Ketika manusia menciptakan sesuatu yang menyerupai dirinya sendiri

Okay, where do we start? What if it's a dream? There was an engineer who dreamed that the machine could talk by itself. Before there was a CGPT, before there was a robot, and even before there was a computer that used GUI. There was a man who was wandering and asking a simple question. Could the machine think?

His name is Alan Turing. He is not just a computer scientist. He is the one who created the Turing Machine, a theoretical blueprint of how computers can work. Not a physical computer, but a mathematical model that shows that a simple machine can run various kinds of algorithms as long as the correct instructions are given.

This idea is the foundation of all modern computers. And he also helped develop the Bombe, a physical machine used to break the Enigma encryption code used by Nazi Germany in World War II. Estimated, Turing shortened the war by 2 to 4 years and millions of lives were saved.

In 1950, Turing wrote a paper entitled "Computing Machinery and Intelligence". He didn't say that machines can think. What he asked was how we can define the word "thinking". That's where Turing's text was born. The idea is this, we talk through text with two entities, one human and one computer.

If you can't differentiate between computers and humans, then the computer is considered thinking. But that doesn't mean the computer is conscious. That's because we don't have any other way to measure it. Unfortunately, the story of Turing is not a happy ending. The person who helped save his country was punished by his country itself because he was... well...

G. He was given a choice between prison or chemical castration. And he chose chemical castration. Then two years later, he died of cyanide poisoning. Two years after Turing died, a group of scientists are in a dark mode in Dartmouth College, New Hampshire. They are not a bunch of people. There is John McCarthy from MIT, there is Marvin Minsky, and there is Claude Shannon. Yes, they are badass.

They brought proposals that were more or less like this. They plan to investigate the creativity of making for 2 months with 10 people and are quite sure that they can make significant progress. Only 2 months and only 10 people in the 1950s when the computer was still like this.

But don't be judged yet, America just came out of World War II and is in a huge euphoria of technology. Computers are just born, jets are just there, so people think there is no such thing as "unbelievable technology". And from here, the word "Artificial Intelligence" appeared.

McCarthy chose that word. Why AI? Why not machine intelligence or something else? But that word is still stuck today. After Dartmouth, optimism was no longer controlled. In 1970, Marvin Minsky said that in 3 to 8 years we will have a machine with the same human intelligence as the general public. Herbert Simon in 1965 also said that machines will be able to do anything that humans can do in 20 years.

These scientists are absolutely geniuses, but they can't differentiate between computers that can play chess and computers that can... ...scratch clothes, for example. And in 1966, a scientist named Wieschenbaum at MIT made a program called ELIZA, one of the first chatbots in the world. Wieschenbaum made ELIZA precisely to show how profound the illusion of conversation with machines.

But what happened was that people believed that Eliza was aware. Wizenbob was surprised. It turns out that many humans were deceived by it. What he did was not AI, but an illusion. This is the first warning that it looks smart and true intelligence is different. But in that era, not many cared. So money was flowing for research and expectations were growing.

On paper, everything looks promising, but there are fundamental problems that no one wants to admit. Computers in that era were not strong enough. Not a little weak, but not really strong. For example, translate a sentence from English into Russian using a computer system.

That's an unreasonable calculation for computers in the 1960s. Scientists call it the combinatorial explosion. The more complex the problem, the more the number of calculations exploded exponentially, and computers at that time were very incompetent. The US government has funded millions of dollars for the Russian language automatic translation project, not an academic project. This is for military advantage in the middle of the Cold War.

And for several years, researchers said it was almost possible or just a moment until finally in 1966 the government formed an independent committee and produced the Alpark Report. In conclusion, it doesn't work, it won't work in the near future, and continuing funding is a waste. Finally, the funds were cut almost overnight.

But the most ironic comes from the AI community itself. Marvin Minsky, who is most ambitious about AI, together with Seymour Pepper wrote the book Perceptron in 1969. There, they mathematically proved that the neural network in its current form cannot solve certain problems.

Technically, their conclusion is true, but many AI communities misread it. From one layer of the neural network has limitations, it immediately becomes all neural networks are useless. Finally, the funds were cut and researchers stopped believing. Then in 1973, the British government asked mathematician Sir James Lighthill to independently evaluate the AI research. Lighthill is not an AI investor, but that is what makes his evaluation considered objective.

