Linkt

A recent leak from the protest group Sora PR Puppets has revealed access to OpenAI's unreleased Sora video model on Hugging Face. This incident raises significant questions about the ethical implications of early access programs in AI development. The protest group claims that OpenAI recruited hundreds of artists for unpaid testing while maintaining strict control over generated content. This raises concerns about the potential exploitation of creative professionals in the tech industry. Key details from the leak include: - The temporary Hugging Face implementation generated 1080p clips faster than reported 10-minute render times. - Users observed OpenAI's watermark on generated clips, indicating the model's authenticity. - OpenAI is reportedly training a new version of Sora to address long render times, with rumored features like in-painting and image generation. This leak clearly highlights the tension between big AI companies and the creative community. But more specifically, OpenAI. While competitors like Runway?are rapidly advancing their AI video tools, Sora has remained largely behind closed doors. The capabilities demonstrated in the leak are impressive but Sora still is way behind their competitors. Could this be enough reason to hold off on the launch of this version of the Sora video model? Share your thoughts! ??: https://lnkd.in/eZDTG2gv?

A recent study demonstrates the power of model pruning, achieving 98.4% accuracy recovery on key benchmarks while reducing 50% of the parameters from Meta's Llama 3.1 8B model. This technique strategically removes unnecessary connections, resulting in a smaller and faster model. Key metrics are: - 2:4 sparsity pattern for effective parameter reduction - 30% higher throughput and 1.8x lower latency, reaching up to 5.0x with quantization - Compatibility with 4-bit quantization (GPTQ) and Sparse-Marlin kernels - Full recovery on fine-tuning tasks like GSM8K and Evol-CodeAlpaca Pruning is essential for optimizing AI models, especially when deploying on resource-constrained devices. By removing non-essential weights, developers can enhance inference speed and reduce memory usage without sacrificing accuracy. Have you thought about integrating model optimization techniques like pruning into your AI projects? Let us know! Chart by Neural Magic

Is Biotech Too Open? Boltz-1’s Release Raises Concerns About Security and Research Ethics The MIT Jameel Clinic has launched a new open-source model that accurately predicts biomolecular interactions: Boltz-1. This model achieves AlphaFold3-level accuracy in modeling complex biomolecular structures. As a matter of fact, Boltz-1 is the first fully commercially available open-source model of its kind. It matches the performance of Chai-1, a renowned closed-source replication of AlphaFold3 in the AI science field. For instance, Boltz-1 demonstrates a 65% LDDT-PLI score on CASP15, significantly outperforming Chai-1's 40%. By providing training and inference code, model weights, and data under the MIT license, Boltz-1 aims to democratize access to advanced biomolecular modeling. This initiative empowers researchers worldwide to innovate in drug design and structural biology. As the Boltz-1 team continues to enhance its capabilities, we look forward to seeing how this model will drive innovation in the life sciences. ?? Read the full article here: https://lnkd.in/gGeh_MKn

AI Revolution or Risk? Mistral’s Pixtral Large Could Reshape Marketing, E-Commerce, and Healthcare. Mistral AI has launched Pixtral Large, an advanced multimodal AI model designed for image generation and analysis. With 12 billion parameters and a 400 million parameter vision encoder, Pixtral Large excels in processing images at their native resolution. This model can handle a 128K token context window, allowing it to process up to 2,000 images simultaneously. It achieves a 52.5% score on the MMMU reasoning benchmark, outperforming many larger models in multimodal tasks like chart understanding and document question answering. From a technical standpoint, Pixtral's architecture enables it to deliver high-quality visuals from textual descriptions, making it ideal for industries such as marketing, e-commerce, and healthcare. Companies can leverage Mistral’s new multimodal model to automate content creation, enhance product visualization, and streamline data extraction from images. This release shows just how narrow the gap between open and closed AI models is getting.

Is this the end of traditional commerce? Alibaba Group’s Accio will change the connection between buyers and sellers. Alibaba has launched Accio, an AI-driven B2B search engine that transforms global trade by efficiently connecting companies or independent buyers and sellers. Utilizing advanced natural language processing (NLP), Accio boosts purchasing intentions by 40% in pilot tests. Built on Alibaba's Tongyi Qianwen model, it offers personalized search results and insights in five languages: English, German, French, Spanish, and Portuguese. Accio quickly understands complex queries and matches them with relevant products and suppliers, streamlining procurement. It integrates with Alibaba's ecosystem to provide comprehensive sourcing tools. By analyzing real-time data, Accio delivers current market trends and insights, enabling informed business decisions. As AI reshapes commerce, Accio will enhance operational efficiency, especially for startups and small businesses in the digital landscape.

