Original link: https://www.indigox.me/indigo-talk-ep07/
In the seventh issue of INDIGO TALK, Dr. Chen Xin, the founder of Gusu Laboratory, was invited. He is a material scientist. After graduating from Stanford University, he has been engaged in new material science research in the United States for more than ten years. With his professional background, we will discuss how this revolution of AI will change scientific research, how large models will be applied in the field of scientific research, which research directions will be accelerated, and what will the new scientific paradigm look like? As always, we will conclude by discussing the impact of the advent of AGI on society.
To listen via Spotify or Apple Podcasts, click here !
To listen to the Little Universe APP, please click here !
Note: The network environment and equipment used to record Chen Xin this time are not very good, so the audio and video quality has been affected. I hope everyone understands!
Guest of this issue
Chen Xin (Materials Scientist Suzhou Laboratory Researcher)
Indigo (digital mirror blogger)
timeline
00:00:25 – Guest introductions at the opening of the program
00:02:39 – What is materials science
00:07:16 – What is the typical new material on the iPhone?
00:09:39 – How does the scientific community think about large language models in AI?
00:23:30 – Comparing materials science and protein bioscience research
00:27:52 – Big models for scientific research
00:32:40 – If the scientific research paradigm shifts to connectionist thinking
00:45:50 – A quick recap of INDIGO
00:46:24 – Some ideas about high-energy physics research
00:49:52 – Which research directions will AI accelerate first?
00:56:00 – How will AI affect production and scientific research in the physical world?
01:02:52 – Discussion on the arrival of AGI
01:09:00 – INDIGO’s final wrap-up
content outline
Chen Xin mentioned in the conversation that many scientific problems are to solve the mapping from a to b in the function. For example, material science is to map the structure of the material to the performance of the material. If I know the structure, I can immediately solve the performance. Then this problem is solved. DeepMind’s AlphaFold used AI’s neural network to solve this mapping at the end of 2021. The spatial structure of the protein can be solved by inputting the DNA sequence. The accuracy rate exceeds 90%, which exceeds the results of scientists’ various experiments and logical deduction. This is an epoch-making change in the history of science, bringing a new paradigm to scientific research.
Such results show that AI has been able to perceive the laws that humans cannot discover through rational cognition, and use this to guide its actions. Humans are no longer the only intelligent species capable of discovering and perceiving patterns in our known world.
The following content is generated by GPT-4 based on the dialogue content of each discussion summary (Prompt: summarize the main points of the following dialogue and answer in list form)
Materials Science and Its Impact on Human Civilization
- Materials include all kinds of atoms and molecules used in production, such as metals, plastics, etc.
- The definition of stages of human development is often related to the utilization of new materials, such as the Stone Age, Bronze Age, Iron Age, etc.
- Energy represents the level of civilization, and the effective use of energy depends largely on the transformation of materials.
- New materials lead to new technologies and production methods, such as silicon semiconductors, optical fibers, and lithium batteries.
- Innovations in the use of materials lead to convenience products such as mobile phones and electronic devices that improve people’s quality of life.
- The progress of civilization is accompanied by changes in the main materials, such as stone tools, bronze, iron, steel, silicon, glass and lithium.
- The emergence of new materials and technologies has triggered paradigm shifts and technological leaps.
Research Paradigms and the Development of AI
- Currently is the era of AI acceleration, focusing on large models such as large language model and large material model.
- The development of large models has benefited from advances in algorithms, technology, and hardware.
- There have been many fluctuations in the development and application of AI in the field of science, from logical reasoning in the 1960s to probabilistic in the 1980s, to the current era of machine learning and large models.
- The era of large models has made the scientific community more rigorous in its reflection on the AI revolution and its future development.
- The history of AI for Science can be traced back to the 1960s, but gradually entered a quiet period in the 1990s.
- With the emergence of NVIDIA GPU and the development of deep learning technology, the application of AI technology in scientific research has gradually recovered.
- Through large amounts of data and neural network training, AI can establish the mapping relationship of complex problems (such as structure to performance).
