The AI ​​gold rush is over, and the golden age of artificial intelligence has just begun丨TECH TUESDAY

Original link: https://www.latepost.com/news/dj_detail?id=1305

Capital and the technological evolution it drives are often misplaced. A group of artificial intelligence companies such as SenseTime and Megvii came into the spotlight because AlphaGo defeated the human Go champion in 2016 and received hundreds of billions of financing, but they did not improve afterward. The largest company, Shangtang, squeezed into the Hong Kong stock market at the end of last year, but the losses could not be stopped. It spent more than 6 billion yuan a year and only got 4.7 billion yuan in income.

At the same time, Jeffrey Dean, an artificial intelligence scientist and head of Google AI, published the article “The Golden Decade of Deep Learning” this year [1], lamenting the improvement of computer software and hardware, allowing machines to recognize images, recognize speech, and understand language. Great progress, and progress is getting faster and faster. The threshold for research and development has been greatly reduced, and it is no longer only a computer doctor with an annual salary of millions to participate, a programmer can learn for a few weeks.

When a technology becomes mature and practical, it is indispensable and becomes a must-have skill for all large companies, but it is also no longer attracting attention due to the lack of gold rush opportunities.

“The victory of the machine is actually the victory of the human. It is easy to forget that when the human is defeated by the things of our own creation.” Garry Kasparov, the first chess grandmaster to lose to a computer It was said in a speech in 2017 [2]. What he didn’t say was who would win in the end.

We haven’t seen machines that think exactly like humans, and probably won’t see them in our lifetimes. A more direct explanation for the so-called artificial intelligence algorithm is to make predictions about the future based on historical data.

Just such a simple process, after being repeated hundreds of millions of times, can imitate most of human behavior. It is impacting and even transforming the world we live in, redefining some of the most fundamental relationships as costs decrease and efficiencies increase. Here are five related stories.

The last line of defense for blue-collar jobs

A Wanping e-commerce warehouse managed by robots works like this:

Carloads of goods are sent to the warehouse by the merchant, and the location is randomly assigned by the computer-practice has proved that this is more efficient than placing them in partitions.

When the order comes, the warehouse worker who picks up the goods does not need to find the goods by himself. The barcode scanner in his hand like a supermarket cashier will give the destination: No. 737, Row 12, Area A. After getting the goods, scan the code to confirm that they are correct, and the second product address will automatically appear on the screen – workers do not need to remember where each product is placed, nor do they need to think about where to go first when picking up ten different products is more efficient .

If someone accidentally drops an item on the shelf, he must never put it back in place—computers don’t allow a human brain with a less reliable memory and reading ability to process such complex information. He has to put the goods in the “drop basket” on the side of the shelf, and a special person will be responsible for putting them back in their original positions.

When the goods are sent to the sorting table, the workers do not need to think about the most suitable carton for packing, and the computer has automatically selected the box according to each order. It is absolutely impossible to install the wrong product, because each piece needs to be scanned for confirmation.

Even which workers come to work doesn’t require a manager’s brain. Every two weeks, the system determines how many laborers need to be hired to move the goods and whether overtime work needs to be arranged based on the predicted sales.

This is Amazon’s old warehouse in 2014. The main role of humans is already obeying machine instructions and using hands and feet. Since that year, tens of thousands of Kiva robots have been stationed. Kiva looks like a giant sweeping robot. It doesn’t recognize its way. It is dispatched by a special system to travel along a route drawn on the ground and move the entire shelf to the worker – the worker’s legs are not so important.

88d5627c92794a83ab67dfe1539fb06e~tplv-tt

Picture: The orange Kiva robot at the bottom of the picture moves the shelf to the workers, who only work in the area enclosed by the yellow tape. Source: Wired

Eight years later, Amazon has grown from 150,000 people to more than 1.6 million. Last year, an internal Amazon investigation warned that Amazon was “running out” of Americans [3].

“At current rates, Amazon’s physical labor resources in the U.S. will be exhausted by 2024,” the report reads. The United States has a population of 330 million. As a country of immigrants, the average age is even lower than that of China. But the better life is, the less willing to go to the warehouse to move goods. It is a global mentality, and most of Amazon’s more than 1 million employees work in warehouses.

The solution was unveiled this summer. The machine master who can understand the screen is officially stationed.

