Chinar Movsisyan

🚀 Looking to streamline fine-tuning for models like Llama 3, Mistral, and Mixtral? Our latest blog shows how Anyscale simplifies LLMs fine-tuning with Ray: 🔑 Key takeaways: • Scalable fine-tuning with LLM-Forge • Model Deployment with RayLLM Read more: https://lnkd.in/gQwD4A2U

35

💬  Learn how to build a custom chatbot with LLMs + RAG and scalable workflows 🤖 RSVP now to join our Chief ML Engineer, Niels Bantilan, for a live workshop online covering: - Using LLMs - Building RAG system for information retrieval - Prompt engineering - Creating scalable workflows using Union.ai (powered by Flyte) Who should attend: Anyone interested in building custom chatbots with LLMs and best MLOps practices should attend. This workshop is designed to be approachable for most skill levels. Familiarity with machine learning and Python is strongly encouraged but not required. By the end of this workshop, you'll be able to build a custom chatbot using LLMs, RAG, and scalable AI workflows for large language models (LLMs). 🔗 https://lnkd.in/gnTajH3Z #llm #llms #rag

13

1 Comment

🚀 Monitoring Your RAG Pipeline Quality with REMi 🚀 Using RAG effectively means harnessing the power of LLMs to deliver precise answers, based on the most relevant and up-to-date data. But even a finely tuned RAG pipeline needs regular monitoring to maintain its quality. 🔍 Meet REMi, Nuclia • The RAG-as-a-Service company’s open-source RAG Evaluation Metrics framework. REMi helps you assess: • Answer relevance: Is the answer useful for the query? • Context relevance: Is the retrieved data relevant? • Groundedness: Is the answer supported by the data? 💡 With Nuclia, you can monitor these metrics, track trends over time, and identify gaps directly from your dashboard. Plus, our API lets you easily collect user feedback for continuous improvement. 👉 Ready to ensure your AI agents deliver reliable answers? Dive into our latest article to learn how REMi works and how it can elevate your RAG pipeline quality: https://lnkd.in/dGkh579w

2

'OpenAI co-founder Ilya Sutskever's new safety-focused AI startup Safe Superintelligence (SSI) raises $1 billion' "Safe Superintelligence (SSI), newly co-founded by OpenAI's former chief scientist Ilya Sutskever, has raised $1 billion in cash to help develop safe artificial intelligence systems that far surpass human capabilities. The company declined to share its valuation but sources close to the matter said it was valued at $5 billion." "Investors included top venture capital firms Andreessen Horowitz, Sequoia Capital, DST Global and SV Angel. NFDG, an investment partnership run by Nat Friedman and SSI's Chief Executive Daniel Gross, also participated." ""It's important for us to be surrounded by investors who understand, respect and support our mission, which is to make a straight shot to safe superintelligence and in particular to spend a couple of years doing R&D on our product before bringing it to market," Gross said in an interview." "AI safety, which refers to preventing AI from causing harm, is a hot topic amid fears that rogue AI could act against the interests of humanity or even cause human extinction." "Sutskever, 37, is one of the most influential technologists in AI. He co-founded SSI in June with Daniel Gross, who previously led AI initiatives at Apple, and Daniel Levy, a former OpenAI researcher. Sutskever is chief scientist and Levy is principal scientist, while Gross is responsible for computing power and fundraising." #SSI #SafeSuperintelligence #AI #artificialintelligence #machinelearning

31

3 Comments

"Learn Your Reference Model for Real Good Alignment," introduces Trust Region Direct Preference Optimization (TR-DPO). Developed by a visionary team led by Alexey Gorbatovski at Tinkoff, this paper pioneers a method that significantly advances the training of language models by updating the reference model during the training process itself. Why This Matters: ➡️ Enhanced Model Alignment: TR-DPO introduces dynamic updates to the reference model, which allow for more accurate and adaptive training compared to the traditional Direct Preference Optimization (DPO). ➡️ Improved Performance Metrics: Through rigorous testing, TR-DPO has demonstrated a 19% improvement in performance over DPO, making it a groundbreaking tool for achieving higher coherence, correctness, and detail in model outputs. Key Insights: ➡️ Dynamic Reference Model Updates: Unlike static methods, TR-DPO continuously adjusts the reference model during training, which keeps the model relevant and highly effective across changing data sets and conditions. ➡️ Human-Centric Results: The model notably excels in producing outputs that are coherent, detailed, helpful, and harmless—key qualities for any application from chatbots to complex decision-making systems. #analyticsvidhya #datascience #generativeai #researchpaper

16

Interested in learning how to build and refactor a high-performance #Rustlang app for real-time data streaming? In our recent workshop, thousands of engineers saw how to: - Create and compile a sample social media app w/Rust - Connect the app to #ScyllaDB and Redpanda Data - Negotiate tradeoffs related to data modeling and querying - Manager and monitor the database for consistently low latencies Check out the full recap here > https://ow.ly/RymT50TyNWk #NoSQL #lowlatency #database #techtips

