Adrienne Jan’s Post

Last two weeks have been amazing for the open-source generative AI community. 1. AI at Meta released a "GPT-4 level" open-source LLM which allows the community to finetune and/or own their models and do all sort of RAG/LoRA wizardry on top of that. Try it here https://lnkd.in/gu_iuZ_q 2. BlackForestLabs has released the largest (12B parameters) and most advanced open-source model for text-to-image generation that excels in realism and human aesthetics. Bonus points, it can generate near perfect humans with fingers as well as text on images. Try it here: https://lnkd.in/gSspYTY4 🚀 Why this is huge for the AI landscape: 1. This reduces the barriers to entry in AI innovation for startups 2. There will be lower dependency on expensive proprietary models and more affordable GenAI solutions for businesses of all sizes 3. Smaller companies can now own their RAG/LoRA/finetuned pipeline if they build on top of open source 4. Slow erosion of the foundational model moat (like OpenAI) This pic summarizes how things went for open source lately. 😁 #GenAI #LLM #Diffusion #Transformers

I'm excited to host the AI4D: ByteDance Open Source Meetup on Thursday, October 3rd, from 5:30 PM to 8:30 PM at 1199 Coleman Ave, San Jose, CA 95110. Join us for a gathering of AI professionals to discuss challenges and solutions in AI implementation. Topic: AI/ML Inference Our keynote speaker is Dr. Alex Wei Chen, Founder and CEO of NEXA AI and creator of Octopus v2, an AI model that's 4x faster than GPT-4 and 70x more energy-efficient. He'll introduce Nexa SDK, an on-device AI toolkit supporting various models and GPU acceleration. Agenda: - 5:30 PM - 6:00 PM: Networking & Registration - 6:00 PM - 6:30 PM: Welcome and AMA with Dr. Alex Wei Chen & Tina Tsou - 6:30 PM - 7:30 PM: Nexa AI Presentation - 7:30 PM - 8:30 PM: Networking Dinner and refreshments will be provided. Space is limited, so RSVP is required. Please remember to bring your photo ID. If you're interested in presenting your company's product, contact Gary at [email protected]. Follow ByteDance Open Source on Twitter: [@ByteDanceOSS](https://lnkd.in/g5PDdYyY) Looking forward to insightful discussions and networking with leading minds in the AI industry! #AI #MachineLearning #OpenSource #Networking #ByteDance ByteDanceOSS

Navigating the latest breakthroughs in AI technology often feels like a leap into the future. In that spirit, let's explore three recent highlights: Microsoft is actively blurring the lines between human and artificial creativity with the integration of GPT technology into their cloud services. This bold move is not just reshaping the way we think about AI but is also revolutionizing how we approach coding and development tasks. https://lnkd.in/dstbvhbv Another fascinating development is the debut of 'gpt2-chatbot', a new enigma entering the already competitive landscape of AI models. Its sudden emergence and sophisticated abilities have sparked intense discussions and curiosity regarding its origins and potential. https://lnkd.in/d6nX68BZ Additionally, GitHub is introducing a new Copilot Workspace, leveraging AI to transform the coding experience for developers. It's a significant step forward for those looking to enhance their workflow with smarter programming tools and support. https://lnkd.in/dGuP3MwG The constant evolution of AI continues to amaze, as it delicately balances between disruption and innovation. With each development, the boundaries of what's possible expand, offering a glimpse into the dynamism of a future powered by artificial intelligence. What else lies ahead? Only time and technology will tell.

New Article Alert: Open-Source AI 🌟 With the tech community still buzzing from the recent legal clash between Elon Musk and OpenAI, it’s more crucial than ever to understand what is at stake. So in my latest article, I delve deep into the heart of open-source AI, exploring how the history of open-source and AI intersect, the current ecosystem, and the different paths we could see in the future. Why is open-source AI not just a trend but a necessity? How to monetize open-source AI models and tools? Who are the key players in this ecosystem? Are open-source AI models really open source and what’s the interest of the big techs to publish them "freely"? 💡 These are some of the questions I try to address. So whether you're an investor, developer, entrepreneur, or tech enthusiast, dive into the full article and don’t hesitate to contact me if you want to discuss about it! 👉 https://lnkd.in/enuyJ5Uz And a special thanks to Olivier & Maxime for the help in writing it 😁