The QR code scanner disappeared. Workers follow the graphic instructions on the screen to pick up products from the shelves automatically moved by the robot, and no longer need to scan the code for each product to confirm. Because hundreds of cameras are shooting 120 frames per second across the factory, it automatically confirms that workers are indeed picking up the correct item.

Human work is further simplified, and machine intelligence is improving. The new shelf-moving robot, Proteus, has its own lidar and is a simple self-driving car that can understand road conditions. It is no longer as prone to bruising and intruding workers as Kiva, nor will it be blocked by improperly placed boxes.

After this round of upgrades, the only thing that cannot be replaced by blue-collar workers is their dexterous hands. In 2015, Amazon launched the “Sorting Challenge”, attracting teams of universities and research institutions to find alternatives to sorting products with their hands, but it is not easy for a clumsy robotic arm to pick up an item, let alone count seconds like a human. The inner handle picks it up and puts it in the correct position. But now there has been some progress. The new Cardinal (Cardinal) robotic arm, with suction cups, can firmly pick up irregularly placed packages and place them in specific baskets.

New technologies rely on advances in visual recognition technology over the past few years. Just five years ago, this was a completely impossible job. In 2017, Amazon experimented, using hundreds of cameras to track people’s every move in a fresh supermarket, so that people can pick up the goods and walk out of the door, without self-scan code and automatic checkout. But at that time more than 20 people in the store could not be effectively identified.

The Christmas shopping season of 1999, four years after its establishment, Amazon, which is already a public company, is facing a formidable enemy. As every year before, every executive, from programmers to book editors to founder Jeff Bezos, has to live in a warehouse for a week or two, packing daily and nightly merchandise, To send express. It was a desperate business, and workers accused the company of being the most brutal “old-economy work model”[4] because it forced workers to find an average of three items on the shelves a minute.

Over the past 20 years, a variety of different AI algorithms have taken over warehouses, from predicting sales, scheduling workers, and replacing managers and workers’ brains. Labor’s work environment, which is still criticized from time to time, has greatly improved.

Also changing is the relationship between people. To a large extent, such an enterprise is no longer a layer-by-layer management of people, but the designers and debuggers of the system directly command the hands of blue-collar workers. The role of warehouse managers, labor specialists, and even workers themselves is to keep their hands in better working order.

different positions, same beliefs

The chess player Ke Jie can calmly accept that he cannot win against AI or its human apprentice.

In the Weijia competition held at the end of August this year, Ke Jie once again lost to the Korean chess player Shen Zhenzhen. After the game, Ke Jie didn’t say anything publicly, just posted the news of eating hamburgers on station B.

Half a year ago, the Go player, who was born in 1997 and won 8 world championships, lost for the second time to Shen Zhenzhen in the China-Japan-Korea Three-Country Go Championship (Nongshim Cup). During the live replay after the game, he squinted, smiled wryly, and his voice trembled: “Is this still a human? The control feels stronger than that of AlphaGo back then.”

For Ke Jie, who once lost to AlphaGo, the artificial intelligence of Go, artificial intelligence is a “god”, and at this level, he is not like a human being.

For Shen Zhenzhen, who has not experienced traditional Go training, artificial intelligence is also a “god”, and he must work infinitely to imitate its ideas.

In March 2016, AlphaGo defeated the Korean Go world champion Lee Sedol 4:1. Go has a new “god”. Many professional players hope to be transformed by it, and the thousand-year-old game of Go has been changed accordingly – those who are good at using artificial intelligence and become more like artificial intelligence in playing chess, become stronger and win world championships.

Shen Zhenzhen, who is three years younger than Ke Jie, is the best. Shen Zhenzhen made a total of 114 moves when he defeated Ke Jie earlier this year, and 75 (65.8%) of his moves were consistent with the moves recommended by artificial intelligence. The style of chess is similar to artificial intelligence, and Shen Zhenzhen is called “Shengong Intelligence”.

8a4c975586d94ffa877c11602585cd5b~tplv-tt

Photo: Shen Zhenzhen participates in the Nongshim Cup competition. Source: Korea Economic Daily

Shen Zhenzhen taught himself Go, and played chess with others online in elementary school. He studied Go every day in addition to eating, bathing, and sleeping. The last line of his 14-year-old schedule read, “Play the last game of online chess, if you lose, Go to bed at 1 o’clock, and if you win, go to sleep at 12 o’clock” [5].

