Golden Ten Data on April 28, according to Danghong Technology, at the "2024 Zhongguancun Forum - UHD Audiovisual Technology Innovation and Development Forum" held on the afternoon of April 27, Danghong Technology and Beijing Economic and Technological Development Zone jointly released the BlackEye large model base and application scenarios. BlackEye integrates long kinds of Depth neural network components, including Transformer, Diffusion and other components, through text, image, video and audio, 3D model and other long modal encoding, decoding, long modal latent short alignment, long modal language reasoning and generation and other technologies, to achieve inference and prediction generation between different modal information.

In 2023, significant progress has been made in the field of generative AI, and vendors will continue to invest, especially in terms of hardware, as reported by the Science and Technology Innovation Board Daily on January 24. It is expected that by 2025, a dedicated AI chip with a performance of more than 100 TOPS will be launched. However, Canalys predicts that 70% of these chips will not be fully utilized due to talent shortages, especially in the areas of high-performance, low-power model optimization, neural architecture search, and tuning for edge devices.

According to a report by IT House on January 18, Google's DeepMind recently published a paper in Nature, demonstrating the latest AI system AlphaGeometry, which is comparable to the world's top students in solving mathematical geometry problems. AlphaGeometry is a neural symbol system that combines language learning and deductive reasoning. The system combines the predictive power of a neural language model with a rule-based deductive engine to find a solution. In the International Mathematical Olympiad (IMO) for high school students, AlphaGeometry answered 25 out of 30 questions correctly, with a correct rate of 83%. For comparison, the average human gold medalist answered 25.9 questions correctly, compared to the previous best record of an AI model successfully answering 10 geometry questions. This result is very close to the gold standard for human entrants, and Deepmind believes it highlights the growing capabilities of AI in mathematics. Deepmind researcher Quoc V Le, who worked on the project, told the Financial Times that the new system is a key step towards building artificial general intelligence (AGI).

According to IT House on January 12, Swarovski today displayed the Optik AX Visio telescope at the CES 2024 venue, which is said to be the world's first binoculars capable of identifying birds, priced at $4,799. It is reported that this 10 x 32 telescope has 10x magnification, up to 15 hours of battery life, and has a built-in NPU neural processing chip, which claims to be able to identify 9,000 species of birds, and can label bird species and information in real time.

According to Jin Shi news, on January 8, local time, AMD announced the launch of the first AI CPU (CPU) on a desktop. AMD noted that the Ryzen 8000G series chips are expected to be available to DIY customers and system integrator partners starting January 31, with systems from individual OEMs expected to be available in the second quarter of 2024. AMD said it will also bring a dedicated AI neural network processor to desktop processors for the first time with the launch of Ryzen AI.

Recently, an artificial intelligence neural network found an unusual feature in one of Raphael's famous paintings, one of the faces turned out to be not from Raphael himself, as reported by IT House on December 31. The painting, titled Madonna della Rosa, has long been debated by scholars as to whether the painting is an original Raphael. While judging the provenance of a work of art requires a synthesis of evidence, a new analysis method based on artificial intelligence algorithms supports those who believe that at least part of the painting came from another artist. Researchers from the United Kingdom and the United States trained a custom AI algorithm specifically on details such as brushstrokes, colors, shadows, and more in Raphael's known work. This method has previously been shown to be 98% accurate in identifying Raphael's paintings, and this time they not only had the AI analyze the entire painting, but also studied the individual faces within it individually. The results show that the faces of the Virgin, the Child, and St. John are all by Raphael himself, with the exception of St. Joseph's face in the upper left corner, which is distinctly different from the others. Researchers speculate that St. Joseph's face may have been painted by one of Raphael's students, Giulio Romano, and while this speculation still needs further research, there is no doubt that AI technology offers a new tool for unraveling classic paintings.

