Vaibhav Srivastav’s Post

Exploring Microsoft AI’s Latest Innovation: WizardLM-2 open source LLM models. I’m excited to share some fascinating developments from Microsoft AI with the introduction of their new large language model, WizardLM-2. There are many models coming out on weekly basis especially this cutting-edge model isn’t just an incremental improvement; it’s a substantial advancement in in LLM open source model , particularly in processing and understanding complex tasks. Highlights of WizardLM-2: 1. Model Variants: • WizardLM-2 8x22B: This model is a powerhouse, leading the charge in Microsoft’s internal evaluations and setting a new benchmark for handling intricate AI tasks. • WizardLM-2 70B: If you’re looking for robust reasoning abilities, this version stands out in its class, pushing the boundaries of what AI can deduce and infer. • WizardLM-2 7B: Remarkably fast and efficient, this model performs on par with others ten times its size, offering a practical solution for more streamlined applications. If you have T4-16 GB GPU , WizardLM-2 7B 8 bit quatized model is best and suitable LLM as of now 2. Advanced Training Techniques: • The use of AI-driven data pre-processing and a progressive learning structure not only improves efficiency but also enhances the model’s performance with significantly less data. •The innovative AI Align AI (AAA) framework allows leading models to teach and enhance each other, fostering a collaborative learning environment. It’s important to note these techniques for LLM fine tuning 3. Performance Metrics: • In terms of human preferences and standardized benchmarks like MT-Bench, WizardLM-2 competes neck and neck with top-tier proprietary models and surpasses other open-source offerings, confirming its superiority across different scales. Open-Source Contribution: • Excitingly, the model weights for WizardLM-2 8x22B and 7B are now accessible on Huggingface, which is a substantial move towards transparency and collaboration in the AI community. Looking Forward: • Microsoft AI plans to release a detailed paper that will delve into the methodologies and technological innovations that power WizardLM-2. Stay tuned #AI #MachineLearning #OpenSource #MicrosoftAI #LanguageModels #Innovation #wizardLM Reference - https://archive.is/36cd2 This post is written during my trip in Hannoi Vietnam , experiencing nice moment of Sunset in West lake

In the realm of AI, less is often more, but Microsoft takes this to a whole new level with their latest innovation: a 1-bit language model! Gone are the days of 16-bit extravagance. Enter BitNet b1.58, where every bit counts, literally. This marvel promises to keep performance on par with its 16-bit cousins while tightening the belt on data diets. Could this be the dawn of leaner, meaner AI models? #MachineLearning #MicrosoftAI #BitNet #llm #genai #aiandml For the full byte 😁 ,check out: Link to article https://lnkd.in/eUACMk4m

Yet another brilliant week for Open Source AI: 1. Phi 3 mini by Microsoft receives a massive upgrade: With the massive increase in scores for a structured generation. 2. ESPNet releases XEUS: A 577M parameter SoTA Speech Encoder trained on 4000+ languages. 3. Kyutai dropped Moshi: 7B parameter speech-to-speech model trained from scratch with an open demo (before Open AI) based on Candle. 4. InternLM pushed out LLM & VLM 2.5: 7B VLM beats GPT 4V and the LLM is the current leader for < 7B LLMs. 5. ChatGLM CodeGeeX4 9B: Yet another SoTA Code LLM from the team with up to 128K context. 6. Minference by Microsoft: 10x speedup and maintains accuracy for prefilling with 1M context. What a brilliant week, I can't wait for the next!

The march of the Small Language Models continues with Microsoft Phi-3 - Mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens and can run on an smart phone, large context - 128k - also The Composable Enterprise is a really achievable with these types of models. Model link here: https://lnkd.in/e_q6Hfkx quantized version here - model is 2.15Gb (mind blowing) https://lnkd.in/e3CcKpY4 #ai #artificialintelligence #llm #llms #transformers #slms #smalllanguagemodels #composablelanguagemodels

Small Language Models (SLMs), such as Microsoft's newest compact model, Phi-2, are significant advancements in the field of AI. They offer several advantages over their larger equivalents, such as efficiency, speed, customizability, and improved security features. Phi-2 with 2.7 billion parameters is about twice as big as its predecessor, Phi 1.5 which has 1.3billion parameters, and continues to perform on par or better than certain larger open-source Llama 2 models, including one with 13 billion parameters. Some of the key features of small language models, including Phi-2, are: - Efficiency: SLMs require fewer computational resources, making them ideal for applications where speed and responsiveness are paramount. - Speed: They can perform computations faster, such as generating texts based on given input, making them suitable for real-time applications. - Customisability: SLMs can be fine-tuned based on specific requirements for domain-specific tasks. - Privacy: Smaller models can be used without external servers, ensuring data privacy and integrity. - Cost: Since the models are much smaller, the costs of managing these models also drops significantly.