Next Wave of LLM 2025: What Founders and Engineers Need to Know 🚀 By a chance, race for dominance in AI innovation is tighter than ever. With Meta's Llama 3.1 scaling unprecedented heights and OpenAI’s O1 redefining reasoning, the landscape is brimming with groundbreaking potential. Here's how you can stay ahead: 📄 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗦𝘁𝗮𝘁𝗲 𝗼𝗳 𝗔𝗜 𝟮𝟬𝟮𝟰 𝗥𝗲𝗽𝗼𝗿𝘁 𝟭️⃣ 𝗧𝗵𝗲 𝗦𝗰𝗮𝗹𝗶𝗻𝗴 𝗥𝗲𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻 OpenAI's O1 model shows the power of inference scaling, achieving unparalleled performance in mathematical and reasoning-heavy tasks. While it's 3–4x more costly than its predecessors, the efficiency gains are a game-changer for enterprise applications. 𝟮️⃣ 𝗖𝗹𝗼𝘀𝗶𝗻𝗴 𝘁𝗵𝗲 𝗚𝗮𝗽 𝘄𝗶𝘁𝗵 𝗢𝗽𝗲𝗻 𝗠𝗼𝗱𝗲𝗹𝘀 Meta's Llama 3.1 and 3.2 prove that open-source isn’t just catching up—it’s leading in accessibility and innovation. With over 440M downloads, the community's commitment to democratizing AI is paying off. 𝟯️⃣ 𝗢𝗻-𝗗𝗲𝘃𝗶𝗰𝗲 𝗘𝘅𝗰𝗲𝗹𝗹𝗲𝗻𝗰𝗲 Microsoft’s Phi-3.5-mini and Apple’s MobileCLIP models bring cutting-edge AI directly to your smartphone, enabling high-speed, low-latency applications without compromising performance. 𝟰️⃣ 𝗤𝘂𝗮𝗻𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆 Innovations like Microsoft’s BitNet and Meta’s MINITRON pave the way for efficient, lightweight models that save memory and energy, making large-scale deployments more feasible than ever. 🔑 𝗪𝗵𝗮𝘁 𝗧𝗵𝗶𝘀 𝗠𝗲𝗮𝗻𝘀 𝗳𝗼𝗿 𝗙𝗼𝘂𝗻𝗱𝗲𝗿𝘀 𝗮𝗻𝗱 𝗟𝗟𝗠 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝘀 𝗦𝘁𝗮𝗿𝘁 𝗦𝗺𝗮𝗹𝗹, 𝗦𝗰𝗮𝗹𝗲 𝗦𝗺𝗮𝗿𝘁: Leverage smaller distilled models like SmolLM or Llama-based miniatures for early-stage applications. They're cost-effective and scalable. 𝗜𝗻𝗻𝗼𝘃𝗮𝘁𝗲 𝗕𝗲𝘆𝗼𝗻𝗱 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲: Explore multimodal integrations—combine text, video, and biological insights to unlock unique business opportunities. 𝗙𝗼𝗰𝘂𝘀 𝗼𝗻 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆: Embrace quantization and pruning methods to reduce costs and environmental impact while maintaining high accuracy. 🛠 𝗔𝗱𝗼𝗽𝘁 𝗛𝘆𝗯𝗿𝗶𝗱 𝗠𝗼𝗱𝗲𝗹𝘀: Use combinations like Falcon’s Mamba-Transformer hybrids to balance speed and accuracy for diverse tasks. 📲 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗲 𝗳𝗼𝗿 𝗢𝗻-𝗗𝗲𝘃𝗶𝗰𝗲 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀: Think beyond cloud. Build LLMs that run seamlessly on devices, reducing latency and increasing accessibility. 🌍 𝗦𝘁𝗮𝘆 𝗢𝗽𝗲𝗻-𝗦𝗼𝘂𝗿𝗰𝗲 𝗙𝗿𝗶𝗲𝗻𝗱𝗹𝘆: Contribute to platforms like Hugging Face to access community-driven innovation while showcasing your work. Want to ride this AI wave? Let’s discuss how these strategies can transform your next project. Connect today and build the future of intelligent innovation! #LLMInnovation #AIRevolution #FoundersTech #StateofAI