Recently, the 37th Annual Conference on Neural Information Processing Systems was held in New Orleans, USA, as reported by IT House on December 31. At the meeting, a scientific team proposed that the artificial intelligence model Transformer has a function similar to the NMDA receptor. In addition to this discovery, they also designed new nonlinear functions that allow the Transformer model to simulate the kinetic performance of NMDA receptors. The NMDA receptor is a "smart" channel that promotes the formation of memories and skills. If glutamate is present in the brain, nerve cells become excited, delivering neurotransmitters to NMDA receptors. At the same time, magnesium ions are responsible for the gatekeeping, and when magnesium ions are squeezed out of the NMDA receptors, the neurotransmitter is allowed to enter the next nerve cell. In the hippocampus, the smooth transfer of matter ensures that short-term memory is effectively converted into long-term memory, and magnesium ions are responsible for selectively allowing matter to pass through. In the new study, the scientists found that there is a process similar to the magnesium ion gatekeeping in the NMDA receptor in the Transformer model, and that adjusting the parameters of the Transformer model can enhance the memory capacity of the model. This finding prompted researchers to think about whether it is possible to develop a method to regulate the memory consolidation process of the Transformer model by studying the mechanism of magnesium ion gatekeeping in the NMDA receptor, and make the model have systematic long-term memory and even consciousness.

As reported by IT House on December 16, Apple recently released the DeepPCR machine learning algorithm, which can accelerate the inference and training of neural networks by processing regular sequential operations in parallel. The team used a parallel cyclic reduction (PCR) algorithm to retrieve the solution, reducing the computational cost of the sequential process from O(L) to O(log2 L), reducing complexity and increasing the speed of operation. Neural networks are already capable of handling complex tasks such as text or image composition, segmentation, and classification. However, due to the computational demand, it can take days or weeks for neural network training and feedback results. In the process of neural network processing, parallelization technology is widely adopted, which can accelerate the training and inference speed.

According to Sina Technology on December 12, 20 papers by Ant Group were included at the NeurlPS conference, the world's top AI conference held recently. It covers multiple artificial intelligence and machine learning frontiers such as computer vision, natural language processing, graph neural networks, image processing, etc., and more than 70% of the papers focus on the field of AIGC (generative AI). Ant Group revealed that in the past five years, Ant Group has published nearly 500 papers in top international academic journals and academic conferences, more than 80% of which focus on the field of AI.

According to a report by Webmaster's Home on December 7, Liquid AI, a startup incubated by MIT, recently successfully completed a seed round of nearly $40 million, with the participation of a number of venture capital firms and well-known angel investors, with a valuation of $303 million. The company's goal is to build an entirely new type of artificial intelligence, known as liquid neural networks. This innovative technology is based on a liquid neural network architecture and aims to provide a more streamlined, explainable, and dynamically adaptable AI system. Liquid AI plans to commercialize the technology to provide customers with private AI infrastructure and platforms in areas such as autonomous driving, weather forecasting, medical data analytics, and more.

According to the qubit report, Google's DeepMind has developed a new material discovery tool, with the help of AI tools, discovered 2.2 million theoretically stable new crystal materials, not only increased the accuracy of predicting material stability from 50% to 80%, but also 380,000 have been put into testing, and the paper has been published in Nature. Google's DeepMind says the research equates to nearly 800 years of knowledge, given that only 28,000 stable materials have been discovered in the last 10 years. This AI tool is called GNoME (Graph Networks for Materials Exploration), which is a Graph Neural Network (GNN) that uses active learning to discover new materials. According to reports, researchers have synthesized 736 materials in the laboratory to prove that the crystals calculated by GNoME can be synthesized. In addition, synthetic materials may also be used as a guide for the design of new materials, or as new datasets to train and optimize other AI models.

Researchers at the University of Chicago, in collaboration with Snap Research, have launched an artificial intelligence method called 3D Paintbrush that automatically generates textures for local semantic areas on the mesh through text descriptions, as reported by Webmaster's House on November 27. This innovation allows users to control texture editing through intuitive, free-form text input, describing their edits for a variety of meshes. Using multi-layered perceptrons defined on a 3D surface, this method creates neural textures that produce smooth changes in 3D, reducing the noise of the texture, and having super-resolution capabilities.

Anatoly Yakovenko, co-founder of Solana, said on the X platform, "Pyth Network is really precious to me. He explained the importance of information monopoly and Pyth in traditional finance. He claims that early in his career as an engineer, he learned a lot, including Haskell, nemerle, f#, ruby, python, clojure, the basics of options and trading, neural networks, FFTs, etc... But he simply didn't have any access to Yahoo Finance's data, and traditional finance was built entirely on information blockade, and Pyth revolutionized that.

