This week marks a very interesting release from our friends at #Anthropic! The new Model Context Protocol -https://lnkd.in/gGeR_P6x - allows for standardization of the tools that enhance #LLM s with user’s or company’s data making them aware of the very recent data or data that is not available on the open web. #RAG This is an important topic for us at LightsOn as we've tackled similar challenges in the consulting engagement with #Cohere. Our team has helped build #Cohere 's Toolkit backend https://lnkd.in/gZX4eDS6, which integrates tools into an all-in-one interface for quickly deploying chat apps. While MCP focuses on standardizing connections, our work emphasizes flexibility, letting developers customize AI applications to their needs. We also helped build 100 (!) open-source connectors to easily enhance your LLM experience with Slack and Gmail data. We will share more next week! Happy Thanksgiving to all! 🦃
Dimitri Gnidash’s Post
More Relevant Posts
-
Some amazing developments and research taking place when it comes to AI models. While the hardware and physical infrastructure side of things has commanded much of investor attention in recent times, the real productivity gains will come when the average company or individual adopts AI. That in turn requires easy to use capabilities that can help the average user do tasks better ... easier ... faster. AI's ability to analyze long videos (1 hr) has been limited so far; requires too much hardware / computational capabilities. However in comes LongVU - released recently by Meta. It's more efficient and more accurate when it comes to analyzing long videos. Try it here (https://lnkd.in/ddJQESPc) Geeky details on this para: in the first instance, LongVU discards / cuts out redundant frames in a video (using Dinov2) i.e. frames that may be duplicative. It then uses the text query to identify frames that are of relevance to the query. It also retains the remaining frames in a much more efficient manner (using spatial pooling / compression - which in itself deserves a lengthy post) while preserving the most important visuals. Others will build on this research. But imagine how LongVU could change the process of going through a security footage video. Or summarizing a lecture video. Or a consultant trying to analyze a video taken over days of a manufacturing process / workflow. Or a medical student going through videos of lengthy surgeries to train and learn a certain aspect of the surgical procedure.
-
9 days to go ... I've organized my agenda for #MicrosoftBuild and I am sharing 2 sessions daily that I am excited to attend, why, and hear from you if I am missing out on anything! If you aren't registered yet don't worry - register for #MicrosoftBuild today (in-person/online options): https://aka.ms/build-promo Session Recommendation 1: Accessibility in the era of generative AI https://lnkd.in/emw2zEpd Looking forward to see how Gen AI is providing a moment to innovate and support people more than ever before. Also keen to learn best practices for the Gen AI apps I build to make sure they are inclusive by design Session Recommendation 2: Take an Azure OpenAI chat application from PoC to enterprise-ready https://lnkd.in/e3zS54u5 Well-architected framework advice for Gen AI apps, sign me up! the pace of moving from POC to production is so fast right now, keen to make sure I understand the best practices for building these into the real world #MSBuild #GenAI #copilots #events #developer Marco Casalaina Ailsa Leen #accessibility #AI Rob Bagby #productionready #waf #architecture
Microsoft Build | May 21-23, 2024 | Seattle and Online
build.microsoft.com
-
Research @RespAI Lab | Technical Executive @MLSA KIIT || ML Enthusiast || HackNITR 5.0 Finalist || KIIT CSE '27
