Although ChatGPT is a “hot trend” right now, many people are probably still confused about what this chatbot really is, how it works. While computer knowledge and artificial intelligence are complex and confusing things, especially for those who are not in the programming industry, technology expert Nguyen Hong Phuc – who has many years of working in intelligence artificial – there was an explanation that could not be easier to understand for this emerging AI chatbot.

Not only providing an intuitive and easy-to-understand view of the emerging chatbot, Mr. Nguyen Hong Phuc’s article also provides a fresh perspective on the influence of ChatGPT on OpenAI as well as a series of other giants in the industry. turmeric. Below is the post of Mr. Nguyen Hong Phuc:

What is ChatGPT really? Easy-to-understand explanations for people who don’t know IT

(The content below is written for common people to understand, so the superior knowledge is popularized, so the academic accuracy is not guaranteed)

Easy explanation:

Easy to understand for the average user, it is simply a website to chat and talk about all kinds of topics with a virtual bot – very easy to understand.

This bot by the company OpenAI was co-founded by saint Elon Musk since 2015, initially with a very loud mission to “prevent the dangers of AI” (remember this paragraph, I will bring it up again to troll later) .

Many years ago we were also very excited about a chat bot like this, which is a Simsimi chicken made by Korean guys, this is also an artificial intelligence (AI) bot, it also continuously learns things. that users teach it, so Simsimi Vietnamese bot is currently the best “funny” chat bot in Vietnamese today.

So is the ChatGPT bot, it is constantly being re-taught with new chat content of users, so after 1 month people started to care about ChatGPT in Vietnamese, it started to reply in more virtual Vietnamese.

Before ChatGPT, we had two very familiar chat bots, but most of us forgot about them because they talked too much, Apple’s Siri and Google’s Assistant. We’re like a chat bot command (understand the command and then execute it) but it’s boring to talk, at first Siri talked quite well but later Apple forgot so it gradually got worse and worse.

How is ChatGPT created?

(This part will be heavy on technical information, but I like to write it because I have spent many years researching and implementing AI in loneliness and no one to talk to).

ChatGPT is an artificial intelligence computer program. Professionally, people often call Model AI in Vietnamese, “artificial intelligence data model”, but in fact it is still numerical data running on a computer, so it is not wrong to call it a program.

The word Model AI consists of two parts: Model (Data Model) and AI (Artificial intelligence). Extraction means Intelligence comes from data (standard translation: :)))) infer that there is a lot of data, it will generate intelligence.

Yes, the process of creating Model AI is a process of steps: data collection, data selection, data labeling for training, training.

Basically, it’s easy to teach AI, you need to create a dataset like this

Question: What’s your name?

Answer: My name is ChatGPT

Question: Which country is Vietnam?

Answer: is the country in the east of Laos

After teaching the AI, it remembers this information (training), and then saves the memorized brain of the AI ​​as Model AI (model checkpoint)

Later, when using, load the brain with the memory containing the above information (inference) into the computer, you just need to ask the corresponding question, the AI ​​will recall the taught knowledge and answer “same”. what it is taught”

Well, the basics of AI are the same as above, anyone who does AI knows this case because it’s so easy. This basic AI creation method has been studied and formed since 1950. So why is it that after more than 70 years, the AIs were still so bad until recently and specifically the ChatGPT, it was “unsurprising”?

Actually, over the past decades, AI has been specialized in many specific jobs such as AI supporting aircraft, AI in combat simulation, AI in games… but almost no big companies have invested in it. AI in the field of languages, it was not until 2017 that there was a breakthrough in technology that made AI training more effective than mutants, especially language AI.

Specifically, written language is the achievement that creates human civilization. Humans contain their knowledge in writing, understanding language (writing) is understanding human knowledge, this is the core point of human civilization. core that makes up language AI, which before 2017 it was very difficult for humans to make computers understand the meaning of a meaningful sentence.

So what’s in 2017?