As a result, the AI research for two decades did not produce anything promised. Hearing this, the British government immediately cut funding, the lab was closed, researchers were unemployed, and the US government was in support. The term AI has become toxic. Researchers who are still interested in this field have been replaced by a machine learning,

pattern recognition, whatever, as long as it's not AI. This first winter is not just about technology that is not ready, but also about the distance between promises and reality that is too far. Researchers are trapped in a rather dangerous cycle. To get funds, you have to make a big promise. To keep getting funds, there must be progress. When the progress does not meet expectations, the government will just say "omong-omong" and this pattern will repeat.

After a difficult time, researchers who are still holding on try a different approach. Instead of making AI learn from scratch, how about putting the knowledge that is already in the computer? Become an expert system. This is basically simple. Intelligence can be made by inserting human rules and logic into the computer. If the symptom is A, then the disease is B. If the condition is C, then the decision is D. Humans who write the rules and the computer runs it.

All AI before had the same way of working. Humans make the rules, then computers just follow and expert systems are the key to that. Meanwhile, there is Hinton who believes in total opposition. Not giving rules, but giving examples. Let the machine find its own pattern. In the 1980s, this sound is not reasonable. And that's why Hinton is considered strange.

But remember this person's name, because we will discuss it later. One of the most famous expert systems examples is Mycin from Stanford. The purpose is to diagnose bacterial infections in the blood. As a result, Mycin got a 65% score in blind tests, while the specialist doctors got 42-62%. In this very specific domain, Mycin is able to beat the real doctor. It was implemented by Xcon in 1980. Its task is only to check the customer's customer order computer components.

compatible or not. The result can save $ 40 million per year. After those examples, there is no more debate about AI value in business. In the early 1980s, the AI industry grew into a billion-dollar business. Then in 1982, Japan announced the 5th generation of

project with an estimated $ 850 million to make a computer that can think and understand human language. America and Europe panicked, so they immediately made a competition program. And this can be said to be the first AI race, long before OpenAI, Google, Anthropic, and others. But behind all that optimism, there is a problem that grows slowly.

Expert system can only work in a very narrow domain. Don't ask MySyn about other diseases, he doesn't know. Don't ask Xcon about other things, he doesn't know either. While the world keeps changing, regulations change, policies change, but the smart system can't learn by itself. Every change must be updated manually by the engineer and often make new errors.

Just like the previous cycle, hype began to meet reality. In 1987, one part of the AI industry began to collapse. Lease Machine. Lease Machine is a special computer made to run AI programs, especially for expert systems. The problem is that ordinary desktop computers from Apple or IBM are developing

very fast and started to be strong enough to run many of the same programs, but with a much cheaper price. So in a short time, the business leaves the machine collapsed. Not because the AI suddenly becomes useless, but because the business that is built around it is not sustainable.

And the 5th generation project in Japan that ended in 1992, 10 years with a budget of 850 million dollars and the result didn't change anything. Can't understand the natural language and can't adapt. The project is over without any revolution. With this, the second winter officially started. If the first winter is like people who are disappointed because

with excessive expectations, the second winter was more like, well, it was a lie again with the hype. But there is one person who didn't give up. His name is Geoffrey Hinton, an English researcher who, from the early 1980s, was sure that the neural network was the right way. During the first winter, he kept working. When everyone moved to the expert system, he kept working. During the second winter, he kept working.

He moved to University of Toronto, partly because it is more tolerant of unproductive research. Neural Network needs three things: enough data, a strong computer, and how to teach the neural network itself. For how to teach the neural network, the answer is in a technique called Backpropagation, which allows the neural network to automatically correct its own mistakes.

Hinton is one of the most popular and developed techniques for the modern neural network in 1986. Unfortunately, no one cared about it at that time. But he was sure that one day all the fuel would be available, so he waited while he continued to work.

While Hinton works alone in Toronto, two things happen in the world that seem to have no connection with AI. First, the internet was born and grew very fast. Suddenly there was a surge of photos, a surge of texts that continued to increase every day. No one planned the internet as a data collection machine for neural networks. But that's what happened. And this is really important, because the hardware alone will not be enough. Neural networks need to learn from examples.

the data is large, he doesn't have enough material to learn. Second, more slow. Every two years, the transistor doubles, the computer becomes more powerful with a lower price. With this, the game industry is booming, and Nvidia aggressively develops GPUs for graphic gaming. Deep learning researchers then realized that the way GPUs work with thousands of parallel calculations turned out to be perfect for neural networks. Neural networks are basically millions of mathematical operations

that can be worked together, just like what the GPU does for graphics rendering. Nvidia doesn't design a special GPU for AI, but without Nvidia's GPU, the revolution that will come will not be that fast. Besides Hinton, there are two other names that need to be known. Yan Lichun, a former

who originally thought he was Chinese, developed the technique of Convolutional Neural Network, a foundation of how computers can see pictures. In the early 1990s, the technology was already used to read handwriting on checkbooks, and Yoshua Bengio, who focused on teaching computers to understand human language.