Qwen from Alibaba Cloud has just launched an impressive open-source AI model that rivals OpenAI's GPT-4o in coding capabilities. The Qwen 2.5-Coder-32B model competes directly with Anthropic's Claude Sonnet 3.5, achieving remarkable results across multiple coding benchmarks. Early testers have noted its ability to perform tasks that Sonnet struggled with. Comments include, “It feels like Claude 3.5 Sonnet” and “I’m blown away by its performance in long context work.” Here are some key highlights from the Qwen 2.5 release: - It nearly matches Claude in overall coding capabilities, despite having just 32 billion parameters. - It outperforms Claude 3.5 Sonnet on HumanEval (92.7 vs. 92.1) and EvalPlus (86.3 vs. 85.9). - Qwen 2.5 excels across various coding benchmarks and fill-in-the-middle tasks. - It is available in both Base and Instruct versions, supporting over 40 languages with a context length of 128K tokens. - The model shows high performance in code repair benchmarks, scoring 73.7 on Aider. - Users can utilize long system prompts (over 16k) with examples and documentation for better results. As open-source solutions continue to evolve, tools like Qwen 2.5 will play a pivotal role in democratizing access to advanced AI capabilities.

Tired of Slow AI? Meet Claude 3.5 Haiku. Launched by Anthropic. Claude 3.5 Haiku is the latest version of its AI model. This new model is designed to enhance coding, tool use, and reasoning capabilities. The company has reported that Claude 3.5 Haiku is the fastest in the Claude family, offering significant improvements over previous versions. It excels in various applications, making it a valuable asset for developers and businesses alike. One standout feature that we have seen is its ability to deliver quick and accurate code completions. This can streamline development workflows, allowing teams to boost productivity effectively. The model also enhances interactive chatbots with improved conversational abilities and rapid response times. This makes it ideal for customer service, e-commerce, and educational platforms that need scalable engagement solutions. Additionally, Claude 3.5 Haiku is great at data extraction and labeling. It efficiently processes large volumes of unstructured data. This capability is crucial for sectors like finance, healthcare, and research. From an AI developer's perspective, the performance metrics are impressive. It surpasses previous versions on many intelligence benchmarks while maintaining a cost-effective pricing structure. Something really helpful for startups and small companies. Starting at $1 per million input tokens, it offers up to 90% savings with prompt caching. At Linkt.ai, we see the potential in today’s AI industry of Claude 3.5 Haiku to transform how startups and tech companies approach AI integration. Anthropic never disappoints us, so we hope that their AI model development continues and scales in today’s competitive industry. ?? Read the full article here: https://lnkd.in/dssprq-E

From China's AI industry development, we bring you a comparative table from Tencent's latest Hunyuan-Large MoE model. This model features an impressive 389 billion parameters and was trained on a staggering 1.5 trillion tokens of synthetic data and outperforms Meta's Llama 3.1 (405B) across various academic benchmarks. This comparative table showcases China's AI development and capabilities in language understanding and generation. Here are some key highlights from this new MoE model: - 236 billion parameters are utilized, with 21 billion activated during generation. - The model employs 160 experts, activating 6 experts during generation for optimized performance. - A comprehensive technical report details scaling experiments and performance metrics. - It has been trained on a total of 7 trillion tokens, with 1.5 trillion being synthetic. - The primary training languages are English and Chinese, enhancing its versatility. - The model is designed to fit on a single H100 Node (8x) using FP8 precision, making it accessible for deployment. - Tencent offers a custom license for commercial use below 100 monthly active users (MAU), though it prohibits usage by EU citizens and companies. Tencent's new model architecture allows for efficient resource utilization while delivering high performance. The ability to activate a smaller subset of parameters during inference means faster processing times without sacrificing accuracy. As the AI landscape continues to evolve, models like Hunyuan-Large, and Meta's Llama 3.1, will continue to play a crucial role in shaping the future of intelligent applications.

OpenAI has announced the partnership between Broadcom and TSMC to develop OpenAI’s first custom chip. The new chip is designed specifically to enhance the efficiency of OpenAI's models, which require substantial computational power. Something similar that we have seen from other companies like Intel AI, NVIDIA and AMD. By tailoring hardware to their unique needs, OpenAI aims to improve processing speed and energy efficiency. So, this development comes at a time when the demand for AI capabilities is skyrocketing. Industry analysts predict that the AI chip market will reach $100 billion by 2027. Companies are increasingly recognizing that specialized hardware can significantly boost performance and reduce operational costs. Moreover, OpenAI's collaboration with established players like Broadcom and TSMC highlights the importance of AI in today’s market. OpenAI’s venture into custom chip development represents a pivotal step forward in optimizing AI performance. At Linkt.ai, we hope to see more companies related to the AI industry like Meta or Google DeepMind create their own chips. ?? Read the full article here: https://lnkd.in/gYHKYmj9

Which AI Provider Leads in Speed? This week, we present an analysis of output speeds over the past two months for various AI providers offering Llama 3.1 Instruct 70B, including Cerebras Systems, Amazon Web Services (AWS), Lepton AI, Google Vertex, Microsoft Azure, Fireworks AI, Groq, and Perplexity. High output token speeds enhance AI performance by reducing response times and improving user experience. Faster speeds allow companies to handle more queries simultaneously, optimizing resource use and lowering costs. Prioritizing these speeds is essential for effectively scaling AI solutions. The chart from Artificial Analysis reveals that smaller, emerging providers consistently achieve higher output speeds. Notably, Cerebras leads with 2155.6 tokens per second, significantly outperforming competitors. Among established providers, Groq ranks highest at 250.7 tokens per second, while Azure trails behind with only 11.1 tokens per second. What is your go-to provider? Let us know!