Steve Ash

Interesting new open source Event processing platform from Microsoft. https://lnkd.in/gPr4Cifu The features that I am interested in are 1. It’s input schema is modelled on Debezium, another open source CDC product. 2. Enables action on an event 3. Write continuous queries using Cypher query lanaguage Combining such a powerful platform with a LLM would be sone pe suhaga (best of both worlds).

5

1 Comment

If you'll be at the #DataAISummit, join Ashok Singamaneni and myself (Scott Haines) as we teach you the ropes of building "Effective Lakehouse Streaming with Delta Lake and Friends". https://lnkd.in/gqmcJmRC We'll be teaching lessons learned (the hard way), and showing how streaming-first is the correct fit for "most" if not all your workflows. Learn to leverage #DeltaLake like an expert, and we'll introduce you to some effortless patterns for making your streaming journey a breeze. You'll also hear from Ashok Singamaneni as we talk about how to leverage painless workflow orchestration with #brickflow, and how to build data trust-networks using #sparkexpectations. #spark #deltalake #sparkexpectations #brickflow #nike #oss #dataai

91

11 Comments

Imagine seamlessly ingesting incremental data from *any* API right into #DeltaLake via Structured #streaming. This article by Irfan Elahi goes deep into the new Custom Data Source API in #apachespark, showing how you can create your own data source for streaming - with checkpointing capabilities to achieve exactly-once guarantees! https://lnkd.in/gEvr5kQJ

17

The Consistent Pitfall of Metaprogramming One of the cardinal principles of clean code is DRY: "Don't Repeat Yourself." When you find yourself needing to do the same thing multiple times, the DRY principle says to develop a function that encapsulates the need and lets you invoke the same behavior from multiple locations. Sometimes, though, you repeat yourself in ways that aren't easily function-ified. Enter metaprogramming: manipulating your program's logic with constructs other than the core language. Code generators, minilanguages, compiler macros, embedded scripting languages, and reflection-based frameworks are all examples of metaprogramming. Metaprogramming is popular for problems like dependency injection or async frameworks, which share the problem of requiring a lot of tedious wiring in between the core bits of application logic. I've written more minilanguages than I'm proud to admit and also a fair bit of reflection magic. And here's what metaprogramming has taught me about the world of mundane non-meta programming: it's all about the tools. I'll elaborate. When your mundane program isn't behaving as expected, you have a host of mundane tools to help you diagnose it. You have an IDE that can jump from method invocations to their definitions. You have debuggers that can set breakpoints and watchpoints to catch conditions of interest. You have stack traces that tell you what larger purpose was being served when something failed. You have profilers that can add up the CPU time spent satisfying a function call. And you have a ton of personal experience using these mundane tools to diagnose mundane bugs. Metaprogramming breaks all of that. Ever seen a stack trace from an async framework? It just says "Failed in AsyncFramework.doSomeAsyncStuff()." What are you supposed to do with that? This highlights the fundamental error that we consistently make when we metaprogram. We develop a brilliant framework for concisely representing correct concepts when everything goes well, but we break all of an engineer's traditional tools for diagnosing the problem when something goes wrong. Things that used to exist in the stack trace now exist in an in-memory graph representation that can't be cleanly printed to stdout. "The code is just data" is a powerful concept, but data is harder to read than code. Metaprogramming is a powerful tool. That's why we reach for it. But sophisticated tools also call for sophisticated diagnostics. When you build a metaprogramming framework, consider not just how beautifully concise and intuitive the code looks when it's working well, but also what mundane tools you've broken for the poor engineer who has to debug this framework when it malfunctions. Replacement tools need to be a part of the framework you provide.

46

4 Comments

🚀 Exciting Announcement from Databricks! 🚀 General Availability of Serverless Compute for Notebooks, Workflows, and Delta Live Tables! 🎉 With this powerful addition, we can now: 🔹 Simplify Infrastructure Management: No more hassles with provisioning and scaling compute resources. 🔹 Boost Productivity: Seamlessly run your data workflows and notebooks without interruptions. 🔹 Optimize Costs: Pay only for what you use with our efficient serverless architecture. This is a game-changer for data engineers, analysts, and scientists who want to focus more on innovation and less on infrastructure. 💡 https://lnkd.in/gt5b_uCF #DataScience #BigData #Serverless #Innovation #DataEngineering #MachineLearning #DataAnalytics #Databricks

4

The new GPT4 turbo model with vision is available on Azure Open AI!

