1. The Qualcomm team demonstrated the phone’s ability to run 100 million scale models in just 15 seconds. They optimized the open source model specifically based on mobile AI hardware and software technology. Compared to relying solely on cloud computing power to process AI models, the computing power of mobile devices can also be fully utilized.
  2. Google launches lifelike virtual fitting. Google has announced a new virtual fitting feature that will allow users to see how real models of different shapes and sizes look in clothing. This feature is driven by a new generative AI model that creates realistic, high-quality images with details such as folds, folds, stretches, and wrinkles. The model was trained using Google’s Shopping Graph dataset, which includes information on products, brands, and reviews.
  3. Microsoft Bing suddenly upgraded its image recognition function. After uploading a picture, you can program and write code, do problems and drawings, and even see a doctor, but it is still in a small scale test.
  4. The impact of AI on human artistic expression is discussed. In the journal Science, academics explore the role of AI in human artistic expression. Generative AI art raises questions about creativity, ownership, and ethics. Some AI tools help in the creative process, but there is also a risk of diminishing the value of human effort and blurring the boundaries of creative ownership.
  5. ChatGPT can’t tell 25 jokes: German academics did a big test on GPT3.5 and found that it can only tell 25 jokes. 90% of the 1,008 results were variations on 25 jokes, with a slight change in wording or phrasing.
    FINE-GRAINED RLHF (fine-grained human feedback reinforcement learning) fine-grained RLHF (fine-grained human feedback reinforcement learning) has a better fine-grained effect than ChatGPT and has saved the language model from nonsense.
  6. Meta plans to commercialize the LLaMA model. Previously, Meta’s open source model, LLaMA, attracted a lot of attention and became the basis for various research projects. Now, Meta is ready to commercialize LLaMA, introducing it into the company’s commercial license, a major development in the field. Companies will have the freedom to use and profit from Metal’s AI models.
  7. Meta has launched MusicGen, a text music generation model, which is available for non-commercial use and the demo has been opened.
  8. OpenAI updates GPT-4 and other models, adds API function calls, and the price is reduced by up to 75%. OpenAI’s CEO revealed the recent development route in the global tour speech, the primary task in 2023 is to launch cheaper and faster GPT-4, longer context window, and the focus in 2024 is multi-mode.
  9. OpenAI warns that Microsoft’s AI is poor at telling facts. In an AI plot twist, Microsoft was warned by partner OpenAI that there were dangers in integrating GPT-4 into Bing search and that more training was needed. However, Microsoft did not heed the warnings and launched the technology earlier this year, leading to a number of “illusions” that damaged the reputation of the Bing search assistant.
  10. GPT-4 full marks passed MIT undergraduate Math exam! This time, not only GPT-4 and GPT-3.5, but also StableVicuna-13B, LLaMA-30B and LLaMA-60B were tested on 228 randomly generated questions from MIT’s data set, with the highest score being GPT-4 with a 100% scoring rate. The most mediocre performer was LLaMA-30B, which scored only 30% of the points.