In 2024, where will the AI predicted to make great progress go？

Time flies, it’s a new year in a blink of an eye, and those unfinished dreams of the old year will turn over a new chapter. At the end of last year, among the top ten scientific advances in 2024 predicted by the British journal Nature, the progress of artificial intelligence and ChatGPT artificial intelligence occupy the top two. The development of artificial intelligence has made numerous breakthroughs in the past year, which has also caused widespread controversy. However, it is undeniable that artificial intelligence has become a symbol of the future, and will continue to approach and influence people’s lives at present and in the future, and will have a broad and profound impact on the progress of human society. So, what achievements of artificial intelligence will bring beneficial development to science? Today, we invite Zhang Tiankan, a well-known popular science author, to talk about this topic.

(1) PANDA made in China is used for early screening of pancreatic cancer.

Artificial intelligence has gradually entered people’s lives, and in the field of disease diagnosis and treatment, artificial intelligence has exerted a strong influence.

Known as the "king of cancer", the 5-year survival rate of pancreatic cancer is extremely low. According to the latest data of American Cancer Society in 2023, the 5-year survival rate of pancreatic cancer patients is 11%-12%, and the latest data of China is 7.20%. If patients can find the disease as early as possible and perform surgery, the curative effect is relatively better. However, pancreatic cancer is often difficult to find out, and most of them are in the advanced stage when they are diagnosed. Therefore, using artificial intelligence technology to help early diagnosis of pancreatic cancer and other cancers is the direction that scientists have high hopes for.

Recently, Shanghai Institute of Pancreatic Diseases, Ali Dharma Hospital, the First Affiliated Hospital of Zhejiang University School of Medicine and other institutions jointly used "plain CT combined with artificial intelligence" (plain CT+AI) to conduct large-scale early screening of pancreatic cancer. In the real world, the sensitivity is 92.9% (the accuracy of judging the existence of pancreatic tumors) and the specificity is 99.9% (the accuracy of judging the absence of tumors). This result was published in the international first-class journal Nature-Medicine, which also published a special commentary saying that artificial intelligence and image-based cancer screening are about to usher in a golden age.

Our research team has developed and trained a pancreatic cancer early screening model panda (with AI) based on artificial intelligence, which uses AI to enlarge and identify the subtle features of lesions in plain CT images that are difficult for naked eyes to identify. The results show that PANDA can accurately detect and diagnose pancreatic ductal adenocarcinoma (the most common type of pancreatic cancer, accounting for more than 90%) and non-pancreatic ductal adenocarcinoma lesions on plain CT, and can be used for screening large-scale asymptomatic patients at any time.

Compared with enhanced CT, plain CT can reduce the radiation dose and eliminate the risk of adverse reactions caused by contrast agents. Clinically, it is difficult for even experienced radiologists to identify pancreatic ductal adenocarcinoma by plain CT, but the combination of PANDA and plain CT has certain advantages in distinguishing common subtypes of pancreatic lesions. The researchers collected the data of 3,208 real patients from several medical institutions to train PANDA, and then used it to verify 6,239 patients in 10 medical centers. The results show that this model has excellent sensitivity and specificity in identifying pancreatic ductal adenocarcinoma.

PANDA is expected to be widely used as a new large-scale screening tool for pancreatic cancer in hospitals or physical examination centers. At present, the model has been used more than 500,000 times in medical treatment, physical examination and other scenes, and the false positive rate is very low. In addition to the diagnosis and screening of pancreatic cancer, the research team is cooperating with many first-class medical institutions around the world to check more cancers, including esophageal cancer, lung cancer, breast cancer, liver cancer, gastric cancer and colorectal cancer.

(2) Generative AI such as 2)GPT will update the iteration.

Looking forward to the development of artificial intelligence this year, Nature magazine lists the emergence and functions of several representative AI modes. For example, GPT-5 will come out and may show more advanced functions than its predecessor GPT-4. At the same time, GPT-4 competitor Gemini (another generative AI tool) and other AI tools will also be launched.

GPT-5 and Gemini are both large language models and generative artificial intelligence, which can be used to create new content and ideas, including dialogues, stories, images, videos and music. Therefore, they can not only generate many languages and translated works, but also generate artistic products for scientific research and teaching.

Generative artificial intelligence model has great potential in human communication and service. Different languages lead to communication barriers for people living in different regions, and simultaneous translation of AI software may help solve this problem. Meta Company in the United States developed Seamless, an open source voice translation model for seamless communication, while Google Company developed Translation 3, an unsupervised voice translation AI system. Among them, Seamless of Meta-Universe Company is a "unified model", which integrates all the functions of the other three deep learning models and can conduct more natural and authentic cross-language communication in real time. While translating words, Translation 3 of Google can also deal with subtle differences in non-text pronunciation such as pause, speech speed and speaker identity.

Even more amazing, artificial intelligence software is expected to be used to "smell" smells and analyze the structure of odor molecules. At present, people can use wavelength to study vision and frequency to study hearing, and can measure and evaluate it through instruments, but they can’t measure or accurately predict the smell of substance molecules according to molecular structure.

Researchers from Google Inc., Monel Chemical Sensory Center and University of Reading in the UK have developed an AI tool that can predict its odor characteristics only based on its molecular structure. It can identify molecules with different appearances but the same smell, and can also identify molecules with very similar appearances but completely different smells.