Hinton, Lee Chun, and Beng Yeo, these three people will later be known as the Godfathers of Deep Learning. They worked in an era when almost no one cared, at the Canadian University which is far from the Silicon Valley hype. In 2006, Hinton published a paper with his student Ruslan, showing that deep learning can be practiced effectively. But unfortunately, there is no big news about this. But for those who follow closely, there is a feeling that something important has just happened.

Every year there is a competition called ImageNet Challenge, a video classification competition. In previous years, the best team had an error rate of around 25-26%. That means out of 100 images, 25 of them were wrong. Then in 2012, one team entered with a different approach.

AlexNet, made by Alex, Ilya, and Hinton. The approach uses deep learning, but what makes it different is the way it is trained. Using two Nvidia GTX 580 GPUs, not a supercomputer, not a special hardware, a gaming GPU that is usually used by gamers. With GPU, the training that should be months, is finished in just a few days. The result, error rate around 15.3%, runner-up 26.2%.

The AI research community was shocked. In one night, everyone moved to deep learning and everyone started buying Nvidia's GPU. Google immediately acquired a startup owned by Hinton for $44 million. Not the product that was bought, but the person. A year later, the acquisition was played for $600 million. Facebook brought Yan Li Chun as the head of the LPR.

The AI researcher's salary exploded, and university professors immediately went out to Silicon Valley. In 2016, DeepMind challenged Lee Sedol, one of the best Go players in the world, with an AI named AlphaGo. Before this, DeepMind was already known for making AI that can play Atari games only from pixels, without being given rules, only given rewards if the score goes up. AlphaGo is a much more ambitious version of that approach. Oh, by the way, Go is a board game

who is 2,500 years old, whose complexity is far above the chain. Then the result, AlphaGo wins with a score of 4:1. On the other hand, Goodfellow's father is hanging out with his friend and gets an idea. What if the two neural networks are asked to compete? One makes a fake image that looks real,

One more, detecting which one is fake. He went home, immediately went to the modding, and the result was GUN. Generative Adversarial Network. Five years later, this person does not exist.com was born. Try opening it. There is a realistic face photo, but the person is not there. And never was. It's a bit creepy actually. In 2015, a group of people in San Francisco made OpenAI. The founder, among others, Sam Au.

Greg Brockman and Ilya Sudskever, yes, the same person who made AlexNet. One of the biggest donors is Elon Musk, with the mission of ensuring that AGI is developed for the interests of all human beings with a non-profit status. Remember this because we will come back here later.

In 2017, Google published the paper Attention is All You Need, which introduced the architecture of Transformers. Before this, the language model is difficult to understand the long context. The simple example is the word "bank" in the sentence "I sit at the riverbank and I go to the bank to take money" has a different meaning. AI ​​has to understand the relationship between words, not just reading the order. Transformers change the way AI sees that relationship.

And almost all of the current AI language models, CGPT, Gemini, Cloud, are built on this architecture. OpenAI adopted this, and in 2019 they released CGPT 2. But they themselves say that this model is too dangerous to be fully released. Is it marketing or true? I don't know. But what's clear is that this is just a warm-up.

In the midst of the 2020 pandemic, OPPO NEi released GPT-3 with 175 billion parameters. It can be coded, write essays, and even make poetry. But GPT-3 is still API only, so its excitement is limited to the tech enthusiasts. The general public doesn't care, they don't even know. Until these two things come.

and start to disturb a certain community. Artists, DALL-E in 2021 and Mid Journey in 2022, just type and the picture will be out. But the problem is not the quality or even AI can make pictures instantly. The problem is the training data using art works from the internet without permission and even

even without paying. One of the most frequently mentioned is Greg Rutkowski, an illustrator from Poland who became a popular prompt in Midjourney. Until the number of AI pictures with his style is more than his own work, even though he never gave permission, never contacted, and of course never paid. Until,

Art Station was flooded with the image of "No AI Art", but at that time it still felt like an ordinary internet drama. Until November 30, 2022, OpenAI released CGPT. What makes CGPT different from GPT-3 is not just the size of the model, there is an additional technique called RLHF . Humans give feedback, for example the answer A is

better than B, then the AI will learn until the answer feels natural. GPT-3 is smart but clumsy, while CGPT is smart and feels like talking to people. Oh yeah, GPT-3 is the AI model, while CGPT is a chatbot application built on the GPT model and optimized so that it can talk to humans.

