AI Innovations: Unveiling the Latest Breakthroughs
Welcome to the June 2024 Edition of Bayes Bulletin!
Uncover the industry's latest breakthroughs, from innovative models to real-world applications. Stay informed and inspired as we navigate through the dynamic landscape of AI that is shaping the future of technology.?
Subscribe for a journey into the future of intelligence!
Latest Models:?
1. NuExtract: A Foundation Model for Structured Extraction
NuExtract is a game-changer for anyone working with text data. This innovative large language model excels at extracting structured information from documents and reports. Imagine a tool that can automatically scan through text, pinpoint key details, and organize them into a clean, JSON format. NuExtract does exactly that!? Even better, it's lightweight and requires minimal computing power compared to other large language models. NuExtract can be used right out of the box for various tasks, or fine-tuned to tackle specific challenges. This powerful tool has the potential to revolutionize the way we handle textual data.?
2. nach0: Multimodal Natural and Chemical Languages Foundation Model
Nach0 is a game-changer for scientific research. Unlike most language models, nach0 isn't just good with words. It's trained on a special mix of scientific papers, patents, and even the codes for molecules themselves. This lets nach0 understand both natural language and the language of chemistry. As a result, nach0 can handle a variety of scientific tasks: answering tricky questions about biology and medicine, identifying key players in research papers, creating new molecules with specific functions, and even predicting how to build them! With this kind of power, nach0 becomes a powerful tool for researchers in drug discovery, material science, and any field where understanding both chemicals and words is key.
Latest Frameworks:
1. PowerInfer-2:
PowerInfer-2 is a highly optimized inference framework that enables fast inference of large language models up to 47B parameters directly on smartphones. A successor to the original PowerInfer, by using heterogeneous computing techniques to cluster operations for different hardware and intelligent pipelining to overlap loading and compute, PowerInfer-2 achieves up to 22x higher speed than previous state-of-the-art mobile frameworks at 11.68 tokens/sec. Coupled with new TurboSparse model architectures providing high predictable sparsity like TurboSparse-Mixtral-47B trained efficiently on just 150B tokens, PowerInfer-2 delivers massive memory savings up to 40% while still outperforming other frameworks' speed - ushering in a new era of large language model deployments on mobile devices.
2. WebLLM:
WebLLM brings the power of large language model inference directly into web browsers with hardware acceleration via WebGPU. This groundbreaking technology is fully compatible with the OpenAI API, allowing you to leverage open source models locally with functionalities like streaming, JSON-mode, and function-calling - all running entirely in the browser without any server support. With extensive model support including Llama, Phi, Gemma, Mistral, Qwen and more, WebLLM allows privacy-preserving AI assistants while enjoying GPU acceleration. Seamlessly integrate WebLLM into your web apps using NPM/CDN or build extensions, leveraging features like real-time streaming, web workers, and easy custom model integration.?
Github repo: https://github.com/mlc-ai/web-llm?
领英推荐
Latest Research papers:
1. From RAGs to rich parameters: Probing how language models utilize external knowledge over parametric information for factual queries
The paper "From RAGs to Rich Parameters" by Hitesh Wadhwa et al. investigates how Retrieval Augmented Generation (RAG) affects language models' (LMs) use of parametric memory versus external context for factual queries. Using techniques like Causal Mediation Analysis and Attention Contributions, the study reveals that LMs, specifically LLaMa-2 and Phi-2, favor external context provided by RAG over their internal knowledge. This "shortcut" behavior is evident as the models rely less on their parametric memory, particularly in Multi-Layer Perceptrons (MLPs), when RAG context is available. The study highlights the need to understand the interplay between internal knowledge and external context in LMs to improve factual accuracy and consistency.
2. Multi-step Knowledge Retrieval and Inference over Unstructured Data:
The paper "Multi-step Knowledge Retrieval and Inference over Unstructured Data" outlines the use of EC’s Cora, a neuro-symbolic AI system, to address challenges in multi-step inference tasks. These challenges, particularly evident in life sciences and macro-economic analysis, include controlling search processes, validating results, and dealing with hallucinated references. Cora combines knowledge extraction algorithms and formal reasoning to produce detailed, reliable answers, surpassing LLMs and RAG baselines. The AI platform utilizes a symbolic reasoning engine and a proprietary knowledge representation language, Cogent, to construct and analyze causal maps. Preliminary evaluations in medical research demonstrate Cora's superior accuracy, citation density, and relevance, highlighting its potential in delivering reliable AI-driven insights for complex queries.
?AI Startup news:
1)? Unitree Robotics introduced its next-generation G1 humanoid robot at ICRA 2024. The previous model, the H1, debuted in 2023 and gained attention through videos showcasing its balance and walking abilities. It was a hit at CES 2024, attracting crowds to the Unitree booth, and later showcased at NVIDIA GTC, where it navigated through busy aisles. Unitree has gained recognition as an early leader in the humanoid robot field, being the first to confidently demonstrate a tether-free robot in public, highlighting its stability and balance.
2) Scale AI, a provider of data-labeling services for machine learning, raised $1 billion in a Series F round from investors like Amazon and Meta. This funding, a mix of primary and secondary, reflects the surge in AI investments, following Amazon's $4 billion stake in Anthropic. Previously, Scale AI raised around $600 million, with a $325 million Series E in 2021 valuing it at $7 billion. Now, despite past layoffs, its valuation has nearly doubled to $13.8 billion. The round, led by Accel, also included investments from Cisco, Intel, AMD, and others, with existing investors like Nvidia and Y Combinator participating again.
Further reading- https://techcrunch.com/2024/05/21/data-labeling-startup-scale-ai-raises-1b-as-valuation-doubles-to-13-8b/
AI Conferences-
1) Databricks Data + AI Summit:
The Databricks Data + AI Summit 2024, held in San Francisco from June 10-13, showcased significant advancements in data management and AI. Key announcements included Databricks LakeFlow for streamlined data workflows, the open-sourcing of Unity Catalog for comprehensive data governance, and major updates to MLflow focused on generative AI and large language models. Enhanced Mosaic AI capabilities and the launch of Databricks AI/BI for integrated intelligent analytics were also highlighted. These developments emphasize Databricks' commitment to innovation and community engagement in the data and AI landscape.
2)?The AI Summit London 2024:
The AI Summit London 2024 featured prominent industry leaders such as Adam Davison, Head of Data Science at the Advertising Standards Authority UK, and Catriona Campbell, UK&I Chief Technology & Innovation Officer for EY. The agenda included keynotes on advancements in language models and AI at scale, with sessions on responsible AI practices led by experts like Alyssa Lefaivre from the Responsible AI Institute and Colin Jarvis from OpenAI. Practical AI applications were discussed by speakers like David Reed of DATAIQ, while the future of AI, including trends like quantum computing, was explored by Corey O'Meara, Chief Quantum Scientist at E.ON.
Stay tuned for more!