Gemma Unleashed: Google's Leap into Open, Responsible AI Development Google recently unveiled Gemma, a groundbreaking series of lightweight, state-of-the-art open models, marking a significant stride in responsible AI development. These models are the culmination of extensive research and technology that also brought forth the Gemini models, with Gemma being spearheaded by Google DeepMind and various teams across Google. Reflecting its Latin meaning, "precious stone," Gemma signifies the invaluable contribution these models aim to provide to the developer and research communities. To complement the release of Gemma models, Google has also launched a Responsible Generative AI Toolkit. This toolkit is designed to support innovation among developers, encourage collaboration, and ensure the Gemma models are utilized responsibly. Gemma models come in two variants, Gemma 2B and Gemma 7B, both available in pre-trained and instruction-tuned formats, offering unprecedented accessibility to developers and researchers worldwide. These models have been optimized for a broad spectrum of AI hardware platforms, including NVIDIA GPUs and Google Cloud TPUs, and are compatible with major frameworks like JAX, PyTorch, and TensorFlow. This ensures that Gemma can achieve industry-leading performance, even on a developer's laptop or desktop computer. Furthermore, Google's terms of use for these models permit responsible commercial usage and distribution, making cutting-edge AI technology available to organizations of all sizes. In addition to providing state-of-the-art performance, Gemma models are designed with Google's AI Principles in mind, emphasizing safety and reliability. Automated techniques were employed to filter out sensitive data from training sets, and extensive fine-tuning and reinforcement learning from human feedback (RLHF) were used to align the models with responsible behaviors. Google has also made efforts to optimize Gemma for various frameworks and tools, ensuring seamless integration and broad accessibility. To support the development and research community, Google offers free access to Gemma through Kaggle, Colab notebooks, and Google Cloud credits, empowering researchers and developers to innovate and create with Gemma. The release of Gemma by Google, following in the footsteps of companies like Meta, highlights a commitment to advancing AI technology in an open, collaborative, and responsible manner. Through Gemma, Google aims to set new standards in AI development, fostering a safe and innovative ecosystem that leverages the full potential of AI to benefit society at large. More info from source: https://lnkd.in/gk7xzDxz

Should AI be open source? Elon Musk re-enters the chat Elon Musk, who had criticized (and even sued) OpenAI for not sharing its source code with the public, is now planning to make his own artificial intelligence platform, Grok, open-source by the end of this week. Open Sesame vs Opening Pandoras Box: The ‘open vs closed’ debate is causing a sharp divide in Silicon Valley. A recent argument between venture capital heavyweights Vinod Khosla and Marc Andreessen captures the essence of the rift: - Khosla argued that the source code for the world’s most powerful AI models should be protected at all costs. “Would you open source the Manhattan Project?” he asked in one recent post. - Andreessen, on the other hand, believes that everyone should have the opportunity to experiment and adapt to AI systems. This will ensure that a single company does not hold all the power. Prevention & Profits. Some major AI companies, like OpenAI, keep their source code secret, citing safety concerns and the potential dangers of unleashing their code. But there’s also undeniably a strong profit motive for keeping your secret sauce a secret. Other AI models, however, like Meta’s Llama 2 and Mistral AI’s model, are open source and allow anyone to tweak and modify their source code. Grok will soon join those ranks. The discussion around the open versus closed AI model is crucial. It will determine whether AI is a technology available to everyone or a tool in the hands of a few powerful companies. #ArtificialIntelligence #OpenSourceAI #TechDebate #AIethics #SiliconValley #AIAccessibility #FutureOfAI Source: Getty Images X Owner Elon Musk.

Gemini AI, developed by Google, is one of the most advanced AI ecosystems, known for its multimodal capabilities. It is designed to understand and generate content across multiple formats—text, images, audio, video, and even code—making it highly versatile. Launched as part of Google DeepMind’s efforts, Gemini models are integrated into various Google products, enhancing tasks like content generation, coding, and complex problem-solving. The Gemini AI ecosystem includes several versions, such as Gemini 1.0 Ultra, which excels at multimodal reasoning, and Gemini 1.5, featuring a highly efficient Mixture-of-Experts (MoE)architecture. This architecture improves performance by selectively activating only the most relevant neural pathways depending on the task. One of the standout features of Gemini 1.5 is its long-context understanding, enabling it to handle up to 1 million tokens, which allows for better understanding and generation of extended and complex inputs. Gemini's practical applications span a wide range, from assisting in **code generation** for competitive programming to enabling advanced reasoning in scientific, mathematical, and technical fields. It is optimized to work efficiently on different platforms, from large data centers to mobile devices. For developers and businesses, Gemini offers a powerful tool for innovation, with APIs accessible through Google AI Studio and Google Cloud's Vertex AI【13†source】【14†source】【15†source】. Additionally, Google's focus on ethical AI development is integrated into the Gemini project, prioritizing safety, bias reduction, and responsible deployment to mitigate risks【15†source】. #snsinstitutions #snsdesignthinkers #designthinking