I trained and deployed a Facial Emotion Recognition model on Hugging Face. I found the platform very intuitive, and Gradio made the deployment even easier. For the model I used a CNN with the Adam optimizer and Categorical Cross Entropy Loss function. 🔴 Try the live model: 🔴 https://lnkd.in/eXeeWHjF 📍 Also check Microsoft Learn's free courses on AI: https://lnkd.in/gRWsH8-q #ai #artificialintelligence #aiml #microsoft #deeplearning #mlsa
Fer - a Hugging Face Space by nowimsoham
huggingface.co
-
🚀 𝗥𝗲𝗮𝗹-𝘁𝗶𝗺𝗲 𝗩𝗼𝗶𝗰𝗲-𝘁𝗼-𝗩𝗼𝗶𝗰𝗲 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗔𝗜 𝗖𝗵𝗮𝘁𝗯𝗼𝘁 I’ve recently developed and deployed a real-time 𝒗𝒐𝒊𝒄𝒆-𝒕𝒐-𝒗𝒐𝒊𝒄𝒆 𝒈𝒆𝒏𝒆𝒓𝒂𝒕𝒊𝒗𝒆 𝑨𝑰 chatbot using cutting-edge technologies. The system transcribes voice input using OpenAI's 𝐖𝐡𝐢𝐬𝐩𝐞𝐫 model, processes responses through 𝐋𝐋𝐚𝐌𝐀 𝟖𝐁 using the 𝐺𝑟𝑜𝑞 𝐴𝑃𝐼, and delivers voice output via Google’s 𝐠𝐓𝐓𝐒—all within a seamless Gradio interface. 🌐 This chatbot offers: ◾Real-time conversation without any audio uploads ◾Powerful AI models for accurate transcription and response generation ◾Fully integrated user-friendly interface Feel free to reach out if you’d like to know more or explore collaboration opportunities! This application is designed for real-time interaction within a Gradio interface, making it simple and efficient for users. Check it out here: https://lnkd.in/dV3Cd5Kk #ArtificialIntelligence #GenerativeAI #ConversationalAI #SpeechToText #TextToSpeech #LLaMA #Whisper #gTTS
VoiceToVoice ChatBot - a Hugging Face Space by adnaan05
huggingface.co
-
I run No-Code & GenAI Workshops | Teaching AI Tools for Hiring | Talent Engineering | Always Building, Always Learning
What a conversation this was hearing about BuildShip's inception into the market on the NoCode Alliance show 😍 The hosts always make my day and one cant help but SMILE JJ Englert David Pal It was also so wonderful to get to hear the Co-Founder Harini Janakiraman story. {Thank you} Adore the versatility of your platform, from building pretty much any automation idea to building powerful API's in minutes to a WhatsApp or AI bots. It was great to see how the right technology in place can really support a business! { was checking out your case studies on your buildShip site.} If you are wanting to get a POC or an MVP out there quickly, take a look at their site: https://buildship.com/ #NocoLoco #BuildShip #VisualLowCodeBackEndBuilder #serverlessAPIs #WorkFlows
Exploring the Future of No Code and AI with Buildship | EP #35
https://www.youtube.com/
-
The journey of developing AI apps is a path filled with potential, creativity, and collaboration. We all know that building something impactful takes more than just skill; it requires a supportive community that fosters innovation. At SnowflakeDevelopers, we are dedicated to empowering builders like you to harness the power of AI in your applications. Explore the latest insights shared by our team that highlight the importance of collaboration in this dynamic space. The lessons learned and experiences shared in this article can inspire you to push your boundaries and tap into the collective intelligence of the developer community. Let’s lead the way in creating meaningful applications that drive transformation. We encourage you to share your own stories of development and collaboration. Your journey could motivate others on their path. #AIDevelopment #Collaboration #CommunityInnovation #SnowflakeDevelopers https://lnkd.in/gdA9w7dk
Calling All Builders: Get Hands-On with AI and Apps at Snowflake’s Dev Conference
snowflake.com
-
Many are familiar with Chatbot Arena, which serves as a useful tool for gauging the performance of LLMs from an everyday human perspective. However, the credibility of the ELO score may be a point of contention. Here's how it operates. Additional details and caveats are available for those utilizing the platform.
