Search AI Products and News

Inflection-2 outperforms Google's and Meta's models in multiple standard benchmark tests, coming in second only to GPT-4. Inflection-2 will be integrated into the company's personal assistant Pi, released in May. Inflection-2 was trained using 5,000 GPUs, achieving a computational capacity of 10 to the power of 25 FLOPs. Inflection-2 excels in benchmark tests across various fields such as science, Q&A, and mathematics. Inflection AI plans to fine-tune the model within the next 12 months.

Researchers have open-sourced a project called AgentTuning on GitHub, which provides a new method for fine-tuning language models. AgentTuning trains and adjusts language models through interaction trajectories from multiple agent tasks to adapt to different tasks and scenarios. This method can enhance the performance and generalization ability of language models while reducing the workload of manual tuning. AgentTuning has been validated in various natural language processing tasks, including dialogue generation, question answering systems, and summarization.

Stability AI has released the portable language model Stable LM 3B, which has 3 billion parameters and is suitable for mobile devices. Stable LM 3B is compact, efficient, and outperforms other language models of similar size. The model can run on edge devices and home computers, with widespread applications in creative writing and other fields. Stable LM 3B shows significant improvements in text generation, achieving higher energy efficiency and reducing energy consumption. The model is released on the Hugging Face platform.

RA-DIT improves the ability of language models to utilize retrieval information and provide relevant results from the retriever through a two-stage fine-tuning process. RA-DIT outperforms existing retrieval-augmented models in knowledge-intensive zero-shot and few-shot testing. The RA-DIT65B model shows improved performance in tasks that require knowledge, demonstrating the effectiveness of fine-tuning. The enhancement of retrieval capabilities brought by RA-DIT has shown good results in knowledge-intensive evaluations, confirming the effectiveness of fine-tuning.

The Japanese government collaborates with tech giants such as NEC, Fujitsu, and SoftBank to invest hundreds of millions in developing a culturally sensitive Japanese language model. The new model will be trained on the national supercomputer Fugaku, boasting at least 30 billion parameters, with performance expected to significantly outshine existing models. The goal is to address shortcomings of current models in the Japanese market, such as generating rare characters and culturally inappropriate communication norms. Researchers have developed the Rakuda ranking system to assess the model's sensitivity to Japanese culture, with GPT-3.5 currently ranking first. The Japanese government also...

Xiaomi has released the large-scale pre-trained language model MiLM-6B, which has 6.4 billion parameters. MiLM-6B has achieved the best performance in its size on both C-Eval and CMMLU. MiLM-6B performs excellently and is suitable for various subjects and difficulty levels.