire opinions of their own because they do not have brains to either want or think such ideas. But you can train them anyway you want, to support the right-wing or to be leftist. But why would one want to do it in the first place?
The goal is to share information, being as objective as possible, and not to “please” anyone or everyone emotionally, so yeah, it’s right to choose to not train a conversational AI to offer subjective opinions. This again is common for most conversational AIs.
Limitations:
The only form of communication is through text:
ChatGPT is an entirely text-based experience. Both the input and the output are in the form of text and only that. You can’t upload files; you can only express your thoughts or opinions or ask questions in the form of text messages. Sadly, there is not even audio. I’d have preferred that even they wouldn’t accept voice input, they’d at least have audio output so that I don’t have to keep my eyes open all the time.
Limitations of ChatGPT as a Language Model:
While its capable of all this, it is a language model at the end of the day. A language model, at its best can only understand what you are saying or asking for or expressing, process your input, and give the most appropriate response as the output. And that, ChatGPT does really, really well.
What you should keep in mind is that at the end of the day, a language model is only a language model.
Translation to Indian Languages Takes Time:
ChatGPT’s Speed is more than decent when it comes to conversing in English, but when we expect the AI to generate responses in Indian languages, the waiting time is much longer.
Natural Language Understanding of Foreign Language Prompts Can Also Be Hard for ChatGPT:
In this example, ChatGPT fails to understand simple prompts in Tamil. This prompt translates to, “(my) friend, tell me a Thirukkural”. I don’t know why exactly a prompt as simple as this failed. Thirukkural is an ancient scripture or work of literature in Tamil. It’s also popular all around the world for how old it is, for its worth as a brilliant piece of literature.
Is it an entity recognition failure or an intent recognition failure? As an NLU trainer myself, I’m curious.
Incomplete Responses When Prompts in Foreign Languages are Used:
Not only is this a grammar issue, that a sentence is left incomplete, but this is also a failure to some extent, as the AI can’t complete the task. Not just once or twice, but often, I noticed that responses were left incomplete.
Also, answers were way too shallow compared to answers given in English. That is understandable because training in a different language for use cases is an entirely new process. But yeah, I’m happy that they have tried to include Indian/Asian languages as well.
Understanding Context:
This is one of the more difficult challenges in conversational AI. There are many different factors that come into play when it comes to making an AI understand the context and give a relevant response. One, picking the best algorithm to do this job, two, training it with enough data, and finally, three, storing conversation history in its memory for the entire conversation, from beginning to end, and the resources this would require, the cost!
In most cases, ChatGpt understands context really well. I tried having long conversations on different subjects, and when I asked what some words mean, ChatGPT could understand the context and give me the meaning of the word in exactly that context, instead of giving any other unnecessary, out-of-context information.
Can’t Handle Tricky Questions:
But in some cases, it can easily miss the context. For example, take this case where the user asks the AI a tricky question, and we can see that the AI does not understand context that well.
Probably, the AI is trained to understand context only with the data it is trained with, and therefore, it cannot understand context in different kinds of user input.
Handling tricky input and “understanding” all kinds of questions and expressions is quite a challenge for most AIs, so this is not something unique about ChatGPT.
It is not updated with daily events around the world:
You might be disappointed if you ask about the latest events or events that have not been documented properly in human history. While ChatGPT is extremely knowledgeable, that is, it’s been trained with a vast amount of data in several fields, it might not have updates or adequate information in all areas of all fields.
Incorrect Responses and Misleading Information:
The makers acknowledge that sometimes, the AI can make mistakes. Come on, if we are looking for something that would never make mistakes, we’re expecting God to chat with us. Maybe you should go to a church or mosque or temple if you want that kind of conversations! ;)
It repeats answers unnecessarily:
More than a few times, I saw ChatGPT giving me the same answers over and over and over. This is a frustrating experience. Again, where or when to stop and knowing how much information is expected are extremely subjective. This is quite difficult even for many of us humans, who have centuries of experience with language, to understand.
Well, when it repeats long answers unnecessarily, you can stop, no?
You can’t interrupt using natural language:
You have to hit the “Stop Generating” button for it to stop “chatting” with you. Here, it does not understand, rather the app does not keep itself open for any kind of input/interruption anywhere in the conversation using language as a tool. But keeping that option would do more harm than good in case of an app like this. Basically, trolls can’t waste their time on spiraling chats to take a conversation anywhere, anytime. In this case, the app has more control of the conversation than the user does. This is a restriction that makes sense.
New Features and Pricing:
ChatGPT was never meant to be free forever. The free trial only helps the makers build a better model with more data and also helps users enjoy the experience and make them want it more. In countries such as the US, paid options for ChatGPT are already available. It won’t be long before paid options would be open in India as well.
“It’s worth noting that OpenAI’s policy and pricing structure may change over time, and the company may decide to discontinue the free version or change the terms of access to it. It’s always best to check OpenAI’s website for the most up-to-date information on pricing and availability.”
Benefits of Paid Version:
As we all know, the paid versions always offer more features. In ChatGPT, the paid version promises higher speed, more frequent updates, higher priority, more relevant responses or better quality responses, dedicated customer support, and access to new features that will be added.
No specific feature has been promised yet, but these are what OpenAI is working on right now, so paid members can expect better results.
- - Improved understanding of context and conversational flow to generate more coherent and human-like responses
- Improved understanding of different languages and dialects
- Incorporation of more diverse and high-quality training data to improve the model’s ability to respond to a wider range of prompts
- Better handling of out-of-vocabulary words and concepts
- More robust and accurate language generation capabilities
What I would use ChatGPT for:
ChatGPT Can Help with Research Work:
Again, since this is a language model, it can’t really reason: it only shares data that it can find, so if there is an abundance of data available for your questions, then ChatGPT can easily summarize the information it finds for you. Here’s an example of a conversation with the new Bing powered by AI:
I am more than happy with this AI. I ask precisely what I want and I get exactly what I expected. Note that this is possible only because the question I asked has a lot of answers across the Internet.
But keep in mind that this “AI” merely only fetches information. You can’t hold the makers of this AI or any other conversational AI in the market right now responsible for giving you incorrect information.
See the example below for example. This ‘information’ is both false and stupid, for these are not biscuits themselves but brands that make different biscuit products, of which some can be vegan and the rest not.
Conclusion:
I’d expect a smart person to know how to use a tool and get the best value out of it instead of complaining about what all it can’t do. If someone would complain why a knife can’t do what a spoon can do, and conclude that the knife is “not sharp enough to do that”, well, you know who is the problem here.
What would you use ChatGPT for? Would you go for the paid version? What do you think you can do with ChatGPT?
