Chinanews.com, December 21 (Reporter Meng Xiangjun) From AlphaGo's "chess skills" to the "car skills" of autonomous driving, from teasing ChatGPT's "talk more if you can chat", to discussing the two-dimensional images generated by AI paintings. No matter how amazing or funny, artificial intelligence seems to be increasingly becoming the focus of people's attention.
But at the same time, voices such as "not smart enough" and "dangerous" about artificial intelligence continue to emerge.
Will some of the popular artificial intelligence applications replace human jobs in the future?
Could they be more "smart"?
What else can be done to make artificial intelligence safe and comfortable to use?
Nie Zaiqing, Professor Guoqiang of Tsinghua University and chief researcher of the Intelligent Industry Research Institute (AIR), analyzed the development prospects of some popular artificial intelligence applications and the problems that need attention.
Data map: Nie Zaiqing.
Photo provided by the interviewee
ChatGPT "has made progress", how far is it from the Turing test?
Open AI previously released an AI chat robot ChatGPT, because of its outstanding intelligent dialogue ability, it can be said to be popular all over the world.
After trying out this robot, Nie Zaiqing had a lot of experience.
He said that the previous application of chatbots in the industry was to use task-based dialogues to extract question parameters through intent understanding and super-value filling, and then answer user questions (smart customer service) or help users complete a smart home. operation (AIoT voice assistant).
This kind of task-based chat robot is quite successful.
It has been able to help users complete many repetitive tasks through voice interaction.
Compared with the previous chat robot technology, the most impressive thing about ChatGPT is that every time it answers a question (it can be an open-domain question), it can answer the question in decent human words, and it will also have a certain understanding of the context. degree of understanding.
If you don't look at the specific knowledge points, it seems that the answer is very professional.
OpenAI's official website introduces ChatGPT as a "language model for optimizing dialogue".
Image source: Screenshot of OpenAI official website
Nie Zaiqing also pointed out that if the questioner does not know the answer to the question, he will feel that ChatGPT "answers all the questions well"; but for some uncommon questions, and the questioner knows the answer, ChatGPT is likely to "use a serious tone and make up wrong answer".
He said that this robot is essentially a language model based on statistics, and should not really understand the open domain problem.
It is actually based on probability in dealing with open domain problems, generating the next value based on the previous value.
To pass the Turing test, we must continue to work hard.
As for how to improve, Nie Zaiqing pointed out that if the correct knowledge that can answer questions can be put into what it says, "it will be amazing."
Behind the AI painting, how far is it from the artist?
Similar to chatbots, there are AI paintings that have recently become popular on short video platforms.
This is an application that pre-records a large number of two-dimensional pictures, uses big data to find similar elements in picture composition, shape, color, etc., and generates new pictures based on AI interpretation.
On the short video platform, one-click operation can convert the photos of netizens into two-dimensional images. The output results are very random, which often makes people laugh and cry, but netizens never tire of it.
"The function of turning any photo into a second dimension is quite interesting." When it comes to AI painting, Nie Zaiqing believes that this is an interesting product development direction for artificial intelligence.
Data map: At the World VR Industry and Metaverse Expo, a woman watched the artwork in "Metaverse Art".
Photo by Liu Zhankun
But at the same time, Nie Zaiqing pointed out that there is logic in painting, and there are still gaps in the transitions, lines, and details of AI paintings compared with real people.
The most important thing is that art comes from the precipitation of life, and the artistic value of a real artist comes from the appeal behind his works.
He further pointed out that although AI may happen to draw some seemingly infectious works, it is still very difficult to require its paintings to have a certain appeal in a specific situation.
Overall, AI painting may pose a threat to painters who focus on commercial marketing, engage in repetitive labor, and do not purely pursue artistic value, but it is too early to replace real painters and artists.
How far is autonomous driving from large-scale landing?
Another artificial intelligence application that people are increasingly hearing about or coming into contact with in life is autonomous driving.
Nowadays, there are more and more vehicles that can drive autonomously or semi-autonomously on the streets, but dangerous situations of related vehicles are often reported in the newspapers.
One side is the development of science and technology, and the other is risk prevention.
If L5 fully automatic driving is to be implemented on a large scale in the society, Nie Zaiqing believes that there are several major "notes".
First, he pointed out that the most important point is the construction and improvement of infrastructure.
If you want to drive autonomously, the road is different from what it is now. For example, all signals on the road in the future will be digitized and sent directly to the car to inform the vehicle that it is a red light, so there is no need to visually recognize the traffic light; In the event of an accident, the road system can automatically prompt all vehicles behind the accident to be aware.
He also pointed out that if autonomous driving and human drivers exist at the same time, a key problem that needs to be faced is that the game between human and car, and human and machine has not yet been clarified.
It is still relatively difficult for a machine to accurately judge every intention of a human driver.
Nie Zaiqing hopes that the infrastructure can be greatly improved. After the improvement, if all the cars on the road can be driven automatically, the accident rate should be greatly reduced.
Data map: At the 19th China-ASEAN Expo, the self-driving bus model attracted public attention.
Photo by Yu Jing
Second, once autonomous driving is implemented on a large scale, a remote control center and "cloud driving" are required to cover all road sections.
Nie Zaiqing said that although people will no longer "drive" cars by then, there must be someone in the cloud who can monitor the road conditions in various places, whether the car conditions are normal, and whether there is any problem.
If there is a problem, the personnel of the remote control center and the "cloud driving" agency must play a role and manage the situation well.
For example, he envisions that the traffic police can handle traffic accidents "on the cloud" in the future, and recover remotely when there is a systemic failure of the vehicle.
Third, to realize autonomous driving, a comprehensive safety testing system and a gray scale internal testing system for new functions are required.
In this regard, Nie Zaiqing pointed out that whether the safety level of unmanned vehicles meets the standards, whether the new functions will cause large-scale damage, etc., must be tested step by step, from the unmanned road section test to the online small-scale A process of continuous release in the road section.
The establishment of such a system must be perfect, it must be able to iterate quickly, and at the same time, it must ensure safety.
Otherwise, the autopilot system will face update problems, and errors may occur once an update is made, so it is also very important to answer this question well.
Fourth, Nie Zaiqing emphasized that a fault-tolerant mechanism for system failures must be implemented.
Assuming that there is a systematic failure of the autonomous driving road system at that time, tens of thousands of vehicles may be affected at the same time.
With a fault-tolerant mechanism, it is possible to roll back the system to the previous version, and it is necessary to ensure that this mechanism cannot make mistakes.
Fifth, before large-scale testing of autonomous driving, it is best to have a mechanism, that is, any autonomous driving decision can be explained.
For example, Nie Zaiqing said that this means that if people need it, they can check every decision of the self-driving vehicle and the specific reasons for the decision anytime, anywhere, and they can also take over the decision-making of the machine anytime, anywhere to form a set of human-machine dialogue and cooperation mechanisms. Only in this way can the safety of autonomous driving be greatly improved.
(use up)