Warning
This article is just my personal opinion about ChatGPT – a hot AI application in recent times. However, this article will have some unpleasant arguments about it. Contraindicated for diehard ChatGPT fans or groups of thinkers who are dreaming of a future where AI can replace humans.
Who is the father of ChatGPT?
Perhaps in recent days ChatGPT is a hot word more than any technology issue published so far. Just looking at the chart of the time it took to reach 100 million users is enough to see how great its appeal is. It has surpassed even the world’s leading entertainment application, Tiktok, to rise to the top position in the list of technology applications reaching 100 million users after only 2 months of release.
The success of ChatGPT makes people curious about its father – OpenAI – a startup in the field of artificial intelligence created in 2015 with the mission to prevent unsafe, dangerous AI technology . One of the notable points is that this company was co-founded by one of the cult and equally scandalous figures in the technology world – billionaire Elon Musk. Although now this billionaire has announced that he no longer holds shares in OpenAI. However, the influence of this scandalous owner is still forever. I will have more detailed analysis in the following sections.
It can be said that from its founding in 2015 until the end of 2017, OpenAI has no reputation in the technology world. That was the time when neural network models based on CNN were on the rise and its most popular application was, of course, Computer Vision. At that time, people did not really pay too much attention to AI. Frameworks are also very young when only Tensorflow 1 and PyTorch did not appear until September 2016. In terms of academic research, OpenAI was even less important than the big research labs of that time like Google Brain or Facebook AI .
But 2017 was really an important milestone and it was the beginning of all the stories to come… Hey, are you still reading, let’s move on. Something good is about to happen.
What breakthrough does ChatGPT have?
Why do I mention the 2017 milestone above, also reminding a bit about the previous times, when the application of NLP was not noticed by most companies except for a few large research labs. Partly because the hardware of that time was not feasible with very large amounts of data, partly because the previous neural network architectures and models were not really powerful enough to model this type of data with so many constraints. hidden in semantics, context like natural language. However, the 2017 milestone is really memorable for those working in the field of AI in general and in natural language processing in particular, when a new divine neural network architecture named Transformer was born. This architecture was invented by a Google research lab, Google Brain, and then announced for the whole world to inherit and use it. Since then, Transformer architecture has really stormed all over the field such as natural language processing, speech processing, image processing… The biggest advantage of Transformer is being able to enable AI models learn based on huge amounts of data without the need to manually label questions – answers like before. The Transformer itself and its training help the model to understand the hidden features in the data instead of just memorizing it mechanically and passively. That is to say, since the time Transformer was born, the training of language models is much simpler, we just need to push a lot, a lot of words into the model so that it can find the features on its own. data representation only.
Going back to OpenAI, it must be said that they are more of an application development company than a research company. They quickly applied Transformer’s achievements and quickly produced all kinds of birth models as you can see in the picture above. These model families are called GPT – Generative pre-trained transformers . Right after Transformer was born in 2017, January 2018 -so fast and furious , OpenAI released the first GPT-1 model for language generation. The architecture of GPT inherits a lot from the original Transformer model and it has gone through versions from GPT-1, GPT-2, GPT-3 and now ChatGPT. With each improvement, this model is like a child learning more and more knowledge, along with the body is also taller and fatter. That is demonstrated in the number of parameters of these models increasing rapidly. You can see it in the chart below.
Whether OpenAI has any breakthrough in developing their language generation models, I personally think not. They benefited from Google’s invention. Inheritance and development are common in the scientific world, which is not a bad thing. But the way they approach these language models is really disturbing. Why do I say that, please read, let’s analyze together.
Why has it been announced only now?
Although there are so many advantages, language models still have a huge disadvantage that no scientist or company can solve, which is the correctness/falseness of the answer. That is, language generation models can produce content that is completely wrong from a human’s moral point of view. And indeed, OpenAI’s GPT models will also share the same fate with many language-generating AI models born after 2017. A lot on the internet you can read news like Facebook’s Chatbot being exploited. death for chatting with each other in a new language or a racist Microsoft Chatbot . That is the reason why Chatbots of big technology companies soon bloomed so quickly .
Going back to OpenAI, the GPT-1 and GPT-2 models are Open Source models, but with their small body, they have not really given a lot of buzz. The answers given by GPT-1 and GPT-2 are sometimes very silly and funny, but they are Open Source models and change when served for entertainment purposes, so it doesn’t matter to anyone. It. OpenAI continues to increase the number of parameters and train more real for their GPT models and until the GPT-3 version is released in 2020, they go in the direction of close source.
Despite all the disadvantages on a GPT-3 chatbbot, they still defied it to release a commercial version of GPT-3 as a text-generating API. I have to use the word defiant because they know that GPT-3 still has many shortcomings, violating the ethical standards of the industry. They defiantly make money on something without control over the information generated. This is really scary.
When GPT-3 was released, there was also a lot of talk about it as a truly intelligent and almighty AI application. While in the middle of the wave of cheering, the COVID-19 storm hit and swept them all away. Take the dreams of an all-powerful Artificial Intelligence of GPT-3 and OpenAI to the remote island.
If it weren’t for COVID-19, we would probably have experienced ChatGPT 3 years ago.
ChatGPT-3 – new bottle of old wine?
