Your Digital Twin, Proxy

• Your personal digital clone for low value tasks

• Gets smarter as you give it commands to learn

• The first truly general AI Agent

Meet your Proxy here

Today:

• Claude's New Tool for Analysis

• Meta Quantized Llama Models

• Cerebras Inference Performance Boost

• OpenAI's Continuous-time Consistency Models

• DeepMind’s New AI Music Tools

• Balaji Exposed OpenAI's Copyright Use

• Perplexity Stands Firm Against Lawsuit

Claude's New Tool for Analysis

Claude.ai has introduced a new analysis tool that allows users to run JavaScript code for data processing and real-time insights. Available in feature preview, this tool turns Claude into a functional data analyst, able to analyze, clean, and visualize data from CSV files. It supports a wide range of tasks, from improving marketing conversion rates to generating financial dashboards, offering precise and reproducible results. 

The analysis tool empowers teams across various fields, enhancing decision-making through data-backed insights. Users can activate the feature by logging into Claude.ai and managing their feature previews.

Meta Quantized Llama Models

Meta introduced its first quantized Llama models, optimized for mobile devices. These models are smaller, faster, and use less memory, making them ideal for on-device AI. By employing advanced techniques like Quantization-Aware Training and SpinQuant, Meta reduced model sizes by 56% and improved speed 2-4 times without compromising accuracy. 

These models are built for Qualcomm and MediaTek CPUs and are open-sourced for developers to create efficient, privacy-focused apps. Meta aims to make AI accessible for everyone, regardless of resources, and continues collaborating with partners to improve performance further.

Cerebras Inference Performance Boost

Cerebras has announced a 3x boost in performance for its Inference engine, allowing Llama 3.1-70B to process 2,100 tokens per second. This makes it 16x faster than the best GPU solution. Optimizations in software and hardware have increased speed and efficiency, making AI models more responsive for real-time applications, from video generation to pharmaceutical research. 

Companies like GSK and LiveKit are leveraging this power for breakthroughs in drug discovery and voice AI. The improved performance enables more complex reasoning and faster responses, transforming AI development and deployment for next-gen applications.

OpenAI's Continuous-time Consistency Models

OpenAI has introduced a new approach called continuous-time consistency models (sCM), designed to simplify, stabilize, and scale generative models. sCM achieves sample quality comparable to leading diffusion models but with significantly fewer steps—just two—resulting in a 50x speed improvement. 

These models generate high-quality samples in 0.11 seconds, ideal for real-time applications in image, audio, and video generation. Trained on large-scale datasets, sCM leverages distillation from pre-trained diffusion models, closing the gap in quality while requiring much less computational power. Despite its advancements, sCM still relies on diffusion models for initialization.

DeepMind’s New AI Music Tools

Google DeepMind has introduced new AI-powered tools aimed at revolutionizing music creation. The latest updates include MusicFX DJ, a real-time music generator allowing users to create unique music with text prompts and intuitive controls. Music AI Sandbox, an experimental toolkit, helps musicians enhance their creative workflows, while YouTube's Dream Track generates instrumental soundtracks for creators. 

These tools, developed in collaboration with musicians like Jacob Collier, offer high-quality, real-time audio generation, making music creation more accessible to everyone, regardless of skill level. The technologies enable users to create, share, and collaborate in new, dynamic ways.

Balaji Exposed OpenAI's Copyright Use

Suchir Balaji, a former researcher at OpenAI, has publicly criticized the company's use of copyrighted data to train AI models like ChatGPT. After spending nearly four years at OpenAI, Balaji left in August, expressing concerns about the company's practices and the broader harm he believes AI is causing to the internet ecosystem. 

He argues that AI systems like GPT-4 are unlawfully using copyrighted materials to generate content that competes with original works. OpenAI, however, defends its approach under the "fair use" doctrine. Balaji calls for regulation to address the legal and ethical challenges posed by AI technologies.

Perplexity Stands Firm Against Lawsuit

Perplexity, a generative AI tool, faces a lawsuit from the Wall Street Journal and the New York Post. The company defends itself, emphasizing its commitment to transparency and innovation in delivering knowledge, citing that the lawsuit is shortsighted. Perplexity argues that media companies resist technological progress, preferring outdated control over publicly reported facts. 

