Databricks Mosaic Research

We're kicking off 2025 with another Compound AI System Meetup with LanceDB in Mountain View on Jan 22! ?? Join us for a deep dive into AI infrastructure and insights from an all-star lineup: ? Lu Qiu ? Allison Wang ? Holden Karau ? Dr. Sharon Zhou Don't miss your chance to connect with experts in data and AI! ?? Mountain View ?? Jan 22, 6:30 PM - 9:00 PM ?? Save your spot: https://lu.ma/5ech3qaz

We're excited to share this blog post on Benchmarking Domain Intelligence: https://lnkd.in/dsca-rnW Evaluating your #AI solutions should be done with tests that match your actual use case. In our research, we observed that the tasks in the widely accepted academic benchmarks for AI system performance differ significantly from the tasks that drive business-relevant results. We developed the Domain Intelligence Benchmark Suite (DIBS) to help Databricks customers build better AI systems for their specific use cases, and to advance our research on models that can leverage domain intelligence. DIBS measures performance on datasets curated to reflect specialized domain knowledge and common enterprise use cases that traditional academic benchmarks often overlook. Thanks to the authors: Pallavi Koppol, Erica Ji Yuen, Kartik Sreenivasan, Yue (Andy) Zhang, Sam Havens, Michael Carbin, Matei Zaharia, Jonathan Frankle

We're back from Vancouver and #neurips2024. Thanks to all the #genai researchers, practitioners and international pop stars who joined us at our Databricks Mosaic AI social last Wednesday night!

Thanks to Singapore Global Network and 1943 for hosting many of our Databricks AI research team (including guest speaker Jonathan Frankle) at their #neurips2024 brunch collective yesterday.

#TaylorSwift may have wrapped up the Eras Tour but we’re still in our Data and AI era! Stop by our booth at #NeurIPS2024 to chat all things research and meet #Brickster Swifties. For more information on our accepted workshops, see our blog post here. https://lnkd.in/gKYMGmzC

New blog post! We explore one method for customizing LLMs — Continued Pre-Training (CPT) — and provide guidance on executing this process effectively: https://lnkd.in/gVD2fAHP Continued Pre-Training refers to a cost-effective alternative to pre-training large language models (LLMs) from scratch. While LLMs are increasingly adept at solving general tasks, they can often fall short on specific domains that are dissimilar to the data they were trained on. In such cases, how do you effectively and efficiently adapt an open-source LLM to your needs? With CPT, Mansheej Paul, Brett Larsen, Connor Jennings, and Cody Blakeney demonstrate how to enhance a small LLM’s factual knowledge performance to match that of a much larger LLM by augmenting the small model’s general knowledge with specialized information. Check out the post for details!

Tomorrow's meetup is going to be terrific! Join us with the registration link in the post below.

As #GenAI projects move from POC to production, Databricks?customers are shifting from deploying single models to leveraging AI agent systems. Naveen Rao, Matei Zaharia and Patrick Wendell describe how modular system design has led to a new development paradigm for intelligence applications: https://lnkd.in/gxamqEQB

We're ready for our next SF meetup, with a great set of speakers: * Anne Holler - Elotl * Bryan Bischof - Hex * Chang She - LanceDB * Daniel Svonava - Superlinked Looking forward to seeing you there! Registration link below.

Had a great meetup at the Databricks Mountain View office! Thanks to Ty Dunn at Continue, Lei Xu at LanceDB, and Sunish Sheth for presenting, and thanks to my friends and colleagues for helping to make this happen: (LanceDB) Jasmine Wang (Databricks) Elizabeth Sapiro Santor, Torey (Markowitz) Bublitz, Olaf Hubel, Ester Shmulyian, and Cat Vo Fantastic insights about compound AI systems and what it takes to do them well: * Continue's approach to AI-enhanced coding really demonstrates empathy for the developer. Also, I'm rooting for them as a component of an ecosystem that can provide every software engineer (and engineering organization) a customizable suite of capabilities, tailored to each specific team, codebase and workflow. * LanceDB's speed, price/performance, and scalability are game-changing! * MLflow's GenAI tracing and evaluation features, and the Databricks Agent Framework showcased how AI-focused organizations can iterate quickly on their applications to deliver the most accurate and relevant capabilities. Already looking forward to the next meetup in SF! Subscribe to the calendar for upcoming events here: https://lnkd.in/gNMeZn7e