All tags
Model: "mixtral-8x22b-instruct-v0.1"
Meta Llama 3 (8B, 70B)
llama-3-8b llama-3-70b llama-3-400b stable-diffusion-3 mixtral-8x22b-instruct-v0.1 vasa-1 meta-ai-fair stability-ai boston-dynamics microsoft mistral-ai hugging-face transformer tokenization model-training benchmarking robotics natural-language-processing real-time-processing synthetic-data dataset-cleaning behavior-trees ai-safety model-accuracy api model-release humor helen-toner
Meta partially released Llama 3 models including 8B and 70B variants, with a 400B variant still in training, touted as the first GPT-4 level open-source model. Stability AI launched Stable Diffusion 3 API with model weights coming soon, showing competitive realism against Midjourney V6. Boston Dynamics unveiled an electric humanoid robot Atlas, and Microsoft introduced the VASA-1 model generating lifelike talking faces at 40fps on RTX 4090. Mistral AI, a European OpenAI rival, is seeking $5B funding with its Mixtral-8x22B-Instruct-v0.1 model achieving 100% accuracy on 64K context benchmarks. AI safety discussions include calls from former OpenAI board member Helen Toner for audits of top AI companies, and the Mormon Church released AI usage principles. New AI development tools include Ctrl-Adapter for diffusion models, Distilabel 1.0.0 for synthetic dataset pipelines, Data Bonsai for data cleaning with LLMs, and Dendron for building LLM agents with behavior trees. Memes highlight AI development humor and cultural references. The release of Llama 3 models features improved reasoning, a 128K token vocabulary, 8K token sequences, and grouped query attention.