Andrej Karpathy delivered an incredible talk on "Introduction to LLMs (Large Language Models)". This talk drew a fascinating parallel between modern Operating Systems (OS) and LLMs, making AI feel like an evolving OS for the future. 🌐🤖 LLM, the kernel process of AI's emerging OS, coordinates memory and computational tools like a maestro, orchestrating resources for efficient problem-solving. ��💡 Here are some striking parallels between LLMs and OSs: 1️⃣ RAM - In LLMs, RAM is the context window, determining the maximum words it can use to predict the next in a sequence. Like a skilled juggler, LLMs switch relevant info in and out of this window for optimal task performance. 2️⃣ File Systems - LLMs store info in embeddings, akin to data persisted in an OS file system. 3️⃣ Peripheral Devices - LLMs are versatile! They generate audio, video, and output to peripheral devices. 4️⃣ Ethernet - Like OSs fetching info online, LLMs tap into the internet through various browsers and search engines. The analogy extends to market dynamics too. Proprietary LLMs by Open AI (GPT), Anthropic (Calude), and Google (Bard) mirror the commercial OS landscape, while open source LLaMA-based versions by Meta add an open-source twist. 🔄 LLMs share security challenges with OSs, tackling prompt injection, sleeper agents, multimodal adversarial inputs, and more. The rapid advancements in LLMs keep them on their toes! For a deep dive, check out the video (link in the first comment). 🎥💻 Stay tuned for more tech updates! 🚀 #llm #openai #operatingsystems #techtalk

🔍 Exploring the Frontiers of Generative AI? The advancements in Large Language Models (LLMs) like Falcon 180B are nothing short of transformative. 🌐 Falcon 180B sets new benchmarks, having been trained on an unprecedented 3.5 trillion tokens with the support of 4,096 GPUs, all facilitated by Amazon SageMaker. In total, that's a jaw-dropping 7 million GPU hours! To put it in perspective, Falcon 180B is 2.5 times more expansive than Llama 2 and operates on 4x the computational strength. 🏥🏦 Implications for Key Industries: For sectors like healthcare, insurance, and banking, the doors to leveraging generative AI have been flung wide open. However, we recognize that embracing these technologies comes with its own set of challenges—chief among them being data governance and security. 🔒 Your Path Forward: Don’t let concerns about data security and governance stall your Generative AI initiatives. 📞 Get in Touch:(602) 644-1942 We're here to guide you in safely and efficiently unlocking the potential of these revolutionary technologies. Contact us now to accelerate your AI journey! https://smpl.is/7s0o2

I flew 11 hours and spent the day at the Pinecone AI Transformation Summit. You only need to spend two minutes to learn my top 9 takeaways 👇👇👇 #1 OpenAI's ChatGPT (GPT-3.5/GPT-4) models are still best-in-class for most LLM use-cases and in developer experience for API integration. #2 Building Large Language Models (LLMs) is prohibitively expensive & it’s unlikely many companies will look to build their own. This means Retrieval Augmented Generation (RAG) techniques will play a key role. #3 Very few have deployed Generative AI into production. The 🔑 to success is mastering the evaluation and iteration cycles using both LLM-assisted and direct user feedback (i.e. "was that response any good?"). #4 There are very few fixed rules when it comes to prompt and parameter engineering - you have to experiment! #5 Coming soon is a fully managed production-ready instance of Pinecone with OpenAI on Azure. This is a really interesting play by Microsoft which could see them hoover up more market share from AWS - "If I'm deploying my LLM on Azure, why don't I also deploy my web app and other services there?" #6 You no longer need a deep maths or data science background in the AI space - if you can call a RESTful API, you too can be an AI engineer.  To be successful though, AI engineers will be able to put together the technology components in a way that's creative, optimised, and scalable. #7 You will be tempted to pass more and more context into prompts as token limits increase - but this will likely not improve results. Check out the paper: "Lost in the Middle: How Language Models Use Long Contexts". #8 Data = competitive advantage. But, unlike past years of machine learning techniques, the beauty of RAGs is that the quality of your embeddings trumps the volume of your data. #9 With LLMs, it's vital to validate each stage in the query, context, and response cycle to ensure you're providing the right outputs.

Prompt Engineering is increasingly becoming a much-desired skill. The ability to interact with Gen AI tools and extract the right result with well crafted prompts requires creativity and understanding of how Large Language Models (LLM) work. If you are involved in any kind of research, it becomes even more essential to know prompting and go above and beyond your regular search engine queries. Here are a few useful links to get you started with Prompt Engineering. https://lnkd.in/gkm3vk8R https://lnkd.in/gbmNe384 https://lnkd.in/gP6ZT4Km Note : Image generated via Microsoft Copilot.

When it is time to tokenize text in C# you can use the RAW BPE class from microsoft, you just need to format vocabulary and merge files and you are done. The .NET Generative AI landscape is becoming very interesting. https://lnkd.in/eHqisDV2