After the COVID-19 wave passed and everything returned to a new normal, it was time for OpenAI to return to their ambitions and once again they defied everything . But this time, they were really smart when they released a Facelift version of GPT-3 but combined with the engineering techniques in making chatbots to create a new thing called ChatGPT and surprisingly, the very way Conversational approach and friendly interface make ChatGPT easily accessible to a large number of users. That’s why ChatGPT became extremely popular. But really, it’s just a new bottle of old wine . Up to this point, I really see that OpenAI is not seeming to follow the principle of preventing the dangers of AI from infancy as they set out when despite releasing products that are below its ethical standards. branch. Are they also influenced by a certain eel family character?
The Deadly Disadvantages of ChatGPT
Many people will think that I wrote this article to get a good view, there will be opinions saying that he can do it. I’m sorry I don’t care about that. What I want to share here are more correct views from a technical person and working in the field of artificial intelligence. ChatGPT is a big language model, that is undeniably, and in itself has the following inherent disadvantages:
Squeeze people with meat
There is no official figure on the number of parameters of ChatGPT, but with the inheritance from GPT-3 with 175 billion parameters, this model is also in the form of a giant giant . In 2020, GPT-3 is the largest language model ever trained, with 175 billion parameters. It is so big that it takes 800 GB of memory to train it. This year, it was overtaken by BLOOM (BigScience Large Open-science Open-access Multilingual Language Model) at the top of the list among the largest models) with 176 billion parameters. This is really a hardware race when it takes 40 GPUs of 24GB type just to accommodate the model. Not to mention that with a huge amount of training data, language models like GPT-3 require huge amounts of computational and hardware costs. I am not understanding the GPT-4 model they are intended to have for a billion billion – triliion of parameters, which is 7 times the current model, how terrible the calculation array will be. This is really a job that takes meat over people that any fledgling technology company has no chance in this race.
Data is not “clean”
It is certain that OpenAI will not be able to have enough time to use humans to read and verify all the data fed into the language model because that is not possible. It also goes against the criteria of developing models by automatically finding hidden features from unlabeled data of language models like GPT. Therefore, when we use ChatGPT we need to accept that we are talking to a training machine from a bunch of mixed data and there is nothing to verify its correctness. And don’t forget a classic principle in computer science Garbage in, garbage out – don’t expect a good output with a pile of garbage as input.
Slash the wind to become a god
Back to the story of right and wrong , which is inherently a very difficult category to distinguish even for ordinary people. And surely more than once you will receive ChatGPT’s answers that are very natural but completely bogus . This is probably the biggest weakness of a language model because it is dangerous to give information that is not real but so natural that it makes people believe it is real.
Update very slow
As mentioned above, training language models like ChatGPT is extremely expensive that’s why it can’t respond to news like it’s simply data. the data is past and not new anymore. Perhaps it is not until there are billions of times more advanced computing technology like quantum technology that it will be possible to help language models like this become real-time updated.
Replacing search engines like Google – still far away
As mentioned above, it’s not that the big guys are stupid enough to let a fledgling startup like OpenAI surpass you. They have huge resources, big data sources, so there’s no reason why they shouldn’t think of developing things like ChatGPT that OpenAI is making. It’s just that they understand that their own children still need to be taught better before taking them out on the street so that they can say whatever they want, if people come home to curse their parents and curse their parents. how to go up.
By itself, giving a single answer like ChatGPT is doing makes the information completely one-sided and not diverse and does not give us the right to choose and learn. It is the skill of searching and digging into the results that gives us the experience, the experience, not the question and answer with a Chatbot that the answers themselves do not know whether it is right or wrong. Then, in the end, we have to search and verify the information ourselves.
Is it a technology goal or a financial one?
Personally, I think OpenAI is a technology startup, so their goal is still revenue and profit. They are different from the scientific research laboratories of other large corporations in that other labs have their own budgets to focus only on developing new technologies. It is because of their financial goals that they defyly offer products that are below industry standards. However, this is really a successful move for them. We can see positive financial moves such as:
- A lot of big enterprises invest and OpenAI and other AI companies
- A lot of stocks of companies working on AI flew to the roof
But certainly with the development and deployment platform is too cumbersome, perhaps ChatGPT needs a lot more improvement to be enough to make a big profit as OpenAI expected. As for investors, they probably have their own opinions about this technology rather than the excessive FOMO like today.
From the perspective of a technology person, I would like to share a concept called Hype Cycle . Accordingly, any disruptive technology must crawl through this curve, including ChatGPT:
- The first is the period of birth
- Then there’s the hype period
- Then comes the period of disillusionment, enlightenment, development, and sustainability
- And finally come to life.
It can be seen that applications like Computer Vision are almost complete and come into reality
How about our ChatGPT. Of course it will also have to go through periods like that. Here is a picture of the future of ChatGPT
ChatGPT or “Dead GPT”?
Imagine what the world would be like if all the information on the internet was created by one hand ChatGPT . We will no longer be able to distinguish what is right from what is wrong and that is when it will engulf us in the sea of misinformation it creates. As a technology lover, I strongly advise you to have a sober look and make the most of ChatGPT. Moreover, there is always doubt and double check with the answers from it.
Don’t let ChatGPT become “dead GPT” for each of us.