➀ Microsoft is accused of scraping user-generated content from Word and Excel to train its AI models; ➁ The 'Connected Experiences' feature is enabled by default and requires specific settings to opt-out; ➂ Microsoft denies using customer data for AI training, stating the feature is for internet-required functions like document co-authoring.

➀ Phison promises a new $50,000 workstation for AI training and AI tuning; ➁ Phison's President Michael Wu explains the shift from $1 million to $1.5 million workstations; ➂ Phison's aiDAPTIV+ Pro Suite simplifies AI model fine-tuning and chatbot creation.

1. Apple uses Google TPUv4 and TPUv5p clusters for AI model training instead of NVIDIA GPUs. 2. Apple utilized 2048 TPUv5p chips for iPhone-related AI models and 8192 TPUv4 chips for server AI models. 3. Apple's research paper suggests the possibility of creating larger, more sophisticated AI models with Google TPUs.

1. Meta's Llama 3 405B model training on 16,384 NVIDIA H100 GPUs experienced half of its failures due to HBM3 memory issues. 2. The 54-day training run encountered 419 unexpected component failures, averaging one every 3 hours. 3. Meta's team improved efficiency with proprietary diagnostic tools and managed 90% effective training time despite disruptions.

1. Microsoft's AI CEO states that most online content is considered 'freeware' for AI training. 2. The lack of transparency in data sources for AI training has led to lawsuits. 3. There is a legal grey area regarding the use of copyrighted content for AI training.