According to NetEase Technology on November 8, three Google researchers pointed out in a paper recently submitted to the Preprint Document Library (ArXiv) that deep neural network transformer, the underlying technology in the field of artificial intelligence, is not good at generalization. The three authors wrote: "When a task or function needs to be outside the scope of the pre-trained data, transformers exhibit various failure modes, and even simple task extensions are found to reduce their inductive power." According to the paper, deep neural network transformers are good at performing tasks related to training data and are not very good at handling tasks beyond that. For those looking to achieve Artificial General Intelligence (AGI), this issue cannot be ignored.

According to titanium media reports, at the "New Generation of Artificial Intelligence Frontier Forum" held by Peking University recently, Huang Tiejun, a professor at the School of Computer Science of Peking University and dean of Beijing Zhiyuan Institute of Artificial Intelligence, said that the current problem with AI security debate is that everyone has a different basic understanding of AI, and many people who engage in scientific thinking do not believe that artificial general intelligence (AGI) will be realized, while another technical thinking, including Yang Likun, believes that AI development is feasible. For the development of AI large model technology in the future, Huang Tiejun gave expectations for the 3-year, 10-year, and 20-year stages: in the next 3 years, large models will be expanded to general intelligence such as vision, hearing, embodied intelligence, and action, and become the "brain" of embodied intelligence. In the next 10 years, mankind will enter the era of intelligence (ability) in an all-round way; In the next 20 years, the current large model will be data-driven static intelligence, and the future will be dynamic intelligence driven by the spatio-temporal environment, and the intelligent carrier will be upgraded to a brain-like spiking neural network. In addition, Al security must be at least as important, otherwise humanity is at great risk.

According to IT House reported on October 20, Suzhou Guoxin Technology announced that Guoxin Technology and Hong Kong Applied Science and Technology Research Institute recently signed a memorandum of cooperation and project research and development support agreement, and the two parties will establish "ASTRI Hong Kong - Suzhou Guoxin New AI Chip Joint Research Laboratory". According to reports, SMIC will invest in the next three years to support the next generation of AI chip technology, and ASTRI Hong Kong will transfer the existing NPU-related technology to SMIC to assist it in launching new products. The joint projects include the development of next-generation AI chips and neural network processors for the development of AI chips for automotive electronics, industrial control, and robotics applications. The two parties said that the cooperation will help both parties to land and explore potential opportunities in the fields of automotive electronics, industrial control and robotics, and accelerate business expansion in the field of AI.

According to the "Science and Technology Innovation Board Daily" reported on October 12, at the 2023 China Mobile Global Partner Conference held today, ZTE released the industry's first large-model-based "Zhiyu" SMS anti-fraud governance system, and cooperated with Shanghai Mobile to carry out SMS anti-fraud governance exploration, and applied large-model technology to the anti-fraud field. According to reports, under the premise of obtaining customer authorization, the "Zhiyu" SMS anti-fraud governance system effectively improves the recognition accuracy, interception rate and recall rate of fraudulent SMS by building an integrated scheme with the anti-fraud model as the core, generalized feature neural network and SNS social feature analysis as a coordinated integrated scheme, and the root cause dimensions such as the intention and semantics of fraudulent SMS.

According to a report from Webmaster Home on October 10, researchers from the Chinese Academy of Sciences and Yale University jointly proposed a new framework called "Thought Propagation", aiming to improve large-scale neural network models (such as GPT-4, PaLM) reasoning ability, enabling it to think analogically more like humans. This framework is inspired by human cognition, that is, when people face new problems, they often compare them with similar problems that have already been solved to derive Solution strategies.

According to a report by IT House on September 29, John Rayfield recently resigned from Intel and joined AMD to be responsible for developing AI workloads on PC chips. Rayfield's previous position at Intel was as the customer AI project director, where he led the research and development of low-power AI technology including neural processing units (NPUs). Related results have been integrated into Intel's upcoming Ultra "Meteor Lake" processing in the vessel. A search on LinkedIn revealed that Rayfield resigned from Intel this month and joined AMD as corporate vice president of AI chips. As of press time, Redfield, Intel and AMD have not responded to foreign media inquiries.