In August 2017, scientists at Google, specifically Google Brain unit, Google’s AI-intensive research unit since 2011, invented an algorithm called Transformer (algorithm name is very similar to robot movie). Michael Bay’s punches).

The Transformer algorithm is very groundbreaking, especially in terms of language AI training. Before this algorithm, humans who wanted to teach AI had to create a training data set with question-answer pairs (labeling data) as mentioned above, and the machine actually only remembers the pair. question-answer rather than “understanding” the meaning of the sentence, there is a huge difference between rote learning and comprehension.

It’s even easier to understand that after 2017 we just need to pour in as much text data as possible, the computer will figure out what it means by itself instead of having to tell them what it means.

Quoted verbatim from google’s transformers publication: “with transformers, computers can see the same patterns humans see”. The translation is not good :)))

Google is very human when making detailed documents about the Transformer algorithm publicly accessible to all. Also provide Open-Source for this algorithm

Suddenly the entire AI science community benefited from Google’s invention. Among them is OpenAI – a company founded in 2015 and did not have any outstanding achievements until after 2017.

After Google announced Transformer, a few months later the first linguistic AIs based on this new algorithm were born.

In January 2018, OpenAI released the first AI based on Transformer, GPT-1, they applied very quickly, faster than Google itself.

GPT stands for Generative Pre-trained Transformer which means “Generative Pre-trained Transformer program”.

This GPT AI was created with the main purpose of “Birth”. Specifically, you will play a word match game with it, you write a sentence, it will read that sentence and then based on the knowledge it is storing in its memory that “generates words” following the sentence you write.

For example:

You enter: Vietnam is

ChatGPT: Vietnam is a country located on the East Asian ocean, in Southeast Asia…

This is what it looks like to be “magical”: you chat a sentence with ChatGPT and it says a sentence back.

In fact, it is not replying to you, but it is playing word matching by “Birth Text” to continue the meaning of the sentence you enter into the chat with it.

GPT-1 is the first generation of ChatGPT. This GPT-1 is a pretty small AI, really small in size and complexity.

In the world of Language AI, people measure the complexity – corresponding to the “intelligent” level of the AI ​​- with a unit called Hyper Parameters, this concept can be roughly explained as this AI understands how deep the meaning of the pile of texts used to teach it is.

To train this GPT AI, the scientists at OpenAI collected a large amount of human written text, mostly from Wikipedia, encyclopedias, major and public newspapers, mass somewhere. about hundreds of GB and several hundred million documents. After collecting, they clean and filter the content. Then bring those texts to the AI ​​to read, force it to read many, many times, each time it reads that block of data, it sees a layer of meaning behind those words, the more times, the more layers of meaning. means.

The more layers of meaning recognized by the AI, the more Parameters the AI ​​has. The AI ​​model GPT-1 has only about 117 million Parameters, GPT-2 (2019) reaches 1.5 billion Parameters, GPT-3 (2020) reaches 175 billion Parameters.

The two AI models GPT-1 and GPT-2 are almost unknown to the public because the text-generation efficiency is not really impressive due to the level of understanding. u The layers of meaning behind the pile of human writing are still too shallow, of course at that time people still didn’t know how deep and effective it was, so the engineers at OpenAI worked hard. teach AI GPT to dig even more layers, until May 2020, the AI ​​GPT has dug up to 175 billion Parameters, the results of Sinh Text now make them even explode when it plays word matching with intelligence. -Knowledgeable as a 10-year-old child in terms of language. They named it GPT-3.

And now it’s time to expose OpenAI

The GPT-3 AI was born, it should have had the same fate as many other AIs after 2017 (a lot of big companies investing in AI like Facebook, Google, IBM, Microsoft also created Language AIs like GPT ), they are all kept in the research room and absolutely inaccessible to the public – ordinary people.

Why are they strictly quarantined?

Trained AIs reach a deep understanding of human written language, leading to a very serious problem that no AI scientist has yet found a solution to.

Calculate “True” or “False” (True or False).

