. . AI: Microsoft's BitNet b1.58: A Paradigm Shift in AI Computation? (2.28.24)

Friends,

Today's edition highlights an innovative method for training and implementing models using new architectures to improve efficiency. It also highlights how early we are on this journey as consumer GPUs use hundreds of watts of power while the human brain operates on a mere 20 watts. While we have a significant journey ahead to match this efficiency, the research is pushing ahead with impressive improvements. In other news, we potentially get a sneak peek into the future of content and work, with Klarna and Alibaba are both using and advancing the capabilities of AI. Happy reading!

Microsoft Launches New Model Architecture for Efficient, Scalable LLMs

First impacted: AI researchers, Software developers
Time to impact: Medium

Microsoft's research team has introduced BitNet b1.58, a new variant of a Large Language Model (LLM) that utilizes ternary weights (-1, 0, 1) for improved cost-effectiveness in latency, memory, throughput, and energy consumption. This model matches the performance of traditional high-precision Transformer LLMs but requires specialized training. The change significantly reduces the need for multiplication in matrix operations and potentially a shift towards AI computation optimized for 1-bit LLMs. [Paper page - The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits] Explore more of our coverage of: Microsoft Research, BitNet b1.58, AI Computation. Share this story by email

Klarna's AI Assistant Does the Work of 700 Employees and Saves Millions

First impacted: Klarna employees, Customer service representatives
Time to impact: Short

Financial services firm Klarna, in collaboration with OpenAI, has developed an AI assistant that it says can perform the work of 700 human workers, potentially boosting profits by an estimated $40 million in 2024. The assistant, which operates 24/7 in 35 languages across 23 markets, has reportedly handled two-thirds of customer service chats, equivalent to around 2.3 million conversations, within a month of its global launch, and has received customer satisfaction ratings on par with human agents. [Klarna says its AI assistant does the work of 700 people after it laid off 700 people] Explore more of our coverage of: AI in Customer Service, Klarna and OpenAI, AI Efficiency Boost. Share this story by email

BigCode Launches StarCoder2 Models, Open-Source and Trained on Large Dataset

First impacted: Software Developers, Data Scientists
Time to impact: Short

BigCode, in collaboration with Nvidia, Hugging Face and ServiceNow, is launching StarCoder2, a range of open LLMs trained on an extensive code dataset, The Stack v2. The initiative is making all models, datasets, and training code publicly accessible, with the largest model, StarCoder2-15B, trained on over 4 trillion tokens across more than 600 programming languages. [StarCoder2 and The Stack v2] Explore more of our coverage of: Open Source AI, Code Learning Models, Hugging Face Initiative. Share this story by email

Alibaba's EMO Framework Boosts Realism in Video Generation

First impacted: Content Creators, Video Editors
Time to impact: Short

Alibaba's EMO framework, as detailed in a research paper on arXiv, proposes a direct audio-to-video synthesis approach for talking head video generation, eliminating the need for 3D models or facial landmarks. The company says this method results in smooth frame transitions, consistent identity preservation, and lifelike animations while achieving a threefold performance improvement in capturing human expressions and individual facial styles compared to previous techniques. Check out the demos in the link below, they are super impressive! [Paper page - EMO: Emote Portrait Alive - Generating Expressive Portrait Videos with Audio2Video Diffusion Model under Weak Conditions] Explore more of our coverage of: Alibaba Research, Audio-to-Video Synthesis, Video Generation Techniques. Share this story by email

Ideogram Launches 1.0 Text-to-Image Model

First impacted: Graphic Designers, Digital Artists
Time to impact: Short

Ideogram has launched its advanced text-to-image model, Ideogram 1.0, and is now accessible on the company's website. The company says that the model, inclusive of a new feature labeled 'Magic Prompt', enhances text rendering accuracy and claims to have 'unprecedented photorealism'. Our initial tests of the system left us pretty impressed. [Ideogram.ai] Explore more of our coverage of: Text-to-Image Technology, AI Model Innovation, Digital Art Creation. Share this story by email

That’s it! More AI news tomorrow.