24

Interested in learning how to build and refactor a high-performance #Rustlang app for real-time data streaming? In our recent workshop, thousands of engineers saw how to: - Create and compile a sample social media app w/Rust - Connect the app to #ScyllaDB and Redpanda Data - Negotiate tradeoffs related to data modeling and querying - Manager and monitor the database for consistently low latencies Check out the full recap here > https://ow.ly/RymT50TyNWk #NoSQL #lowlatency #database #techtips

64

🌟 LLMs Compared: GPT-4o, Gemini Pro 1.5, Llama 3, Mistral 8x22B, Claude 3 ❓Why This Matters: 🔹LLMs are revolutionizing everything from content generation to conversational AI, enhancing efficiency and strategic decision-making. 🔹However, choosing the right model for your use-case based on areas where each of these models excel is a challenge since we have so many great choices available. 🔹Learn more in this free hands-on session on Wednesday, June 20 @ 10-11 AM PST that will explore the diverse applications and transformative potential of LLM such as OpenAI's GPT-4o, Google's Gemini Pro 1.5, Meta's Llama 3, Mistral AI's 8x22B, Anthropic's Claude 3 is crucial for professionals across industries. 📚 What You'll Learn: ➡️ Key use-cases of leading LLMs including GPT, Gemini, Llama, Mixtral, Claude ➡️ Practical applications of LLMs in different sectors and their impact on business and research ➡️ Future trends in #GenAI and predictions for the next generation of LLMs ➡️ Insights into the ethical considerations and challenges facing the deployment of LLMs 👉🏼 Registration link in the comments. #artificialintelligence #llms #ad

46

1 Comment

Here’s a tip: If OpenAI’s new model #o1 is too slow for you, you can use gpt-4o to do the first few steps , and then finish up with o1 For example: I like to create high-quality blog articles with AI - which requires tons of processing and 20-30 steps. So I use 4o for this step. I then use o1 in the final step to act like the editor of New York Times and optimize the article for publishing. This gives the best of both worlds.

13

2 Comments

Meta just released Llama3, and tonight at ZAAI.ai we’ve been hacking away and testing its boundaries! First impressions below 👇 Key Features of Llama3: - Model Variants Available: 8B and 70B - Training Data and Strategy: Trained on an impressive 15 trillion tokens (!) using a custom-built 24k GPU cluster. Also, they have used 4x the amount of code related tokens compared to llama2. - Context Length: 8k. - Performance Benchmarks: The Llama3-8B has shown to outperform the Llama2-70B in various scenarios, specially those requiring logical reasoning and coding. It also shows best in class performance for models of similar size: e.g. Gemma and Mistral 7B or Gemini pro 1.5 and Claude 3 Sonnet. - Llama 3 Instruct has been optimized for dialogue applications and was trained on over 10 Million human-annotated data samples with combination of fine-tuning and reinforcement learning. 🔍 Testing Experience: From our initial testing, Llama3 has showed a strong grasp of knowledge, significantly improved over Llama2 (cutoffs are also getting closer to the release dates - end of 2023 for the 70B and March 2024 for the 7B). The capabilities for reasoning and programming tasks are robust. However, its creative writing outputs are not as refined, possibly due to a substantial positivity bias — something that might limit its use in unrestricted creative scenarios. Despite these limitations, the 8B model’s size is ideal for fine-tuning, which we anticipate could address some of the creative writing challenges. The current 8K context length is not great and honestly is somewhat below expectations. It was mentioned in the release notes that it will be expanded in future iterations, much like what was seen with the Mistral-7B upgrade from 8K to 32K. In our experiments, Mistral models still feel a bit more realistic and the chain of though is more coherent. Nevertheless, for logical puzzles and programming tasks llama3 is possibly the best in their class. It almost feels that they were a bit too optimized for those tasks (specially for a general-purpose suite of models). 🎙️ Insights from Zuckerberg's Interview: Zuckerberg’s recent podcast appearance covered the integration of a version of Llama3 into Meta's suite of products. This integration aligns with ongoing developments, including the exciting news of a 400B model (still in training) currently matching GPT-4’s performance metrics, set to release later this year. Another good day for open source! Stay tuned for more updates as we continue to test it. #AI #Llama3 #OpenSource #MachineLearning #ArtificialIntelligence #Meta #TechnologyNews #Innovation

44

🚀 Just slashed our ML feature store latency by 95% - here's how we did it The silent killer in many ML pipelines? Inefficient feature stores. We discovered this the hard way when response times jumped from 50ms to 2 seconds in production. Key optimizations that transformed our performance: Implemented feature computation graphs to eliminate redundant calculations Deployed smart multi-level caching with hot/warm/cold feature strategies Built intelligent batch processing with dynamic sizing The results were dramatic: ✨ Feature computation 📉 CPU utilization 💰 Storage costs 🎯 Model accuracy Check out our detailed technical breakdown and implementation guide here: https://lnkd.in/d3pf6e8P