9

With new BigQuery ML functions, it is now easier than ever to monitor model skew and data drift in BigQuery. These functions are super useful if you perform your model training directly in BigQuery ML, but also if you use BigQuery for your data and feature management and preprocessing. You can read more in the blog post by Mike Henderson and Yinguang (Ian) Zhao – thank you! https://lnkd.in/gBGMyCsg cc Omid Fatemieh, Xi Cheng, Vaibhav Sethi, Michael Kilberry, Irina Farooq, Tomas Talius

75

1 Comment

No one likes running out of reservation time when running an AI and ML on AWS Here’s how to take your AWS capacity blocks game to the next level… If you don’t want to reserve too much buffer time time which directly means overspending, reserve time closer to what you think you need and use a checkpoint + restore based strategy to protect from losing your work in the case that you need more time than your reservation allows. On the flip side, reserve some extra time and start running another training exercise without fear of getting interrupted halfway through via having a checkpoint to restore from. Met with a large bank recently who loves this idea and it compliments well their idea of using AWS as a cloud bursting environment. MemVerge Memory Machine engine is a perfect solution for these types of cloud checkpoint restore use cases. Would love to hear thoughts and comments from others. Have you used capacity blocks and run out of time mid-training or fine tuning / RAG? Background on what these are in case you’re coming into this brand new. “Capacity Blocks for ML allow you to reserve highly sought-after GPU instances on a future date to support your short duration machine learning (ML) workloads. Instances that run inside a Capacity Block are automatically placed close together inside Amazon EC2 UltraClusters, for low-latency, petabit-scale, nonblocking networking.” #AWS #HPC #AI #MemoryMachine

6

1 Comment

It gives us great pleasure to announce that Apache Spark in BigQuery is now officially available in Private Preview. This feature expands BigQuery’s processing capabilities by offering Apache Spark as an OSS processing engine under a unified control plane (admin, monitoring, billing, security, governance). It is the first step in delivering on the Unified Data Platform vision and was a key announcement at Next ‘24 #googlecloud #bigquery #spark

38

2 Comments

Very excited that IBM released Apache 2.0 licensed #granite models on Hugging Face - you can find more info on these models at: https://lnkd.in/epTZw-xS * comes in various sizes: 8B, 2B dense and 1B, 3B MoE * datasets and mixtures are transparently shared * power law learning rate and two phase training * mup for improved hyper parameter selection - developed by Microsoft and popularized by Cerebras Systems * comparable performance as other OSS models of similar sizes like #llama3.2.

93

🌟 Upgraded Claude 3.5 on Vertex AI: A New Era of AI-Powered Applications 🌟 🎉 Google Cloud and Anthropic just dropped some exciting news: Claude 3.5 has arrived on Vertex AI! This release brings cutting-edge conversational AI capabilities directly to Google Cloud’s enterprise-ready ecosystem, enabling organizations to seamlessly leverage advanced AI for various business applications. Here’s why this is a game-changer: read details in my colleague Nenshad Bardoliwalla ‘s excellent blog on this below 🚀 Top 5 Reasons to Explore Claude 3.5 on Vertex AI 🚀 1. Effortless Integration & Enhanced Customization • Vertex AI makes it simple to deploy and scale Claude 3.5 across diverse enterprise environments. By integrating directly into the Google Cloud ecosystem, businesses can fine-tune the model to suit industry-specific requirements, ensuring better, more accurate responses tailored to their unique needs. 2. Increased Performance for Natural Language Understanding • Claude 3.5 offers improved contextual understanding for complex queries, with the ability to handle intricate conversations seamlessly. Whether it’s customer support or analyzing insights from unstructured data, Claude’s advancements in NLP offer accuracy and depth unmatched by previous iterations. 3. Higher Standards in Security & Compliance • Operating on Google Cloud’s secure, compliance-friendly infrastructure means Claude 3.5 is ready for regulated industries like healthcare, finance, and more. Security certifications on Vertex AI help companies maintain high standards in data privacy, making it an ideal choice for handling sensitive information. 4. Cost-Efficiency Meets High Scalability • As more businesses adopt conversational AI, the need for cost-effective solutions grows. Vertex AI provides scalable infrastructure that adjusts to demand, making Claude 3.5 accessible without excessive overhead. Companies of any size can now tap into AI capabilities previously reserved for larger players. 5. Innovation in Multimodal AI • Claude 3.5 supports multimodal capabilities, enabling more versatile applications that go beyond text-based interactions. Imagine using voice, images, and other inputs to engage in richer interactions, opening up possibilities in fields like retail, where visual and verbal inputs are crucial for product recommendations. 💡 What’s Next? For businesses, this isn’t just an upgrade—it’s a step toward AI-driven transformation that enhances customer experiences, streamlines workflows, and drives innovative new services. The Claude 3.5 integration with Vertex AI paves the way for AI that’s reliable, robust, and ready to handle real-world complexities. Curious to learn more? Explore Google Cloud’s announcement here to see how your organization can leverage the power of Claude 3.5 on Vertex AI. #AI #MachineLearning #VertexAI #Claude3 #Innovation #GoogleCloud #EnterpriseAI #NLP #DigitalTransformation