The AI race is continuing to heat up. Google, which has been trailing leading AI labs such as OpenAI, has been firing on all cylinders and has released some incredible models over the last two weeks. Among these models is Gemma, Google’s new state-of-the-art open-source model. Gemma is a lightweight family of AI models that shares architectural and infrastructural components with Google’s larger, cutting-edge Gemini models. Gemma was optimized to be a small model that can be run with minimal compute resources, thereby lowering the costs of fine-tuning (specialization) and inference (using models to generate outputs). Gemma’s efficiency allows it to be run directly on laptops and computers, without the need for cloud computing or data centers, which makes it accessible and affordable for organizations, researchers, and individuals. Despite this small size, Gemma achieves best-in-class performance among models of similar size and even outperforms significantly larger models on a variety of benchmarks that test general knowledge, reasoning, math, and coding skills. What does it mean to be open-sourced? It signifies free access, the ability to run the AI model on your servers which enhances data security, and the freedom to edit and customize these models at will. In comparison, closed source models such as GPT-4 have very limited customizability and access that is controlled by OpenAI, which means that all data flows through their cloud services, reducing control over data security and privacy. By allowing developers and researchers to access, experiment with, and alter the source code, open-source models are a primary driver of democratization, innovation, and competition. Making increasingly powerful LLMs accessible allows for experimentation and innovation to happen anywhere in the world, regardless of resource constraints. This also enhances the ability of new competitors to enter markets that previously relied on closed-source models due to performance requirements. We have already observed the profound effects that open-source models can have both on scientific progress and business applications. Meta, the previous leader in open-source LLMs, released two open-source models in 2023. Some researchers called these releases the most important developments in AI because they supercharged the broader research community with access to powerful systems that can be used to discover new science and techniques. Indeed, many breakthrough research papers in 2023 leveraged Meta’s open-source LLaMA models. As the AI race continues to accelerate, the continuation of a strong open-source ecosystem is vital in fuelling innovation and ensuring the broad availability of powerful AI tools. Open-source models benefit everyone, from individuals and researchers to companies and governments. And the better the open-source models get, the faster innovation will happen, creating a positive feedback loop that is a net benefit for all.

🚀 Excited to announce the release of Meta’s Llama 3.1, our most advanced open-source AI models yet! 🚀 Llama 3.1 is pushing the boundaries of AI capabilities with unmatched flexibility, control, and performance. This latest release is set to transform the field of AI with some groundbreaking features: Key Highlights: - Expanded context length to 128K, enabling advanced use cases like long-form text summarization and multilingual conversational agents. - Support for eight languages, enhancing global accessibility and usability. - Introducing Llama 3.1 405B, the first frontier-level open-source AI model. Exciting Features: - New models, including Llama 3 8B, Phi-3 Mini, Granite, and Gemma models from Meta, Google, and IBM. - Enhanced AI ecosystem with over 25 partners, including AWS, NVIDIA, Databricks, Groq, Dell, Azure, and Google Cloud, offering services from day one. - Improved safety and security with new tools like Llama Guard 3 and Prompt Guard. Why Llama 3.1 Stands Out: - Competitive performance with leading foundation models across various tasks. - Open-source advantage allows developers to fully customize models for their needs and applications. - Innovative applications from synthetic data generation to model distillation. Join the revolution and explore the potential of Llama 3.1: - Real-time and batch inference - Supervised fine-tuning - Continual pre-training - Retrieval-Augmented Generation (RAG) Download the models, read the paper, and start building the future of AI today. 🔗 [Read the Llama 3.1 Paper](https://lnkd.in/dqAnWPPw) 🔗 [Visit the Llama GitHub Repo](https://lnkd.in/d67FDP-a) 🔗 [Download Llama 3.1 on Hugging Face](https://huggingface.co) #AI #OpenSource #MetaAI #Llama3 #MachineLearning #ArtificialIntelligence