AI can’t understand what is “True” or “False”

AI can see many layers of meaning in a sentence, but cannot “understand whether the meaning is right or wrong”. Because right – wrong is relative, for humans it is fragile and controversial even between humans and humans.

In addition, the huge amount of textual data that scientists at OpenAI collect to train AI is not all “right” and contains information that is “correct” with the standards of human society. people, because the amount of data is already too large for their ability to select, for example they can collect the records that the earth is round, and at the same time can also collect the records that the earth is flat . Data, they contain both true and false information in them. Then when the AI ​​reads and re-reads those texts to find layers of meaning, it also finds both “true” and “false” meanings, but the AI ​​does not have the sense to recognize the meaning. which – which information is true and what is the meaning – which information is false. AI simply remembers it all. When asked later, it simply answered from its memory that information, regardless of right and wrong.

Companies like Google, Facebook, IBM, Microsoft have repeatedly announced breakthrough Language AIs in answering human-typed questions, but quickly deleted that AI. You can find articles about this all over the internet from major newspapers. It’s mostly because the AI ​​that answers some questions is biased towards a meaning of “Wrong” that is unacceptable in terms of current human social norms such as respect for gender, respect for religion, respect for ethnicity, the accuracy of events that have occurred, the truths that people have agreed to be true…

Large companies all follow the standard of information accuracy, they think that the AI ​​can’t solve the perception of Right – Wrong, it is best not to go public. Like if you haven’t taught your child to study properly, lock him at home, don’t let him out on the street.

GPT-3 is the same, it also creates paragraphs that violate the standard of “Right-Wrong” of people, even to the point of unacceptable.

But OpenAI despite the fact that they were founded with the principle of “preventing the dangers of AI in its infancy”, this principle was preached by Elon Musk at the TED technology event when announcing the establishment of OpenAI.

They ignore the danger when the AI ​​GPT-3 creates false passages.

They are the first AI company to provide API access to GPT-3 AI to the public, just pay and use it. Something that no major tech company has offered until now.

They commercialize an uncontrollable AI of “True – False”.

The press at that time was also quite interested in promoting their AI GPT-3, and other small and medium companies in the market were also excited to apply GPT-3 to technology products.

GPT-3 was on the verge of becoming popular, when the global Covid-19 pandemic broke out, the epidemic situation became more and more tense from mid-2020, the pandemic information stream engulfed information about GPT-3.

The AI ​​GPT-3 and OpenAI were forgotten by the public until the end of 2022. OpenAI decided to do a marketing game to see if it could revive interest in Language AI anymore?

So they tweaked the GPT-3 AI to ChatGPT, making it easier to use, instead of coming up with the look of a web page where people type text, edit parameters, and then get back a text that connects the word. , then ChatGPT comes in the form of a Chat program, with a chat box to enter a question, the AI ​​ChatGPT plays a word generator game with that question, but in the form of an answer.

Just a small change in UI/UX but AI becomes much easier to communicate with.

Fortunately, they succeeded, they revived the public’s curiosity about AI, pushed the public’s imagination towards AI, and formed a clear image of AI in the public’s mind. “a robot that answers every question the user asks”. In just 1 month, everyone was talking about AI, and AI became the equivalent of ChatGPT.

Excellent, marketing only needs results, nothing more.

Summary of ChatGPT’s success formula in the past 1 month: A Language AI trained deeply enough to produce meaningful sentences that convince readers + the moral defiance of an AI technology company + Appropriate UI/UX (Chat) = ChatGPT.

Currently, the success pressure of ChatGPT OpenAI is forcing big AI companies such as Google, Microsoft, IBM, Facebook, to accept lowering the ethical standards of the industry, to keep up with the race where they are suddenly left behind even though they are Possessing breakthrough technologies and huge computing power, they are releasing AI language models for dialogue for public use. Google this March will release LaMDA (2021), Microsoft next June will release MEGATRON (2021) which is a giant bot with 530 billion parameters…

So we already understand what ChatGPT really is, but can ChatGPT be used in everyday work, or even threaten to replace the work of many people? Wait for the next one, and here it comes :)