Santiago Valdarrama的动态

DeepSeek torpedoed the U.S. AI stocks today, but ?- DeepSeek is open source ?- Gen AI is in early days If you invest in AI stocks and worry, consider: ?- U.S.-based cloud providers can deploy DeepSeek alongside other Gen AIs. ?- OpenAI, Anthropic et al are full of smart people that will learn and leverage DeepSeek's open source code. ?- Just a guess, but I'd bet that the algorithms are more effective with higher performing chips (like Blackwell) DeepSeek is more catalyst than roadblock.

The AI stack is still in its infancy ?? . This means that some parts of the stack have yet to consolidate into a selection of tools, libraries, and frameworks, opening doors for business and investment opportunities. At the same time, other components of the stack have developed to the point where there are industry best practices, leading tooling, and cloud service providers. Questions that AI Engineers and IT Leaders ask themselves are: - Should I use open-source or close source models for my AI application? - What even is open source? - What database should we use? Graph, Vector, Relational or non-relational etc , relational - LangChain or LlamaIndex? Here's a resource that guides you towards answering these questions above and more: https://lnkd.in/ewhyxJSm #ArtificialIntelligence #GenerativeAI #AI #aistack

We have a great upcoming session on Data Governance in the age of AI: Balancing innovation with responsibility, Nov 28th, 4pm GMT with some great speakers. Covering some really interesting topics: - Differentiating data governance from AI governance. ? - The evolving role of data governance in an AI-driven world and - Establishing scalable frameworks that support ethical AI Join for some great insight! #thoughtworks #ethicalAI #responsibleTech #AI

?? BREAKING: C3 AI Launches C3 Generative AI for Government Programs with Google Cloud! ?? Are you tired of endless service delays and long wait times in government processes? What if there’s a way to streamline operations and serve the public more efficiently? Imagine this ?? Federal, state, and local government agencies can now eliminate service delays and cut down wait times, thanks to C3 Generative AI for Government Programs! ???? Here’s the deal: ?? Accelerated Processing: No more frustrating delays in service. ?? Improved Efficiency: Reduce those annoying wait times. ?? Seamless Integration: Powered by the robust capabilities of Google Cloud. With the power of AI, government services are on the brink of a revolution. ?? Did you know? Streamlined services can save countless hours and significantly reduce operational costs. ?? Why settle for the old way when you can jump into the future? Curious to know more? Check out the full story here: [Morningstar on C3 Generative AI](https://lnkd.in/gGWhRNfy) P.S. Ready to embrace the future of government services? Click the link and see how AI can transform public service delivery today! ??

1/ Google Cloud is expanding the grounding capabilities of Vertex AI to enable more accurate and correct AI applications. This includes a dynamic grounding feature that lets the AI model decide whether to use search results or its own training knowledge. 2/ The new Grounding with High-Fidelity mode in the Experimental Preview is designed to reduce hallucinations by relying only on context when generating answers and specifying a source for each sentence. It uses a fine-tuned Gemini 1.5 flash model. 3/ Starting in Q3 2024, users will be able to link AI models to data sets from third-party providers such as Moody's, MSCI, Thomson Reuters, and Zoominfo to improve factual accuracy.

Attention CTOs, CDOs, and Data & AI specialists—as #AI advances, so do data governance challenges. How can you navigate ethical, compliant strategies to support AI-driven innovation? Join our online panel with Thoughtworks, Amazon Web Services (AWS) and BBC to gain essential insights for 2025. Register now to join the discussion—and invite your data and AI teams to be part of it: https://ter.li/l6egg2

Introducing our latest AI search model: Rerank 3.5! Rerank 3.5 delivers state-of-the-art performance with improved reasoning and multilingual capabilities in 100+ languages to precisely search complex enterprise data like long documents, emails, tables, and code. Rerank 3.5 can be easily implemented with just a few lines of code to help businesses enhance their search and retrieval-augmented generation (RAG) systems. Customers like Notion have used our rerankers to help users quickly get relevant answers to their most important questions. It’s available today on Cohere’s platform and Amazon Web Services (AWS) Bedrock and Sagemaker, with additional cloud platforms coming soon. Learn more: https://lnkd.in/e6nVA7ws

Super exciting Cohere announcement with our latest model release with both (a) getting live on AWS Bedrock (b) retrieval and accuracy performance - in VB's words this "will fundamentally reshape how global enterprises handle information retrieval" Full article here: https://lnkd.in/e_Z_ZfNU #GenAI #AWS #AWSBedrock #Cohere

This is huge! We launched our latest Search Model this week. Rerank 3.5 delivers state-of-the-art performance with improved reasoning and multilingual capabilities in 100+ languages to precisely search complex enterprise data like long documents, emails, tables, and code.

I made a quick simple test in Arabic using the multilingual model (rerank-multilingual-v3.0) API and got good results (similar to jinaai/jina-reranker-v2-base-multilingual). What I noticed also was small relevance scores for harder queries (even smaller in Cohere). Earlier, I used Jina reranker on the MMARCO translation dataset (Arabic) and tried to create a knowledge distillation model (COLBERT) with AraBERT as base model. I could retrieve documents against queries but relevance scores were close. I then picked data for training where scores were well-spread (32 docs per query) and trained models with minor improvements. Long story short, translation does not work for this task in Arabic. For those willing to invest in this field (time, talent and money), my advice is to work on original quality data and a native Arabic tokenizer.