1. “Impression AI” mainly has three functions in the Evernote desktop platform, which are “Creating something out of nothing”, “Everything in Life”, and “one-click mind-map generation”.
  2. 24x faster than HuggingFace: UC Berkeley’s flagship open source God LLM inference system, VLLM, is harnessing PagedAttention to be 24x faster than HuggingFace/Transformers, and halving the number of Gpus.
  3. Neusoft has launched a number of AI+ medical applications for the medical field, including the big model in the medical field, the Flymark medical image annotation Platform 4.0, and the Web-based virtual endoscope.
  4. Mediatek announced that the report on the AI large order received from Google is an erroneous report based on market rumors that has not been fully verified, affecting the company and investors’ interests. If any damage is caused by wrong reporting, the company’s rights will be protected according to law.
  5. LTX, a subsidiary of U.S. financial technology company Broadridge, recently announced the launch of a chatbot APP based on GPT-4’s big model for corporate investment.
  6. Palo Alto Networks, a US cybersecurity company, warned that it had found a malicious Android app that mimics ChatGPT. The company also noted that as of June 19, ChatGPT’s official app is only available for iPhone, and there is no Android client.
  7. Changhong TV’s large model “Changhong Super Brain” was officially released, with three core capabilities of perception, understanding and creation. At the same time, Changhong launched the world’s first artificial intelligence TV, released the world’s first multi-mode interactive TV, and realized the world’s first wakefree TV interaction capability.
  8. The AI code generation star project gpt-engineer night fire, a few days more than 27 k star, it is a tool, according to the instructions of the generated code AI you only need to words, You can build the entire code base directly.
  9. The latest Revelations of GPT-4 parameters: The well-known American hacker George Hotz revealed in an interview that GPT-4 is composed of 8 220B models, and the 8x220B expert model is trained with different data/task distributions and 16-iter reasoning
  10. Meta has announced a new AI model called Voicebox, which it claims is the most versatile model for speech generation to date. Meta says that while there are many exciting use cases for generative speech models, they are not making voicebox models or code public at this time due to the potential risk of possible abuse.
  11. Research by MIT and Microsoft has found that GPT-4 can self-correct error codes, while GPT-3.5 can’t. Similarly, other studies have shown that only large, mature models seem to be capable of emergence