Despite the lawsuit, Perplexity highlights its collaboration with other media, such as TIME and Fortune, through revenue-sharing agreements. The company remains open to working with News Corp. but vows to defend itself while continuing to offer transformative learning tools.

🧠RESEARCH

microsoft/OmniParser

OmniParser is a screen parsing tool designed by Microsoft to convert UI screenshots into structured data, improving large language model (LLM) agents for user interfaces. It uses a finetuned YOLOv8 model for detecting clickable areas and a BLIP-2 model for generating icon descriptions, enhancing interaction and functionality understanding. Users should apply critical reasoning when using the tool, and it's not recommended for sensitive content analysis or workplace scenarios.

deepseek-ai/DeepSeek-V2.5

DeepSeek-V2.5 is an advanced language model combining general and coding abilities. It improves on DeepSeek-V2 through enhanced instruction following and human preference alignment. The model excels in tasks like coding, conversational AI, and text generation. It supports commercial use and is highly optimized for efficient inference, especially using large-scale GPUs.

CohereForAI/aya-expanse-8b

Aya Expanse 8B is a multilingual language model developed by Cohere For AI, supporting 23 languages, including Arabic, Chinese, French, and Spanish. It features an 8-billion parameter transformer architecture optimized for text generation and various multilingual tasks. The model is open for research under a CC-BY-NC license, and fine-tuning is available for specific use cases like writing assistants and question-answering systems.

MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models

MIA-DPO (Multi-Image Augmented Direct Preference Optimization), a method for training Large Vision-Language Models (LVLMs) to better predict human preferences across multi-image inputs. By extending single-image data with grid or pic-in-pic formats, MIA-DPO reduces the need for costly data annotation and enhances model performance. It uses attention values to filter out incorrect responses without human involvement.

WorldSimBench: Towards Video Generation Models as World Simulators

WorldSimBench, a dual evaluation framework designed to assess video generation models, referred to as World Simulators, from both visual and embodied perspectives. It includes Explicit Perceptual Evaluation, using human feedback to judge video quality, and Implicit Manipulative Evaluation, assessing how well generated videos translate into correct actions in dynamic environments. The framework covers tasks like autonomous driving and robot manipulation.

🛠️TOP TOOLS

Genmo - Open-source AI video generation 

Video Design by ElevenLabs - Generate a unique voice from a text prompt alone.

Watermark PDF - Add custom text or image watermarks to your PDF documents securely and easily.

Short Generator - Create Viral Videos in minutes!

Duonut - Conversational AI Survey with Real-time Follow ups and Response Summary for your users

📲SOCIAL MEDIA

🗞️MORE NEWS

• Google Photos is enhancing transparency by showing when AI tools like Magic Editor and Magic Eraser were used on photos. Starting next week, AI edits will be visible in the app's metadata, improving clarity around photo modifications.

• Google has launched a free SAIF Risk Assessment tool to help organizations assess AI security risks. The tool generates custom recommendations for mitigating risks like data tampering, making AI system security simpler and more efficient.

• Siemens and Microsoft have expanded their collaboration by enhancing the Siemens Industrial Copilot, integrating advanced AI with Microsoft Azure. This tool helps over 120,000 engineers streamline processes, reduce downtime, and address labor shortages, boosting industrial automation efficiency globally.

• Cloudflare has introduced Workflows, a durable execution engine designed for building reliable, scalable, multi-step applications. Now in open beta, developers using Cloudflare Workers can create workflows with minimal setup. Workflows handles retries, state persistence, and task orchestration, even in the event of failures or network issues. 

• OpenAI's Noam Brown emphasized the shift towards "system two thinking" in AI, where slower, more deliberate reasoning mimics human problem-solving. Highlighting OpenAI's new o1 model, Brown demonstrated how just 20 seconds of strategic AI thinking can match the performance gains of scaling models by 100,000x. 

• The ongoing debate between open-source and closed-source AI models remains unresolved, with closed models like those from OpenAI and Anthropic maintaining a slight edge in quality. Nvidia's open-source Nemotron, built on Meta's Llama, has gained attention recently, sparking questions about whether open-source models are catching up.