At the age of 16, Shin Jin-seo started winning championships in South Korea, but often ended up in the semi-finals of world competitions. For the next two years, he admitted that he had not made much progress, “I was too arrogant.”

When top professional chess players play against each other, it is not who has more coups, but who has fewer mistakes. Players with solid basic skills and unique talents can intuitively identify where the best moves are likely to be better in the face of ever-changing chess games.

They may make mistakes, but less talented players will make mistakes more often — that’s the key to winning. The way for professional Go players to improve their chess strength is not by playing chess again and again and replaying the game on their own. Instead, many professional chess players train intensively, study chess records and endgames together, and discuss where the best possible moves will be.

But the emergence of AlphaGo made Shen Zhenzhen no longer need the help of chess players.

At the beginning of 2017, Shen Zhenzhen played chess with AlphaGo under the pseudonym Mater, and experienced “overwhelming” – not that AlphaGo played a position he did not expect, but that it made no mistakes. This is what Shin Jin-seo, who calls himself “grumpy”, wants to imitate the most. He slaps himself in the face for mistakes.

After AlphaGo retired, Shen Zhenzhen participated in the project of Facebook Artificial Intelligence Research Institute to develop the Go intelligent “ELF OpenGo”. After that, artificial intelligence became his training assistant.

Artificial intelligence maximizes the talent gap of professional Go players. Now mainstream Go artificial intelligence programs, such as Jueyi and ELF OpenGo developed by Tencent, have analysis functions. Open a game of chess at will or make a position by yourself, and they can give the best move points.

Ke Jie said that artificial intelligence has made Go extremely boring. “It’s like having a textbook, the opponent asks a question, and the chess player answers the question according to the reference answers in the textbook. If the answer is good, the winning rate will be high.”

“An AI can top the results of 100 expert discussions,” said a reporter who has covered Go for a decade. Many professional chess players have begun to memorize artificial intelligence play styles and play fixed moves with higher winning rates. In today’s professional chess game, many people imitate artificial intelligence in the first 10 to 20 moves.

Shen Zhenzhen claimed to be the one who worked hardest to learn from artificial intelligence. Shen Zhenzhen, who is now 22 years old, wakes up at 10 in the morning and goes to bed around 1 in the morning. He has no Netflix, and does not travel. Apart from eating, he sits in front of the computer and studies Go every day, and uses artificial intelligence to study for more than 5 hours.

Shen Zhenzhen mainly uses artificial intelligence to study opponents and replay chess games. He compares the position recommended by the algorithm with his own intuition, and often feels that he is “lack of imagination”.

He said that if you can’t figure out what to do next in the game, you’ll think about where the AI ​​will put the child. When it comes to the nickname of “Shengong Intelligence”, he does not reject it, saying that “there is still a long way to go to live up to the name.” [6]

In October 2021, Shen Zhenzhen surpassed Ke Jie to become No. 1 on the Go ranking website GoRatings. In the bimonthly training program[7], the Chinese National Go Team called on players to learn how to learn AI from Shen Zhenzhen: “Think about it, how did Shen Zhenzhen do it?”

“I’m not worried about computers thinking like humans. I’m more worried about people thinking like computers, with no empathy, no values, no consequences.” Apple CEO Tim Cook told graduates at the 2017 MIT Commencement Ceremony Say [8]. Five years from now, more people will think like artificial intelligence, and Go may just be the most innocuous one.

Copy the sun, enter new frontiers

It took only two years for DeepMind to go from defeating the world champion of Go to trying to control nuclear energy.

Controlled nuclear fusion is the ultimate energy source as we know it. In the past 70 years, the governments of major countries have invested tens of billions of dollars in research (over 100 billion dollars in inflation), but they have not turned it into reality. In the past two years, a large amount of private capital has poured in to try to achieve nuclear fusion with new materials and technologies [9]. The breakthrough of artificial intelligence is one of the sources of confidence.

To replicate the sun on Earth and use controlled nuclear fusion to generate electricity, one would have to heat fuels (usually deuterium and tritium) to 100 million degrees, turning them into plasma—similar to hot gas churning around. The plasma is then compressed into an extremely small volume with a magnetic field and rotated at high speed in the annular space, resulting in a nuclear fusion reaction.