22

For all their success, LLMs have trouble following prescribed sequences of operations, whether operational workflows or API dependencies. At this year's North American Chapter of the Association for Computational Linguistics (NAACL 2024), Amazon researchers showed how to address this using dependency graphs and constrained decoding. #generativeAI #llm #AWS #graphs #api #agents

10

I'm happy to share the Preview launch of Fine-grained DML in BigQuery. With Fine-grained DML, BigQuery tracks table mutations at row granularity reducing write amplification to improve efficiency and latency of DML statements. Deleted rows are garbage collected (GC) using a hybrid background/foreground approach that amortizes the cost of rewrites across many mutations to the table. The system offloads GC whenever possible to the background, but piggybacks on user mutations when necessary. At read time, the row level bitmaps are applied to skip deleted rows. Metadata is indeed Big Data! We use distributed processing techniques to manage and query this Big Metadata. The bitmaps are maintained compactly so that large scale mutations, where a majority of rows in the table are rewritten, continue to work without metadata bottlenecks. https://lnkd.in/gTkRm9qc

370

12 Comments

Awesome to see the Meta #Llama 3.2 models on #Azure #AI Model Catalog through managed compute inferencing #MaaS and soon on serverless API deployments while MACC Eligible providing customers with Enhanced Security and Compliance, Content Safety Integration, Simplified Assessment of LLM flows, Client integration, and Simplified Deployment and Inference. https://lnkd.in/g4Zpb3Dm

9

1 Comment

https://lnkd.in/gUpFEbby One of the limitations of current generation #LLM is it's inability to consume non linear data at scale. #SpreadSheetLLM Attempts to get rid of this limitation by focusing on most commonly available non linear data, that is tabular data.

5

1 Comment

Anthropic's enhanced Claude 3.5 Sonnet model is now available on Vertex AI, offering developers a powerful platform to build and deploy cutting-edge AI applications. Why choose Claude 3.5 Sonnet on Vertex AI? Enhanced performance: Experience significant improvements in coding, reasoning, and creative writing tasks. Seamless integration: Leverage Vertex AI's robust features for MLOps, data management, and model monitoring. Accelerated development: Move your AI workloads from concept to production faster with fully managed infrastructure and advanced developer tools. Get started with Claude 3.5 Sonnet on Vertex AI and unlock new possibilities in AI development. #Claude3.5 #VertexAI #GoogleCloud #Anthropic #AI Get all the details, here →

3

Fascinating to me how engineering innovation at the tooling level is equally magical to and keeping up with the LLM frontier model innovation! Super exciting progress with the #Azure #AI Model Inference API as a unified API to awesome models from Cohere, Meta, Mistral AI, Core42 and more without having to change orchestration code while integrating with Azure GenAI Gateway within API management for a scalable architecture. https://lnkd.in/ghutPvBp

28

I have published a new tech blog post, the second in a series, detailing my experience optimizing S3 bucket listing code using Amazon Q Developer through three rounds of interactions. Despite encountering initial bugs, Amazon Q Developer’s suggestions, such as reducing API calls and using list comprehensions, ultimately led to fully runnable optimized code with a 150% performance improvement. This highlights the value of Amazon Q Developer in iteratively optimizing code. 🏂 🏄‍♂️ 🏄‍♀️ https://lnkd.in/gCJDuFhw #AmazonQ #AWS #GenerativeAI #GenAI

9

More details about the mighty phi-3 family of SLMs. Such a key development as GenAI Dev becomes more about always optimizing and orchestrating with LLMs routing to LLMs and SLMs continuously to balance complexity, latency, cost and more! https://lnkd.in/gGQYqNKt

9