[July 29th Big Model Daily] Information Olympic 8-minute AI video goes viral President Bach: Thanks to China’s black technology

Olympic 8-minute AI video goes viral President Bach: Thanks to China’s black technology

https://news.miracleplus.com/share_link/34949

During the Olympic Games, an AI technology-enhanced short film “She Never Fades” co-produced by Alibaba and the International Olympic Committee attracted widespread attention. The short film uses AI technologies such as image super-resolution, coloring, and interpolation based on generative technology to show the important role and achievements of women in Olympic history. Through the stories of female athletes such as Susan Lenglen, Zhang Shan, and Catherine Switzer, the article shows how they broke through and fought hard in the Olympic arena, promoting the development of women’s sports. Alibaba Cloud’s AI technology team’s work in restoring historical images has enabled these precious historical moments to be presented to modern audiences in a clearer and more vivid form.

Llama3.1 training fails once every 3 hours on average, temperature fluctuations affect H100 GPU cluster throughput

https://news.miracleplus.com/share_link/34950

During Llama 3.1 model training, failures occurred once every 3 hours, and most failures were caused by hardware problems in the H100 GPU cluster, especially GPU and HBM3 memory failures. To improve effective training time, the team reduced task startup and checkpointing time, and used tools such as PyTorch’s NCCL flight recorder for rapid diagnosis and problem solving. Despite hardware failures and environmental factors such as temperature changes that affect throughput, the Llama 3.1 team maintained more than 90% of effective training time. Meta is facing technical challenges of large-scale AI clusters, including power, network design, parallelism, and reliability.

NVIDIA Llama 3.1 Synthetic Data Technology and RAG System Fine-tuning Practice Guide

https://news.miracleplus.com/share_link/34951

NVIDIA’s technical blog post introduces the use of Llama 3.1 405B large language model to generate synthetic data to improve the performance of AI models in different industries. Synthetic data does not create new information, but creates different data variants by transforming existing information. This process can help the model improve in accuracy. The article details the methods of using LLM for knowledge distillation and self-improvement, and how to apply synthetic data through pre-training, fine-tuning and alignment steps. In addition, it introduces how to use synthetic data generated by LLM to improve other models, such as RAG pipelines, and shows a specific method for evaluating the performance of RAG pipelines through a case study.

Original link: https://developer.nvidia.com/blog/creating-synthetic-data-using-llama-3-1-405b/?ncid=so-twit-933996&linkId=100000275486093

Next-generation AI database: Infinity version 0.2 leads a new era of hybrid search and sorting

https://news.miracleplus.com/share_link/34952

Infinity database version 0.2 introduced sparse vector and tensor data types, realized multi-way recall and hybrid search functions for full-text search, vector search and tensor search, and greatly improved the retrieval quality of the RAG system. The article emphasizes the limitations of pure vector search in precise semantic expression, and introduces how to optimize search results through hybrid search (combining full-text search with sparse vector search). In addition, the article also discusses the characteristics of the ColBERT ranking model and how to implement end-to-end ColBERT applications in Infinity, including support for Tensor data types and processing solutions for very long texts. Through the evaluation of the MLDR dataset, Infinity demonstrated its excellent performance and ranking quality, especially in comparison with Elasticsearch, Infinity showed excellent full-text search performance.

TDCon2024: Exploring the infinite possibilities of time series big data

https://news.miracleplus.com/share_link/34953

Time series data is becoming increasingly important in the AI ​​era. With the surge in data volume and the demand for real-time analysis, traditional processing tools have become difficult to cope with. Based on many years of technical research and development experience, Mr. Tao Jianhui, the founder of Taosi Data, founded Taosi Data and developed TDengine, which is specifically aimed at time series big data processing problems. Since its open source in 2019, TDengine has quickly attracted the attention of developers around the world and has been widely used in multiple industries. At the TDCon2024 conference, industry experts and elites gathered in Beijing to discuss the role of time series data in promoting the digital transformation of enterprises.

Related links: https://www.zhihu.com/people/jefftao

Israel’s new chip revolution

https://news.miracleplus.com/share_link/34954

The Israeli chip industry, with its high-tech industry as the economic pillar, occupies an important position in the global semiconductor field. Despite limited geographical and demographic conditions, Israel has the highest number of engineers per capita and the largest number of high-tech enterprises per unit area in the world. Israel not only has a deep history of the semiconductor industry, but also has a complete industrial ecosystem chain, attracting global technology giants including Intel and NVIDIA to establish R&D centers in their home country.

In cutting-edge technology fields such as AI, quantum computing and photonic computing, Israel has produced many innovative start-ups, such as Hailo Technologies and Quantum Transistors, which have played a key role in promoting the new revolution in the chip industry. The Israeli government has cultivated a large number of high-quality talents for the semiconductor industry through educational policies and capital support, and has supported the development of start-ups through policies such as legal protection and tax incentives, promoting Israel’s unique competitiveness in the semiconductor industry. These measures have ensured Israel’s leading position in the global semiconductor industry and promoted the country’s technological innovation and economic growth.

AI is completely defeated by human doctors! Research finds that large models make sloppy and unsafe clinical decisions, with a minimum accuracy rate of only 13%

https://news.miracleplus.com/share_link/34955

Large language models (LLMs) such as Llama 2, Clinical Camel, and Meditron, although they perform well in medical licensing exams, are much less accurate than human doctors in actual diagnosis. The study is based on the MIMIC-IV database, involving 2,400 real cases of 4 common abdominal diseases, and evaluates the performance of LLM in clinical diagnosis, compliance with diagnostic and treatment guidelines, interpretation of laboratory results, and robustness to information changes. The results showed that the LLM’s diagnostic accuracy was 73%, while the doctors’ accuracy was 89%. In the diagnosis of cholecystitis, the LLM’s accuracy was as low as 13%. In addition, as the case information increased, the LLM’s diagnostic accuracy decreased, and it might recommend examinations that posed a risk to the patient’s health. The study also found that the LLM performed poorly in following diagnostic guidelines and arranging necessary examinations, and might make hasty diagnoses without a full understanding of the case. Therefore, the study believes that the current LLM requires a lot of clinical supervision from doctors, and recommends that AI experts work with clinicians to further develop and optimize models suitable for clinical practice.