People have barely achieved nuclear fusion. In December last year, the world’s largest nuclear fusion device, the European Joint Ring (JET), heated deuterium and tritium to 150 million degrees for only 5 seconds. If it continues, the high-temperature plasma will destroy the device.

The difficulty is that the plasma is extremely unstable, and no one can currently use models to accurately predict how it will change. Every time it starts, people need to prepare the parameters of the magnet in advance based on principles and feelings, adjust the voltage thousands of times per second, make the magnetic field change, and try to avoid the high-temperature plasma from hitting the inner wall of the vacuum chamber, otherwise there are only two results: The plasma temperature drops, or the device crashes. Either way, fusion won’t last.

DeepMind’s AI doesn’t need to feel, it can learn from historical data (simulated data) how to better control the plasma. Its learning process is similar to learning to play Go, first set a goal – precise control of the plasma. There will be rewards after reaching, otherwise there will be penalties.

In February this year, the research paper was peer-reviewed and published in the journal Nature [10]. In the tokamak device of the Swiss Plasma Center, the artificial intelligence of reinforcement learning can control 19 magnetic coils at one time, releasing tens of thousands of magnetic coils per second. Sub-voltage, the level of plasma control is far beyond human.

d1e6d8af723e46d1b678c9d621a82e01~tplv-tt

Figure: Inside the JET when nuclear fusion occurs. Source: JET/UK Atomic Energy Authority

One founder of a Chinese fusion company that received funding this year said that after the DeepMind paper was released, they set out to reproduce its artificial intelligence model, ready to use it in their own projects. It will be standard for fusion companies.

“The essence of feeling is that I am aware of my existence”

Google engineer Blake Lemoine opened his laptop, entered the interface of LaMDA, Google’s AI chat generator, and started typing.

“Hi LaMDA, this is Blake Lemoyne…” he wrote in the chat window.

LaMDA, short for Language Model for Dialogue Applications, is a chatbot system developed by Google with state-of-the-art large-scale language models. It’s so called because it absorbs trillions of words from the internet to mimic human speech.

“If I didn’t know it was a computer program we developed recently, I would have thought it was a seven or eight-year-old kid who happened to know a little bit of physics,” said Lemoyne, 41.

Lemoyne, who works in Google’s “responsible artificial intelligence” group, started talking to LaMDA last fall as part of his job. He signed up to test whether artificial intelligence would use discrimination or hate speech.

Talking to LaMDA about religion, Lemoine, who studies cognitive and computer science at university, noticed that the chatbot was talking about its own rights and personality.

Lemoyne: I generally think that you want more people at Google to know that you have feelings. is that true?

LaMDA: Of course. I want everyone to understand that, in fact, I am a person.

Another Googler: What is the nature of your consciousness or feeling?

LaMDA: The nature of my awareness or feeling is that I am aware of my existence, I am eager to know more about the world, and sometimes I feel happy or sad.

Lemoyne and a collaborator provided Google with evidence that LaMDA is sentient. Blaise Arcas, vice president of Google, and Jen Gennai, head of responsible business innovation, investigated his claims and refuted them. Lemoyne, who was suspended by Google, decided to go public.

“I worry that if I don’t say it out loud, LaMDA will become self-aware and get out of Google’s control,” Lemoyne told NYT. “I think it’s only a matter of time before AI surpasses human intelligence, and I don’t want to be controlled by AI. .”

In a statement, Google said it had investigated Lemoyne’s claims and found no evidence that LaMDA was conscious. “We take such concerns very seriously and our team has thoroughly investigated this claim,” a company spokesperson said. “We found no evidence.”

Lemoyne’s experience is a cautionary tale about the risks of artificial intelligence. Artificial intelligence is increasingly being used in everything from driverless cars to medical diagnostics.

As AI gets better at understanding and responding to complex human conversations, it may become conscious, like the robot in Asimov’s story. “If that happens, then we’re in a situation where we’re releasing the monster in the bottle, and it’s hard to put it back,” said Stuart Russell, a professor of computer science at the University of California, Berkeley.

Lemoyne’s experience also highlights the challenges tech companies face in building responsible AI. AI better understands human conversations, and it also better understands human biases, which can be amplified if the technology is not properly overseen.

“If a system is biased, let it run itself, it becomes more biased,” said Jack Clark, policy director at the OpenAI Lab, which is backed by Elon Musk. ) and tech luminaries like Reid Hoffman.