4

1 Comment

Thank you, Sion Smith, for having us! 🎙 In this talk, Yingjun Wu introduces RisingWave, a scalable streaming database that utilizes #S3 as its primary storage. From underlying technology to real-world scenarios, Yingjun illustrates how RisingWave supports heavy traffic workloads in enterprise-grade production environments. #risingwave #streamingdatabase

3

Massachussetts Attorney General's Advisory on AI to developers and others: https://lnkd.in/diZgtniY. The advisory underscores the need for strong guardrailing solutions and why we at ZenGuard AI are tackling this problem head on. Yes, we educate our users, but it makes it so much easier when AG echoes our warnings. * Data security is paramount, and the Data Security Law, G.L. c. 93H, still applies to all AI systems. * With all the benefits of AI, it poses risks like bias, lack of transparency, and data privacy issues. LLMs are opaque to consumers acting as a black box, leading to financial, reputational and ethical risks - when AI does not function as intended or meet expected standards. * AI systems must not falsely advertise their capabilities or fail to perform as advertised. * Protections around personal data security is doubled down - developers must safeguard PII/IP and other personal information (with necessary notification requirements). Failure to comply with these standards can lead to violations under the Massachusetts Consumer Protection Act. Developers must ensure their AI systems are robust, secure, unbiased, and transparent to avoid potential legal challenges. Talk to us how we can help you to ensure that your AI systems are secure and safe at hello@zenguard.ai. #massachussetts #genai #llmguardrails #llmsecurity

12

🚀 Stop wasting your GPU resources on blanket Fine-Tuning for LLMs. Instead, be strategic and only fine-tune the layers that truly matter. The results might just surprise you! 🎯 We recently fine-tuned the unsloth/phi3-medium-4k with a clear objective: Enhance the model's skills across various domains while maintaining or boosting its performance in German RAG Eval, all while minimizing GPU resource usage. 🔧 We aligned it with DPO using Spectrum QLoRA Fine-Tuning – an innovative method that cuts down on resource consumption during LLM training. The outcome? A remarkable improvement in the model’s capabilities! Our benchmarks speak volumes. The newly introduced mixeval, with a 96% correlation with Chatbot Arena, showcases our success. Its regular updates make it a gold standard for LLM evaluations. (Huge thanks to Jinjie Ni and Team for your incredible work!). 🔍 Benchmark Highlights: ⏩ GPT4ALL AVG: SauerkrautLM-phi3-medium (79.51) vs. Unsloth/Phi3-medium-4k-instruct (74.62) ⏩ Ger Rag Eval AVG: SauerkrautLM-phi3-medium (97.1) vs. Unsloth/Phi3-medium-4k-instruct (96.7) ⏩ Mixeval Hard AVG: SauerkrautLM-phi3-medium (56.7) vs. Unsloth/Phi3-medium-4k-instruct (54.3) 👉 Model and more benchmarks available on Hugging Face: https://lnkd.in/emu33-Bd 🔧 Fine-Tuning Details: ⏺ Base Model: Unsloth/Phi3-medium-4k-instruct ⏺ GPU Time: 15.5 hours ⏲ ⏺ GPUs: 4x RTX 6000ADA ⏺ Costs: $70.68 💴 ⏺ Training Method: Spectrum QLoRA DPO ⏺ Selected Layers: 50% ⏺ Data: 70k samples 💿 ⏺ Epochs: 1 ⏺ New Model: SauerkrautLM-phi3-medium 💡 Key Takeaways: ⏩ Fine-tuning only a fraction of an LLM (50% in our case) can significantly enhance a model. ⏩ Targeting 50% of the layers allows for higher learning rates. We used 5e-04, avoiding gradient explosion. ⏩ A reduced warm-up ratio (0.03) enabled earlier intensive training phases. Training on just half the model dramatically cuts down training time and costs. A big shoutout to David Golchinfar, Fernando Fernandes Neto, Lucas Atkins, and Eric Hartford for contributing Spectrum to the open-source community. 🌟 👉 Find Spectrum here: https://lnkd.in/eTXeaQGn #AI #MachineLearning #DeepLearning #LLM #FineTuning #GPU #Benchmarking #TechInnovation

41

Exploring AGI, the challenges of scaling and the future of multimodal generative AI https://lnkd.in/esu4F4Ry

5

💡Blog Post Alert: Explainability (XAI) techniques for Deep Learning and limitations💡 As AI continues to transform industries, understanding how models make decisions is crucial. In our latest blog, we explore key explainability techniques for deep learning models, including visualization-based methods like Grad-CAM and distillation methods like LIME. While these tools help bridge the gap between complex AI systems and human understanding, challenges like computational complexity and data-specific limitations still remain. Check out Part I of our blog series to learn more about the current state of XAI and the road ahead for AI transparency. Stay tuned for Part II on the opportunities explainability can unlock! Read more 👉 https://lnkd.in/gzqY9pD2 #AI #ExplainableAI #MachineLearning #AITransparency #DeepLearning #AryaXAI

9