#LLM eval is hard, understanding how approaches apply to your situation is key .........Here's 5️⃣ take-aways on the LMSYS Chatbot Arena from their recent report 👇👇👇 After almost a year of operation, running over 240k battles from 90k+ users LMSYS released a paper last week diving into the operations of the Chatbot Arena analyzing battle data and sharing insights on the strengths and weaknesses of the evaluation approach. 𝐖𝐡𝐚𝐭 𝐢𝐬 𝐢𝐭: If you haven't used it before, LMSYS Chatbot Arena is a conversational platform where users can input a prompt, which is then processed by two randomly chosen LLMs, either #opensource or proprietary. Users then vote on which model response is best. These user votes contribute to ELO scores, effectively ranking LLMs on a leaderboard based on their performance. 𝐈𝐧𝐬𝐢𝐠𝐡𝐭𝐬: 1️⃣ 77% of the user conversations are in english While not surprising this may diminish the value of the leaderboard for those looking to understand model performance in non-english languages. Most other major languages represent just 2% each of the total conversations. 2️⃣ There's a plan to release 100k pairwise preference dataset from battles So far #opensource community has mostly relied on LLM generated preference datasets, this could be a valuable resource to improve open models. 3️⃣ User votes in the arena closely align with Expert preferences A potential limitation of the arena is that the voting data by users may not be high quality as there's no guarantee users are qualified to make accurate judgements. However, when user votes were compared with preferences of topic experts there was agreement in 70-80% range suggesting this is not an problem. 4️⃣ Conversations diversity is high with over 600 topics After clustering user prompts a diverse range of over 600 topics were discovered. The top 16 topics which included tasks like math, coding, medical Q&A and writing individually represent less than 1% of total prompts. Some had voiced concerns that with users being allowed to choose their own prompts diversity might be low, this appears not to be the case. 5️⃣ Average conversation turns per battle is just 1.3 Given the non real-world scenario most interactions are are short. As a result the model ranking may not be a great representation of performance for more complex multi-turn type use cases. Leaderboard 👉https://lnkd.in/g2Edc3sk Arena 👉 https://chat.lmsys.org/ Paper 👉 https://lnkd.in/gY4jSWP6
LMSys Chatbot Arena Leaderboard - a Hugging Face Space by lmsys
huggingface.co
-
🚀 Tom Warren of The Verge about Microsoft is transforming the #Copilot with new voice and vision capabilities, creating a more personalized AI assistant 🤖 ✨ Key features include: 📰 Virtual news presenter mode 👀 Copilot Vision: see what you see 🗣️ Natural voice conversations 🗒️ "Think Deeper" for complex queries Designed for a warmer user experience, the new Copilot is rolling out on mobile, web, and Windows today. 💻📱 This is just the beginning of a dynamic AI journey! #msftadvocate #AI #VoiceAI #VisionAI Read the full story here: https://lnkd.in/dTqAfPci
-
-
🚀 Tom Warren of The Verge about Microsoft is transforming the #Copilot with new voice and vision capabilities, creating a more personalized AI assistant 🤖 ✨ Key features include: 📰 Virtual news presenter mode 👀 Copilot Vision: see what you see 🗣️ Natural voice conversations 🗒️ "Think Deeper" for complex queries Designed for a warmer user experience, the new Copilot is rolling out on mobile, web, and Windows today. 💻📱 This is just the beginning of a dynamic AI journey! #msftadvocate #AI #VoiceAI #VisionAI Read the full story here: https://lnkd.in/dTqAfPci
-
-
🚀 Excited to share my latest development project! 🚀 Over the past few days, I’ve been working on an AI-driven image generation tool that leverages the power of Flux1.1. The tool generates stunning architectural visuals based on creative prompts. I’ve been using ChatGPT to assist with prompt generation—combining the two seemed like a natural next step! 🔧 How it works: Flux1.1 manages the image generation process, delivering photorealistic results. OpenAI helps craft creative prompts, guiding the generation toward the desired architectural style. 💡 The amazing thing about this project is the speed at which it generates images! I’m looking forward to exploring more enhancements to this tool to assist with design ideation. (personal work unrelated to practice) #AI #GenerativeDesign #Architecture #TechInnovation #OpenAI #BlackForestLabs #FluxAI #ImageGeneration
Co-Founder of Altrosyn and DIrector at CDTECH | Inventor | Manufacturer
3wGiven MCP's emphasis on standardized connections and Cohere Toolkit's focus on flexibility, how do you envision the interplay between these paradigms in a multi-modal LLM architecture leveraging both structured data from knowledge graphs and unstructured text from sources like Reddit? Could a hybrid approach combining MCP's deterministic routing with Cohere Toolkit's dynamic plugin system optimize performance for real-time applications requiring both factual accuracy and nuanced understanding?