LLM News and Articles Weekly Digest — April 30, 2024
Latest News
Apple Releases OpenELM (HF, Github, Paper)
OpenELM, a suite of eight open-source LLMs, has been introduced by Apple. These models are optimized for single-device operation and cater to text generation tasks, with parameter capacities varying from 270 million to 3 billion.
Apple And OpenAI Are Reportedly In Talks For iOS 18 Integration
Apple is currently exploring partnerships with various AI providers, including OpenAI and Google, to integrate generative AI technologies into the upcoming iOS 18.
Microsoft launches Phi-3, its smallest AI model yet. (X, HF)
Microsoft has revealed its latest language model, Phi-3 Mini, boasting 3.8 billion parameters. Additionally, they’ve announced upcoming versions including Phi-3 Small and Phi-3 Medium, with 7 billion and 14 billion parameters respectively. The training methodology for Phi-3 Mini imitates the progressive learning phases of children, employing a curriculum that spans from elementary to advanced structures and ideas.
Elon Musk’s AI venture, xAI, is poised to finalize a $6 billion investment, valuing the company at $18 billion pre-money. Backed by investors like Sequoia Capital and Future Ventures, this funding round reflects substantial investor trust, partly attributed to Musk’s track record and networks established through his ventures like SpaceX and Tesla.
Cohere open sourced their chat interface : Cohere Toolkit (X)
Toolkit is a collection of prebuilt components enabling users to quickly build and deploy RAG applications.
Snowflake releases Arctic, an open LLM for Enterprise AI
Snowflake AI Research has introduced Arctic, an economical enterprise AI LLM showcasing a Dense-MoE Hybrid transformer architecture with an impressive 480 billion parameters. Trained at a cost of under $2 million, Arctic demonstrates proficiency in tasks such as SQL generation and coding. The model is completely open-source under the Apache 2.0 license, offering unrestricted access to both model weights and code.
Play.ai (previously play.ht) releases conversational Voice AI platform (X)
Powered by the new Large Dialogue Model (LDM), Play LDM 1.0 makes the speech generation contextual; handles turn-taking, interruption, voice energy and emotion modulation for a natural, fluid, human conversations in real-time.
Memory is now available to all ChatGPT Plus users. Using Memory is easy: just start a new chat and tell ChatGPT anything you’d like it to remember.
OpenAI slapped with GDPR complaint: How do you correct your work?
Articles
Joe Spisak talk about Llama3 on Stage at WandB Fully connected (Full Talk, TLDR)
SigLIP and Llama3 based multimodal model
The Bunny series of multimodal models are a powerful set of open models that perform well for their size on the MMMU benchmark. This is the team’s first open release of models based on Llama3 8B.
Some Technical Notes About Llama 3
Meta AI’s Llama 3 stands out with its advanced capabilities, boasting a 128K-token tokenizer and grouped query attention. Trained on an extensive 15 trillion token multilingual corpus, it incorporates Rotary Positional Encoding and Key-Value caching to enhance inference efficiency significantly.
Short Story Review of Large Language Models: A Survey
Unlocking Llama 3: Your Ultimate Guide to Mastering Llama 3!
Papers and Repositories
OpenLit
OpenLIT is an OpenTelemetry-native GenAI and LLM Application Observability tool designed to make the integration process of observability into GenAI projects possible with just a single line of code.
Thank you for reading !
The blog is originally published in https://shresthakamal.com.np/blog/2024/newsletter-edition-4/
If you have any feedbacks or suggestions , please do comment.
