Recently, a new study from the University of Michigan found that an energy-efficient method for training large language models can achieve similar results in the same amount of time while reducing energy consumption by 30%. This method can save enough energy to power 1.1 million American households by 2026. The researchers developed a software tool named Perseus that identifies the critical path, which is a series of subtasks requiring the longest time to complete. Then, Perseus reduces the processor speed on non-critical paths so that all tasks can finish simultaneously.

In the era of driven artificial intelligence, the demand for high-quality data by large language models (LLMs) like GPT-3 and BERT is increasing. However, manually sorting this data from the web is not only time-consuming and labor-intensive but also often difficult to scale. This poses significant challenges for developers, especially when large amounts of data are required. Traditional web crawlers and data scraping tools have limited capabilities in extracting structured data; although they can collect webpage data, they often cannot format it in a style suitable for LLM processing.

With the rapid advancement of technology, artificial intelligence (AI) has entered our kitchens to help generate recipes. Companies like Apple, Google, and Samsung have rolled out AI-powered recipe generators, allowing users to request dinner party menus based on available ingredients through Siri. However, this has become a nightmare for food bloggers who are deeply invested in recipe development. To compare the merits of 'real recipes' versus AI-generated recipes, numerous chefs on YouTube and TikTok have embarked on a unique challenge.

MianBi Intelligence has recently launched MiniCPM 3.0, a high-performance AI model that operates on devices with just 2GB of memory, heralding a new era of edge AI experience. The MiniCPM 3.0 model features 4 billion parameters and outperforms GPT-3.5, providing AI services on mobile devices comparable to those of GPT-3.5. This allows users to enjoy fast, secure, and feature-rich local AI services without relying on cloud processing, ensuring a smoother and more private intelligent interaction experience.

Google launched the Gemma22B model with 2 billion parameters, demonstrating exceptional performance across various hardware environments, particularly outperforming models like GPT-3.5-Turbo-0613 and Mixtral-8x7b in the chatbot arena. The model is optimized with the NVIDIA TensorRT-LLM library for efficient deployment in diverse scenarios, from edge devices to powerful cloud environments. To enhance safety, Google introduced the ShieldGemma safet....

Microsoft has announced the launch of serverless fine-tuning capabilities for the Phi-3 small language model, enabling developers to easily adjust model performance on the Azure AI platform without managing underlying infrastructure, and it is free for initial use. The Phi-3 model, with 3 billion parameters, is suitable for enterprise developers, offering efficient performance at a lower cost. Although it has fewer parameters compared to OpenAI's GPT-3.5, it performs similarly across various applications. Serverless fine-tuning simplifies operations, reduces barriers, and supports the fine-tuning of small to medium models.

Recently, a model named Arcee Spark based on Qwen2 has undergone fine-tuning on 1.8 million sample data, featuring a 128k token context. The release of Arcee Spark has attracted widespread attention, particularly sparking a surge of interest among professionals in the field of artificial intelligence. In benchmark tests such as MT-Bench, it has performed exceptionally well, achieving the highest scores among similar models and even surpassing GPT-3.5 on multiple tasks. It is reported that Arcee

Google successfully attacked the OpenAI GPT-3.5-turbo model at a cost as low as 150 Yuan. The attack method was simple; by making fewer than 2000 queries through the API, the entire projection matrix and hidden dimensions were obtained. This attack reveals that even large language models may face security threats. OpenAI has modified the model API to prevent similar attacks from happening again.

Experiments have found that OpenAI GPT-3.5 exhibits racial bias in resume screening. Test results indicate that GPT-3.5 is more inclined towards certain ethnic names, which could impact hiring decisions. The experiments also revealed gender and racial preferences for GPT in different job positions. OpenAI responded that companies usually take measures to mitigate bias when using its technology.

60% of GPT-3.5 outputs have plagiarism issues. A Copyleaks report indicates that 60% of outputs exhibit plagiarism. GPT-3.5 shows different similarity scores across various subjects. The plagiarism issue has led to legal disputes between OpenAI and The New York Times.

Plagiarism detection company Copyleaks released a report stating that 60% of the output generated by the GPT-3.5 model contains some form of plagiarized content. Copyleaks uses a proprietary scoring method that considers factors such as text similarity, minor alterations, and paraphrasing to calculate a 'similarity score.' The highest similarity scores were found in fields like computer science, physics, and psychology, while the lowest were in drama, humanities, and English. OpenAI has stated that it has taken measures to limit inadvertent memory usage, and its terms also prohibit deliberate reverse-engineering of content. Aside from the New York Times lawsuit, creators are also attempting to assert rights regarding AI training.

The research shows that large language models like ChatGPT often repeat misinformation such as conspiracy theories and harmful stereotypes. They found that GPT-3 frequently makes mistakes, contradicting itself within a single response while repeating harmful misinformation. Researchers employed four different survey templates covering six categories of facts and misinformation.

Mistral recently launched the Mixtral8x7B model, which has caused a huge stir in the AI community. This model surpasses the well-known GPT-3.5 and Meta's Llama2 family, stirring up significant buzz in the open-source AI field. The performance of Mistral8x7B has left early users amazed, but it has also raised concerns about safety. The released Mistral open-source model Mixtral8x7B outperforms the well-known GPT-3.5 and Meta's Llama.

A study by the University of California found that GPT-4 did not pass the Turing Test and performed worse than human intelligence. The GPT-3.5 model lost in conversations with the 60-year-old AI chatbot ELIZA. The research used the Turing Test to explore the human-like characteristics of AI models, sparking academic interest. Investigators' judgments were primarily based on language style and social-emotional traits, with no significant prediction based on educational level.

["Microsoft CEO Satya Nadella announced at the Cloud Developer Conference that Azure will launch updated generative AI models.", "The new models include GPT-4 Turbo and GPT-3.5 Turbo, providing more advanced and efficient AI tools.", "GPT-4 Turbo is three times more cost-effective and two times more productive, introducing new features like parallel calls and JSON mode.", "Azure OpenAI services expand to eastern Norway, including new regions."]

["Microsoft paper reveals that the parameter count of GPT-3.5 is only 20B, a huge gap from the official claims.","CodeFusion's small model performs extraordinarily well with just 75 million parameters.","Amid controversy, deep reflections on model scale and performance.","The competition in the AI field is fierce, with ChatGPT and ClaudeAI vying for supremacy.","GPT-4 reviewer Terence Tao uses GPT-4 to solve mathematical problems."]

OpenAI Chief Scientist Ilya conducted a deep interview before the release of GPT-3. He discussed the evolution of neural networks and their analogy to the human brain. He is confident about the development and future of deep learning, stating that breakthroughs will require large-scale computation and persistent efforts. He believes there are still many beautiful and mysterious properties of deep learning yet to be discovered.

['Kunlun Wanwei's self-developed TianGong large model achieved an 80% accuracy rate in the Benchmark GSM8K test.', 'The inference capability of the TianGong large model exceeds that of GPT-3.5 and LLaMA2.', 'The TianGong large model performs exceptionally well across multiple datasets.', 'The problem-solving approach of the TianGong large model is simpler and clearer.', 'The TianGong large model will provide strong intelligent support for users and enterprises.']

ELYZA has released the Japanese LLM 'ELYZA-japanese-Llama-2-7b' based on Meta's Llama 2, with a parameter count reaching 7 billion, performance comparable to GPT-3.5. The model has undergone additional pre-training and unique post-training, achieving the highest scores in a 5-level manual evaluation. Although it has not yet reached the level of closed LLMs, it is already on par with GPT-3.5. ELYZA has successfully developed LLMs in other languages, including English.