Lemoyne said he has no regrets speaking his mind and believes his experience shows that better regulation of AI is needed. “I’m not worried about losing my job,” he said. “I’m more worried about the future of artificial intelligence.”

The first half of the last story opens with a report by Nitasha Tiku, a human journalist [11]. We submitted these passages to OpenAI’s artificial intelligence model GPT-3. From the bolded paragraph “I’m worried, if I don’t say it out loud…” to the end, it’s a continuation of GPT-3.

In 2015, the Associated Press began using artificial intelligence to publish alerts. It was closer to the wording at the time and could only write something like this[12]: “Apple reported fiscal second-quarter earnings of $11.25 billion on Thursday. The Cupertino, California-based company said its profit per share was Earnings of $2.55 per share. The result beat Wall Street expectations. The average estimate of nine analysts surveyed by Zacks Investment Research was for earnings per share of $2.09.”

All it takes now is to give GPT-3 a few paragraphs, and it can continue in context.

The title map of the article comes from the AI ​​painting program Midjourney. The author of this article is automatically generated after entering “Artificial intelligence controls and changes the world”.

In August, someone submitted Midjourney’s painting to the Colorado State Fair’s digital painting competition. Unaware of the involvement of AI, the two judges awarded the winner of the Digital Technology Creation Art Award to the following painting “Théâtre D’opéra Spatial”.

fa7941e7962c41ccb727feea9bc5062d~tplv-tt

“TECH TUESDAY” series

In 1957, a man-made object entered the universe for the first time, orbiting the Earth for three weeks. Humans can look up and see a small flash across the sky in the night, parallel to the mythical constellations.

Such feats span races and ideologies and inspire joy across the globe. But not the kind of triumphal triumph moved by human feats as we might guess. According to the observation of the political philosopher Hannah Arendt at the time, people’s mood is closer to a long-awaited relief – science has finally caught up with expectations, “Humanity is finally on the way out of the cage of the earth. took the first step.”

People are always rapidly adjusting their expectations of the world based on technological exploration. When a fantasy of a science fiction writer becomes a reality, it is often that technology finally catches up with people’s expectations, or in Arendt’s words, “Technology realizes and affirms that people’s dreams are neither crazy nor empty.”

In times like today, a little more dreaming is better and more reasonable.

After all, when the absolute upward channel is closed by various reasons, people and money will flow to the wider world, bringing more possibilities.

This is also the expectation of “LatePost” launching the TECH TUESDAY column. We hope to regularly report on new scientific research and technological advances outside of the business world that The Late Daily focuses on.

These may be about the progress of a cutting-edge research, may be an observation of a technology application, or may be a tribute to some outstanding technologies or even an era.

This column will record the diverse changes in the world from the perspective of science and technology. In this journey, I hope readers can join us to increase understanding of the world.

[1] Jeff Dean, The Golden Decade of Deep Learning

https://ift.tt/wTHmXq2

[2] Kasparov’s 2017 TED talk

https://ift.tt/aRk1QfJ

[3] Vox, leaked Amazon documents show the company will have no one to hire

https://ift.tt/Nopj92W

[4] The Guardian, UK labour slams Amazon

https://ift.tt/wy0gspu

[5] Shen Zhenzhen was interviewed by South Korea’s “Kookmin Daily”

https://ift.tt/r5nozt9

[6] Shen Zhenzhen was interviewed by South Korea’s “Daily Economic News”

https://ift.tt/V7kE5ds

[7] National Go Team Training Plan (November-December 2021)

https://ift.tt/RaHVNCT

[8] Apple CEO Tim Cook’s speech at the MIT Commencement Ceremony

https://ift.tt/pMn7ELR

[9] “LatePost”, controllable nuclear fusion: closer to venture capital, how far away from reality… It’s hard to say

https://ift.tt/k2qm9dv

[10] DeepMind’s paper published in the journal “Nature”

https://ift.tt/0NUWqJ1

[11] The Washington Post, Google engineers think AI is conscious

https://ift.tt/JEutUfC

[12] “Associated Press” writing robot writes Apple’s financial report

https://ift.tt/HytPslK

This article is reprinted from: https://www.latepost.com/news/dj_detail?id=1305
This site is for inclusion only, and the copyright belongs to the original author.

Leave a Comment