Meta 70B Llama Outperforming GPT-4

Meta has unveiled Code Llama 70B, a leap in open-source code generation models, surpassing GPT-4 with a HumanEval score of 67.8. Building on Llama 2’s architecture, Code Llama 70B, a large language model (LLM), specializes in code-related tasks, offering enhanced code generation and natural language understanding of code.

Code Llama 70B is trained on 1TB tokens of code and code-related data, setting a new standard in the field. It’s designed to generate code from text prompts, supporting popular programming languages like Python, C++, and Java, and is especially proficient in Python, owing to additional fine-tuning on 100B tokens of Python code.

This model offers three versions: a foundational code model, a Python-specific model, and an Instruct model fine-tuned for natural language instructions. The Instruct variant shows improved performance in generating contextually accurate and safe code outputs, making it particularly useful for code generation and debugging tasks.

Code Llama models handle up to 100,000 tokens of context, improving code generation for larger projects. They also exhibit versatility in serving different latency requirements, with the smaller 7B and 13B models offering faster, real-time code completion

other announcement:

OpenAI released new embedding models (text-embedding-3-small/large), updated GPT-4/3.5 Turbo, reduced prices, and enhanced API management tools. text-embedding-3-small/large showed improved MIRACL (44.0%/54.9%) and MTEB (62.3%/64.6%) scores, with scalable dimensions and cost-efficiency.

Google released Bard Gemini Pro, matching GPT-4 in human evaluations. It ranked second on Chatbot Arena, surpassing GPT-4 03/14 and 06/13, but trailing GPT-4 Turbo. The model’s performance is notable, though based on 3,000 user ratings, indicating potential for shifts with more data

Adept introduced Fuyu-Heavy, the third-most-capable multimodal model, excelling in UI understanding and outperforming Gemini Pro in the MMMU benchmark. Despite its smaller size, it matches text-only benchmarks of larger models and integrates text-image processing, indicating high efficiency and potential for diverse applications.

Google’s Lumiere, a space-time diffusion model, generates realistic videos using STUNet architecture for coherent motion, outperforming existing T2V models by synthesizing 80-frame, 5-second videos at 16 fps, trained on 30 million videos, and addressing temporal consistency challenges.

Hugging Face and Google Cloud partner to integrate open AI models and cloud services, enabling AI development through Vertex AI, GKE, and enhanced hardware (TPU v5e, A3 VMs, C3 VMs) for training, tuning, and deploying AI applications.