“Top Ten Frontier Technology Trends Report 2023” released: Intelligent bodies are hot, 3D generation emerges, spatial computing and other keywords become keywords
China National Radio, Beijing, December 15 (Reporter Niu Guyue) On December 14, the MEET2024 Intelligent Future Conference was held in Beijing. The conference summarized and looked forward to the latest developments in intelligent technology and officially released the “Top Ten Frontier Technology Trends Report for 2023”.
The report was initiated by QuantumBit Think Tank. Based on long-term tracking of the science and Frontier Technology field and in-depth exchanges with nearly 100 start-ups, industry research institutes, and investment institutions, it nominated the top ten cutting-edge Frontier Technology trends for 2023. The top ten trends in 2023 involve six major areas: AI technology, autonomous driving, space computing, biotechnology, space exploration, and energy innovation.
1. AI Technology
01. Intelligent Agents: A new paradigm of human-computer interaction has been opened by big models
This year’s big model research and development boom has also driven the development of AI agents, becoming their powerful “power engine”.
Currently, many academic studies have taken large models as the cognitive core of AI agents. The reasoning technology and powerful natural language understanding capabilities of large models give agents powerful learning and transfer capabilities, providing a breakthrough technical solution for the research of agents.
This year, there have been breakthroughs in the architecture of intelligent entities, replacing the original reasoning methods of thinking chains and thinking trees with more flexible thinking algorithm technology.
In terms of intelligent agent interaction, the AI town proposed by Stanford also allows the public to see the benefits of multi-agent collaboration, achieving more efficient and intelligent decision-making and actions.
The report predicts that the development of intelligent agents in the future will have two directions:
Autonomous agents and agent simulation will correspond to the needs of the B-end and C-end respectively, and can improve the efficiency of complex processes and provide emotional value. In the near future, software embedded with autonomous agents is likely to change the existing usage mode, from “users adapt to software” to “software adapts to user habits”, and truly become a “personal assistant”.
02. 3D generation enters the emergence stage: new algorithms and new models emerge, and quality, efficiency and controllability change with each passing day
In this year when AGI has become extremely popular, the demand for 3D data in industries such as games, film and television, and XR has continued to grow. Many groundbreaking new algorithms and models have emerged in the field of 3D generation, and academic research breakthroughs in key performance aspects such as quality, efficiency, and controllability are advancing with each passing day.
“3D Gaussian Splatting for Real-Time Radiance Field Rendering”, which was named the best paper at SIGGRAPH 2023 this year, introduces a new method for fast and high-frame-rate real-time rendering – **3D Gaussian Splatting**.
The emergence of this new expression and algorithm is a major breakthrough in the field of 3D generation, and its performance (such as speed, quality, and controllability) is better than NeRF on some data sets.
In October this year, ByteDance’s research team launched a new multi-view diffusion model, MVDream, which can generate geometrically consistent multi-view images based on given text prompts. This new model can apply multi-view priors generated for 3D, greatly improve the stability of existing 2D lifting methods through fractional distillation sampling, and significantly improve the 3D consistency problem.
At the same time, several 3D generation optimization models have appeared this year.
03. Unification of segmentation models: Computer vision is about to usher in the “ChatGPT moment“
The Segment Everything Model (SAM) is an AI model launched by Meta this year. In October, a related paper won the Best Paper Honorable Mention at ICCV23. Influenced by SAM, the CV field and related research will be “hot again” in 2023.
Before the emergence of SAM, all other image segmentation models we could see were proprietary models, and their good performance could not be reflected in other fields. The greatest contribution of SAM is that it can quickly segment unseen images and unify the previously scattered image segmentation models. Some CV practitioners believe that the emergence of SAM is the reason why the field of computer vision has entered the “GPT era”.
04. Embodied intelligence brings new imagination: a new terminal in the ultimate scenario of AGI
This year is the year when humanoid robots and large models emerge simultaneously, and embodied intelligence (EmbodiedAI) has become a new hot spot in AI research.
Several key technologies in the field of embodied intelligence have made breakthroughs in 2023. In June, Fei-Fei Li’s team released a new embodied intelligence achievement, VoxPoser, which can analyze targets and environmental obstacles in 3D space, allowing robots in the real world to perform tasks directly without training.
DeepMind released a new visual language model RoboticTransformer2 (RT-2) this year, which can automatically learn from the Internet and robot data, and successful learning can be converted into general instructions for robot control.
Various breakthrough bionic robot products have emerged around the world. Under the influence of large models, embodied intelligence will enter a new stage.
2. Autonomous Driving
05. End-to-end autonomous driving has become a consensus: BEV+Transformer reconstructs the technical route
This year’s CVPR best paper was awarded to end-to-end autonomous driving, which almost represents a consensus: end-to-end autonomous driving is the future of the industry.
With the development of generative AI, the generalization ability of large language models represented by ChatGPT has been greatly improved, and end-to-end autonomous driving technology has therefore received widespread attention.
End-to-end autonomous driving technology is characterized by the neural network of all modules, has low dependence on rules, and has the ability of intelligent emergence and cross-scenario application potential.
In 2023, urban NOA will become a new direction for autonomous driving manufacturers to compete, and BEV technology based on the Transformer model will become the only choice at present.