- Since 2015, data-driven research paradigms have emerged in AI for Science.
Contrasting research in materials science and protein bioscience
- AlphaFold is a landmark work that can predict the three-dimensional structure of proteins from DNA sequences.
- The protein structure is relatively simple from a mathematical point of view, it is a linear polymer chain, each chain has 20 kinds of amino acids.
- The significance of the development of AlphaFold is very important, because new molecules can be designed to block viruses by predicting protein structures.
- The protein data set comes from Protein DataBank , an international non-profit organization where scientists from all over the world can share protein structure data.
- The informatization of biological materials is relatively simple, but the informatization of materials science is much more complicated, because the molecular structure is more random and free.
- Current materials science research needs to narrow the scope of research to a family of materials based on a benchmark material.
Big model about scientific research
- Current AI such as ChatGPT can learn common sense, sometimes better than some human common sense.
- Many people lack scientific common sense, even educated people may lack scientific common sense because they are no longer exposed to science.
- Explore whether it is possible to teach AI scientific common sense through large-scale models to provide support for scientific research.
- With the help of the ideas and methods of large models, it is possible to achieve master’s or doctoral level intelligence at the general knowledge level within 2-3 years.
- Facebook released a scientific innovation model called Galactica , which was judged unusable due to the hallucinations caused by the accident.
- Discusses AI21 ‘s plans to launch training models for scientific papers in 2024.
If the scientific research paradigm shifts to connectionist thinking
- GPT is based on natural language, but natural language has limitations and may not be the best way to describe scientific problems.
- There have been attempts to use formal languages or programming languages to describe scientific problems, but there has been no apparent success, perhaps because the amount of data is not large enough, or the model is not large enough.
- The material world can be divided into different levels, such as molecular level, atomic level, subatomic level, etc., and it may be necessary to establish connections between these levels.
- Subject classification may be an artificial division, which does not conform to natural laws, and the connection between knowledge in different fields should be sought.
- With the development of artificial intelligence, it may bring profound changes to academia, education, work and social organization.
Some insights into high-energy physics research
- The study of matter may be stratified according to the granularity level of the macroscopic world, rather than the existing physical and chemical levels.
- Use the connectionist model to connect matter and transform data.
- The world inside the nucleus is the real quantum world.
- The Standard Model physics is complex and the particles are named arbitrarily, but it might be a model.
- Without enough experimental data to observe details, training AI in microphysics may become a research tool.
- Regarding the divisibility of a particle, it may take a huge amount of energy to break down.
- Particle accelerators are getting bigger and bigger, but the scale required for measurement is also getting bigger. When the Planck scale is reached, the accelerator may need to be as big as the Milky Way.
- In recent decades, human beings may not be able to reach a higher level of material research, but AI may bring unprecedented discoveries.
Which research directions will AI accelerate first?
- The application of AI to weather forecasting has achieved remarkable results, improving accuracy.
- AI has great application potential in the fields of medicine and new materials.
a. Accelerate the development of new drugs and reduce social costs.
b. There may be at least 10 AI-developed drugs in the next 10 years. - Applications of AI in scientific fields such as astronomy are also evolving.
- Energy and new materials in the field of medicine will be the fastest growing direction of AI.
- The development of new materials can help solve energy utilization and storage issues.
- The application of new materials also includes electronic devices (such as AR/VR and micro-light-emitting arrays), which may promote the realization of metaverse concepts.
How will AI affect production and scientific research in the physical world?
- AI big models are evolving rapidly in the digital world.
- Human beings tend to underestimate the development speed of the digital world and overestimate the development speed of the atomic world.
- The AI leap is beyond many people’s imagination, and intelligence appears rapidly.
- AI scientists believe that AI can have a greater impact on the physical world. Huang Jiaozhu predicts that the next wave will be the effect of AI on the physical world, worth 100 trillion US dollars.
- The ARK Fund predicts that the value of AI to the world will be about 75 trillion US dollars in 2030.
- AI’s impact on the physical world is in its infancy, but there are already some signs, such as Nvidia’s computational lithography technology.