In just 5 days, you get 1 million users. In 2 months, you get 100 million users. While Instagram needs 2.5 years, TikTok needs 9 months. CGPT is one of the fastest product growth in the history of technology. And this spread is organic. People try it themselves, surprise themselves, and then tell others too. In Google, the condition is "Code Red". Their main business is search engine. And CGPT gives a new way to search for information. Google soon released BART.

But the first demo answered the question wrong and the stock fell to $100 billion in just one day. Microsoft All In, add an investment of $10 billion and integrate CGPT into Bing. Then there's Entrovik, made by a former OpenAI person who is concerned about safety and released Cloud. Then there's Meta, which has been released as an open source for a long time.

In 2023, the OPA America Writers Guild started working. One of the main requirements is the protection of AI. Writers are afraid that the studio will use AI to draft stories and pay people cheaper just to revise. A few months later, SAKAFTRA also joined MOGO. The actors were worried that their faces and voices would be scanned once and used forever. Now suddenly every week there is a new model and also a new drama.

Friday, November 17, 2023, the OpenAI Board of Directors decided that Sam Altman was fired. The reason is that Altman was considered not consistently candid. They feel he is not always honest. And in a few hours, almost 700 OpenAI employees signed a threat letter that if Altman doesn't come back, they will all resign. On Monday morning, three days after being fired,

Altman returned to being the CEO. While the board that pressed him, they were the ones who came out. The most awkward, Ilya Shutskever, the co-founder of OpenAI, joined the vote to press Altman on Friday. Then on Monday, he signed a letter asking Altman to return to being the CEO. Until now, the real reason has never been explained transparently. Some say it's about safety, power, or OpenAI's movement is too fast. There is no satisfactory answer.

Then there is Elon Musk, the co-founder of OpenAI and the biggest donor at the beginning, who agreed with his non-profit mission. But in 2018, he left. The official reason is because of the conflict of interest with Tesla. In the other version, he wanted to control it bigger, but the board didn't agree. Then as a response in 2023, he launched XAI and ROG. And in 2024, he launched OpenAI with the accusation of violating the non-profit mission.

OpenAI replied by publishing long emails from Elon Musk that showed that he had suggested that he was the one who held full control. Angry because he was too profit-oriented, but he had to be the main leader. Well, how is it?

In January 2025, a Chinese AI company, Deepseek, released a new model that is equivalent to the best open AI model and Google, but with a much cheaper training cost. Partly because they don't have access to the latest Nvidia chip due to the US embargo, so they have to be more efficient.

As a result, Nvidia's stock dropped almost $600 billion in a day. Need to note, this is not necessarily because of deep seek, but because the sentiment of investors and other aspects also plays a role. But that means if the advanced CGPT AI can be made cheaper, the barrier to entering the AI race is drastically reduced.

Meanwhile, there is an impact of AI that is more directly felt by ordinary people, namely deepfake. In early 2024, a fake photo of Taylor Swift spread to millions of people before it was deleted. It's not Taylor Swift who made the point. The point is that this can happen to anyone. The tools are there, free, easy to access, and the law in almost all countries is not ready for this.

In the programmer's group, there are questions that make it crowded. Is programming dead? There are Git Hapopilot, GCPT, Coursor, suddenly there are tools that can write code and make applications from text descriptions only. The short answer is not dead yet. What changes is the type of work that is needed. If it's just a repetitive boilerplate code, AI can overcome that. If the architecture of complex systems, tricky debugging, business context,

it still needs humans. What may be lost is the barrier to entry that was so high back then. And Jorvi Hinton, a person who has not given up for decades, who is most responsible for the Deep Learning revolution, left Google in 2023 so that he could be more free to talk about the danger of AI. He said he regretted joining this development, not because his technology was wrong, but because he was not sure humans were ready to deal with its consequences.

The person who is the most stubborn to defend his beliefs during the AI Winter is now one of the most stubborn to warn the danger of what he does. Oh yes, one more thing, everything I tell in this chapter, open AI buddhism, drama Elon, deep seek, deep fake, Hollywood strike, all happened in less than 2 years and it hasn't touched half of it.

Every story is worth making a video of. So I'll close it here first. From Alan Turing to CGPT, the history of AI is a story about humans who keep trying to make machines smarter. But the patterns are always the same. The hype rises, expectations grow, then reality is hit.

Now AI can be used by millions of people every day, but is it different this time or are we just repeating the same cycle? Because the question is from 1950, can the machine think? The answer is still unclear of course. Okay, that's it as usual, cemiu! And anyway, bye!