BEV+Transformer can directly generate high-quality 3D data that contains more useful information and can generate more reasonable driving instructions. It is expected to gradually become the mainstream paradigm of autonomous driving perception.
In addition, BEV-based objects can complete trajectory prediction and route planning through a top-down view, and can avoid problems such as scale and occlusion under the image perspective, which is also the key to achieving end-to-end autonomous driving.
3. Spatial Computing
06. Spatial computing is clearly defined: consumer-grade products are launched, and the XR full-stack link is opened
In June, Apple officially launched its first spatial computing device, **Apple Vision Pro**, at the WWDC2023 Developer Conference. The advent of this representative consumer-grade spatial computing product truly brought the concept of “spatial computing” to the public and quickly popularized it.
Spatial computing increases the dimension of computing, making the content more three-dimensional and able to almost 100% restore information.
It has become a consensus that spatial computing will become the next generation of computing and will completely transform the traditional information presentation and interaction paradigm.
The outbreak and rapid development of AIGC this year has also begun to solve the problem of rapid production and creation of digital content, realizing a closed loop in the field of spatial computing, and opening up the XR full-stack link.
At the same time, the two major technical routes of spatial computing, VST/OST, are in parallel. After the commercialization of representative spatial computing products, it will drive the gradual improvement of XR developer ecology, content ecology and supply chain. XR devices are also changing to the definition of “spatial computing platform”, which will further promote the accelerated growth of **XR**, **digital twin** and **metaverse**.
4. Biotechnology
07. mRNA opens a new quadrant: providing new solutions for precision medicine and opening a new chapter in biomedicine
This year, the mRNA field achieved remarkable achievements, and related technologies won the Nobel Prize in Physiology or Medicine.
In the past, due to the instability of mRNA technology and human immunity, it could not be applied to new drug development. Until the breakthrough of lipid nanoparticle delivery technology and base modification technology, the difficulty of mRNA vaccine application was solved and rapid mass production was achieved.
Because mRNA has advantages in safety, number of targets, efficacy, manufacturing process and R&D cycle, it can be applied to preventive vaccines, therapeutic vaccines, therapeutic drugs and other fields.
At the molecular design level, a new linear design method was proposed to solve the search problem among numerous mRNA coding sequences. The search range was reduced through a dynamic programming algorithm to achieve rapid positioning of mRNA coding.
In terms of delivery technology, in June, the team of Professor Drew Weissman and Professor Michael J. Mitchell of the University of Pennsylvania synthesized a new **adjuvant lipid** to enhance the delivery efficiency of mRNA.
08. New stage of brain-computer interface testing: product reliability breakthrough, AI improves data decoding capabilities
It has been 50 years since the concept of brain-computer interface (BCI) was proposed. Currently, brain-computer interface technology has entered multiple medical scenarios and has entered the human clinical trial stage.
This year, we have seen progress in the field of brain-computer interfaces focused on increasing the speed of neural decoding, which is closer to the speed of text or speech output by normal people.
Nature published two consecutive papers reporting on new brain-computer interface devices. Neuralink announced in late May that it had received FDA approval to launch its first human clinical study; the world’s first non-human primate interventional brain-computer interface trial was successful in Beijing, marking a breakthrough in my country’s brain-computer interface technology.
5. Space Exploration
09. Reusable rocket enters “China orbit”: engineering difficulties are overcome, and commercial aerospace enters a closed loop model
At the current stage, the recovery and reuse of launch vehicles has become a development trend in the world’s aerospace industry, and many Chinese commercial aerospace companies are also developing their own reusable and reusable models.
The development and scale of space Frontier Technology has always been constrained by launch costs. Reusable rockets can achieve larger-scale and shorter launch cycles under limited production capacity compared to single-launch rockets.
In July this year, the internationally recognized new generation of reusable rocket engine technology – liquid oxygen-methane engine achieved an engineering breakthrough, and Blue Arrow Aerospace’s “Zhuque-2” became the world’s first liquid oxygen-methane carrier rocket launched into orbit.
In April and November, Starship, recognized as “the world’s most advanced reusable rocket”, conducted two test flights, solving key problems such as the separation of the thruster and the rocket, which is of great significance.
6. Energy Innovation
10. Milestone of controlled nuclear fusion: Successful ignition opens up commercial imagination space
This year, there have been continuous breakthroughs in the field of nuclear fusion:
In April, my country’s EAST device achieved 403 seconds of steady-state long-pulse high-confinement mode plasma operation at high-power temperature, setting a new world record for steady-state high-confinement mode operation of a tokamak device.
In August, my country’s new generation of artificial sun, “China Circulation 3”, made significant scientific progress, and for the first time achieved high confinement mode operation under a plasma current of 1 million amperes.
In recent years, there have been continuous breakthroughs in the underlying technologies of nuclear fusion, especially breakthroughs in high-temperature superconducting materials, which are expected to accelerate the commercialization of clean and low-cost nuclear fusion energy.
As a new energy source, nuclear fusion has been defined by industry and academia as the “ultimate energy source” for mankind due to its extremely high energy density and clean characteristics. Once nuclear fusion can achieve large-scale power generation, it will subvert the energy supply form of human society.
In addition, these nominated Frontier Technology fields are worthy of attention: advanced computing power, satellite Internet, synthetic biology, quantum computing, etc.