- AI may speed up human improvements to the physical world, but not faster than the digital world.
- Gradually pay attention to the trend of AI’s transformation of the physical world.
- AI can provide better experimental plans and assistants for physical experiments.
- Digitally connect large scientific installations with AI, and use AI to conduct scientific research experiments.
Discussion on the arrival of AGI
- The arrival of general intelligence is inevitable, it is only a matter of time.
- The arrival of AGI may have both positive and negative impacts on humans.
- Humans may be the bootstrap of AGI, guiding its development.
- Human use of technology and facing new things will change values and moral standards.
- Values that cannot predict the future need to be adjusted continuously with technological progress.
- Welcome to the arrival of AGI, because it will have the potential to help humanity enter the next stage.
- Regardless of whether future humans are based on silicon-based or carbon-based life forms, or a combination of them (such as cyborg), humans will need to adapt to technological progress.
related reference
Blue LED – Inventors of the blue LED Shuji Nakamura, Hiroshi Amano and Isamu Akasaki won the 2014 Nobel Prize in Physics
OLED – Organic Light Emitting Diode
Micro LED – micro light-emitting diode ( miniLED, OLED, MicroLED detailed explanation )
Polymer – Polymer is an important material in science and engineering, widely used in plastics, rubber, fibers, adhesives, coatings, glass, and biomolecules such as proteins, nucleic acids, and polysaccharides.
Corning ‘s Gorilla Glass (Gorilla Glass)
AlexNet – is a Convolutional Neural Network (CNN) architecture that gained a lot of attention in 2012 due to its remarkable performance in that year’s ImageNet Large Scale Visual Recognition Challenge (ILSVRC). This neural network model was designed by Alex Krizhevsky , Ilya Sutskever and Geoffrey Hinton . AlexNet is mainly used in computer vision tasks, especially in image classification and object recognition.
ResNet – The full name is Residual Network (residual network), which is a deep neural network (DNN) architecture. It was originally proposed by Kaiming He et al. in 2015 to solve the “gradient disappearance” and “degeneration problems” in deep learning. ResNet is widely used in fields such as computer vision and image recognition tasks because it can build very deep models and maintain high performance while training.
Back Propagation – is a training algorithm widely used in deep learning and neural networks. It adjusts the weights of each layer (that is, the strength of connecting different neurons) according to the error of the neural network in order to improve the performance of the model.
AlphaFold – is an AI system developed by DeepMind that applies deep learning techniques to predict protein structure, that is, predict the three-dimensional shape into which proteins fold. Prediction of protein structure has important implications for drug design, disease research, and the fields of biology, but the task has historically been viewed as a major challenge for the scientific community.
Protein Databank – wwPDB (Worldwide Protein Data Bank) is a global scientific organization that provides researchers with an important resource on biological macromolecules such as proteins and nucleic acids. Their main task is to collect, organize and provide access to the three-dimensional structural information of these biomacromolecules. This information is important for understanding fundamental processes in biology, drug design, and other fields.
Galactica – A research model trained by Meta Corporation to help researchers study.
AI21 Labs – is an Israeli start-up company focused on the research and development of artificial intelligence (AI) technologies, especially in the fields of natural language understanding (NLU) and natural language generation (NLG).
Wolfram｜Alpha – Founded by British scientist Stephen Wolfram, it focuses on computational intelligence and technology development.
Wittgenstein & Symbolism & Connectionism
Standard Model – It is a theory describing the three basic forces of strong force, weak force and electromagnetic force and the basic particles that make up all matter. It belongs to the category of quantum field theory and is compatible with quantum mechanics and special relativity.
PaLM 2 Med – a medical-specific version of PaLM 2
Stainless Steel for Starship – 300-series stainless-steel
Helion – Fusion company backed by Sam Altman
Retro Bio – Sam Altman’s investment in life science and technology companies that can add ten years to human life
This article is transferred from: https://www.indigox.me/indigo-talk-ep07/
This site is only for collection, and the copyright belongs to the original author.