This AI system is called odor map. After the big data training of odor molecules, it is not only suitable for identifying known odor substances, but also for identifying odor substances with very similar structures, and can also describe a large number of unrelated molecules with different molecular characteristics. From the application scenario, odor spectrum can be used not only in food and agriculture to predict and discover new compounds, spices and foods, but also in the research and development of chemical products and cosmetics, biomedicine and other fields, and its "smell" sensitivity is more reliable than that of dog’s nose.

(3) There is a new breakthrough in dealing with antibiotic-resistant bacteria.

Nature magazine predicts that this year, Deep Thinking Artificial Intelligence will release a new version of the artificial intelligence tool AlphaFold. Previously, researchers have predicted the 3D shape of protein with high accuracy by using the old version of alpha folding, and the new alpha folding will simulate the interaction between protein, nucleic acid and other molecules with atomic accuracy, which will open up a new way for drug research and development.

Artificial intelligence can exert its talents in chemistry, biomedicine and pharmacy. The main reason is that it can quickly discover new substances and invent new materials.

Everyone will get sick. In the past, human beings used too many antibiotics for this purpose, and the antibiotics used in animal breeding have greatly increased the drug resistance of many bacteria. As early as a few years ago, the World Health Organization warned that if human beings use antibiotics without restraint, infectious diseases will fall into the dilemma of no drugs (antibiotics) available in the future.

Perhaps, the intervention of artificial intelligence can solve this problem. According to the research results published by a research team of Massachusetts Institute of Technology, artificial intelligence has identified a new type of antibiotics from tens of millions of compounds, which can kill the superbacteria commonly seen in clinic-MRSA.

The research team generated training data by testing the antibiotic activity of about 39,000 compounds against methicillin-resistant Staphylococcus aureus, and then input these data and chemical structure information of the compounds into a deep learning model. At the same time, the research team also trained three additional deep learning models to predict whether these compounds are toxic to three different types of human cells. Finally, researchers found that some compounds can not only kill bacteria, but also have the least adverse effects on human body.

The research team screened out about 12 million compounds, and the artificial intelligence model identified five different types of compounds according to the chemical substructure in the molecule, and predicted that they could fight against methicillin-resistant Staphylococcus aureus. Finally, the artificial intelligence model selected two compounds, which were considered as the best antibiotic candidates.

After that, the research team used two mouse models to verify the antibiotic candidates, one was the skin infection mouse model of methicillin-resistant Staphylococcus aureus, and the other was the systemic infection mouse model of methicillin-resistant Staphylococcus aureus. The results showed that these two antibiotic candidates significantly reduced the number of methicillin-resistant Staphylococcus aureus. That is to say, the compounds selected by the artificial intelligence model are effective for local and systemic infections of methicillin-resistant Staphylococcus aureus, and are suitable for further research and development and treatment of serious septicemia-related drug-resistant bacterial infections.

Methicillin-resistant Staphylococcus aureus is a kind of Gram-positive bacteria with thick cell wall. Further research shows that these two antibiotic candidates can destroy the cell wall of the bacteria, but they will not cause substantial damage to the human cell membrane, which is safe for human cells. We have reason to believe that artificial intelligence can screen out more antibiotics for treating diseases in the future.

(4) It is necessary to study whether AI will produce consciousness.

Interestingly, Nature also included the debate and discovery about consciousness in this year’s scientific progress, arguing that this is not only a study of human consciousness based on brain and nerve, but also a breakthrough in human psychology and philosophy research. Of course, the biggest controversy about this research is still whether artificial intelligence will produce consciousness and how humans treat artificial intelligence that may have consciousness.

Previously, Templeton World Charity Foundation funded 30 million dollars to study and verify two main theories of consciousness generation, one is integrated information theory, and the other is global neuron workspace theory. Researchers from different laboratories used functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG) and electroencephalogram (EEG) to study the neural correlation between the content and duration of visual experience of 256 human participants. The results showed that there were information reflecting the content of consciousness in visual cortex, ventral temporal cortex and subfrontal cortex. The sustained response in occipital cortex and lateral temporal cortex reflects the stimulation duration; There is content-specific synchronization between frontal lobe and early visual area.

In short, these results confirm some predictions of integrated information theory and global neuron workspace theory, but neither theory can explain how consciousness is produced. Nature magazine predicts that before the end of this year, scientists will have a new understanding of the neural basis of consciousness and publish new experimental results.

Associated with this, does artificial intelligence really produce consciousness? Some scientists worry that if artificial intelligence evolves consciousness, it may change from human control of artificial intelligence to artificial intelligence control of human beings, which will be the end of mankind. Although more people are dismissive of the hypothesis that "artificial intelligence produces consciousness", members of the Mathematical Consciousness Science Association (AMCS) call on the United Nations to provide more funds to support the research on consciousness and artificial intelligence. According to the association, there is an urgent need for scientific investigation on the boundary between conscious and unconscious systems, which involves ethical, legal and security issues. These problems make it very important to understand the consciousness of artificial intelligence. For example, if artificial intelligence develops consciousness, should people be allowed to shut it down effectively and simply after use?

At present, some researchers in the world predict that the consciousness of artificial intelligence will be realized within 5-20 years. But the fact is that this conclusion lacks research support. In 2023, no funding was used to "study artificial intelligence to generate consciousness". Researchers of the Mathematical Consciousness Science Association believe that only by knowing what can make artificial intelligence conscious can we evaluate the impact of conscious artificial intelligence systems on society, including the dangers they may cause and how to deal with them.

In any case, with the rapid development of science and technology today, it is of great significance for the future of human social civilization to study whether artificial intelligence will produce consciousness and draw exact conclusions. At that time, it is not too late to evaluate whether artificial intelligence is a favor or a burden for human social civilization.