Paul McNamara

Llama 3 70B is already in the top 6 of all LLMs, showcasing better accuracy than much larger, proprietary models. It’s a testament to the power of data quality and showcases that open, small models can deliver leading performance – perfect for #edge deployments. With the release of new models like Phi-3 (mini 4B, small 7B and medium 14B), this trend is clearly accelerating.   Open-source small-to-medium sized models trained on high quality data can deliver market leading performance comparable to the largest proprietary models – great news for bringing #AI everywhere at the edge.   View the public leaderboard below.   #IAmIntel #OpenSource #LLM   https://lnkd.in/g5K_7P5d

160

2 Comments

One of my Wharton students founded a company that is still in stealth mode. During a lunch, he asked me if I thought it was necessary to build a physical prototype to increase the likliehood of success seeking subsequent funding. My answer was, “Yes”. He gently reminded me that digital twins can more accurately portray reality than reality can. It was a very good point. I stuck with my assertion that there was something psychologically powerful about being able to experience a physical object, but I had to admit that his strategy of going digital made sense. Thoughts? Some things are more clearly better twinned than others. Here’s a great example: The Futurum Group Wharton Executive Education | CTO Program Wharton Executive Education | Digital Transformation for Senior Executives Program

7

1 Comment

Making AI simple. The meaning and context of every word and everything can be expressed as a vector. These vectors can be 1 click generated, searched for proximity to get answers in languages we talk. The already generated open source vector sets can be imported. Not to be missed talk by Edward & Nipun.

3

Data streaming has become costly, with vendors effectively acting as resellers of cloud providers' networking and storage—often adding margins without delivering real value to customers, especially in networking costs. At StreamNative, we’ve focused on identifying the most expensive components of running data streaming platforms and addressing the main barriers to interoperability between Pulsar, Kafka, and other protocols. Three primary challenges stand out: ↗️ Expensive Intra-Cluster Replication: Leader-based data streaming technologies (like Kafka, Pulsar, and others) rely on leader-based replication mechanisms to ensure data availability. However, this process incurs substantial networking costs. ↗️ Duplicated Data Copies: Data streaming’s power can become an anti-pattern when overused, leading to unnecessary data duplication. Customers often end up with 4-5 copies of the same data across the lifecycle—on Kafka or Pulsar’s local disk, in S3, loaded into Snowflake, and potentially into Databricks or other analytical tools. This adds unnecessary complexity and costs. While “shift-left” thinking is excellent for governance, it can lead to expensive and redundant storage and compute inefficiencies. ↗️ Costly Data Transfers Between Streaming and Lakehouse Systems: Integrating data streaming with batch processing often involves transferring data between systems, resulting in high networking costs and additional redundancy. To address these issues, we built the Ursa Engine—a data streaming engine evolved from #ApachePulsar, becoming leaderless & headless, to augment Lakehouses with Kafka-compatible, sub-second latency streaming capabilities. Ursa combines stream storage and table storage within a single storage layer. You can store data once and access it via streaming APIs (Kafka or Pulsar) or table APIs (open table formats like Delta Lake and Iceberg). More importantly, you can leverage your existing governance tools to manage data as streams or tables without duplication. We’re thrilled to announce that Ursa Engine entered public preview at the #DataStreamingSummit yesterday! It’s available for AWS BYOC customers, supports the Kafka protocol for data streaming, and the open Lakehosue formats for data storage. Users can achieve sub-second (200-500ms) latency while realizing up to 10x cost savings. If you're interested, please try it out or reach out to us for more information. For more details, check out our announcement blog post: https://lnkd.in/gvfRqqn7

20

Multi cloud GenAI needs, with all the building blocks for modern applications, from MongoDB...

10

With Cohesity Gaia, JSR Corporation is uncovering the hidden potential in their data. We're not just protecting their data—we're making it work for them. And this is just the beginning! Get the scoop from ComputerWeekly.com: https://lnkd.in/eU5ewbyt #AI #genAI #generativeAI

Cloudian recently announced our integration with #AWSMountpoint which allows data stored on Cloudian #datalakes to be mounted as local file systems with seamless, high-throughput access, enabling Cloudian users to leverage the limitless capacity and scalable performance of Cloudian HyperStore through a familiar file interface. Now see how easy it is to connect #AWS Mountpoint to a Cloudian HyperStore bucket in this technical demo with Henry Golas, Director of Technology at Cloudian. Common use cases for AWS Mountpoint include large-scale #machinelearning, #ai, #autonomous vehicle simulation, #genomic analysis, data ingest (ETL), and image rendering. Learn more about the integration: https://bit.ly/3PzUS5h

25