Meta AI's latest project, Llama 3, focuses on training with synthetic data rather than human-written answers to optimize model performance in areas such as code generation, mathematical reasoning, multilingual processing, and long text handling. Llama 3 generates synthetic data for code generation through three methods, employs research methods for mathematical reasoning, and collects high-quality human annotations for multilingual pre-training. Furthermore, Llama 3's training is enhanced using Brave Search, Wolfram Alpha, and Python interpreter.

Cybercriminals are starting to use generative artificial intelligence for attacks, including Meta's Llama2 large language model. The group known as Spider is using Llama2 to generate attack scripts targeting a North American financial services company. Security firms warn that the current ability to detect AI attacks is limited, but trends of malicious use of AI are expected to increase. The report notes that attempts to maliciously use large language models may still exist in 2023, but the generated results could be ineffective.

Tsinghua University's latest research introduces the AgentTuning method to enhance LLM's agent capabilities. The AgentInstruct agent dataset has been constructed, containing high-quality interactive trajectories and CoT reasoning chains. A hybrid instruction fine-tuning strategy was adopted, using AgentInstruct in conjunction with general instructions. Results indicate that AgentLM outperforms Llama 2 significantly in both held-in and held-out tasks. The 70B version of AgentLM excels in agent tasks.

Meta has released the Llama2-Long model, which does not increase computational demand when processing long texts while still maintaining excellent performance. The model employs continuous pre-training, improved positional encoding, and data mixing strategies to reduce computational overhead by up to 40%. It performs outstandingly on both long and short tasks, surpassing other long-context models and has the potential to revolutionize the field of natural language processing. The model shows significant improvements in encoding, mathematical, and knowledge-intensive tasks, even surpassing GPT-3.5. The release of the Llama2-Long model provides a powerful solution for handling long texts.

By expanding the vocabulary, selecting data, and employing a multi-stage training strategy, the Chinese version of LLaMA2 was trained in 15 hours at a cost of a few thousand yuan. The performance of LLaMA2 on various Chinese tasks has significantly improved, reaching an advanced level compared to models of the same scale. The building process, code, and weights are all open source, making it transferable to other languages and fields. Achieve low-cost large model training by training the Chinese version of LLaMA2 in just 15 hours with a cost of a few thousand yuan. All training code and pre-trained weights are open source and can be directly applied to other languages and fields for rapid low-cost large model training.

['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.']

IBM announced that it will host MetaPlatform's AI language model MetaLlama2 on its enterprise AI platform WatsonX. Llama2 is the commercial version of Meta's open-source AI language model launched in July this year, distributed through Microsoft's Azure cloud services. Currently, in watsonx.ai, AI builders can utilize models from the IBM and Hugging Face community.

IBM plans to host Meta's Llama2 chat model with 7 billion parameters on watsonx.aistudio. IBM has a collaborative foundation with Meta in the field of AI open innovation. The Llama2 model will be part of IBM's strategy to offer both third-party and proprietary AI models.