Biplab Mahadani

#GenAI models necessitates training on extensive datasets, which in turn requires robust #cloud storage solutions. As Gen AI continues to advance, demand for cloud services would surge correspondingly, creating a symbiotic growth relationship b/w the two. https://t.ly/9Rg-C

9

Snowflake's flexibility is a double-edged sword for budgets. To gain control: Understand who is using v/s abusing the system Right-size compute resources for workloads Educate users on cost-aware querying practices By combining these strategies with regular spend analysis, you can transform Snowflake from a budget buster to a cost-effective powerhouse. Remember, predictability is key to long-term success in the cloud. #dataactionability #snowflake #finops #dataobservability

32

2 Comments

Interesting demo for those who utilize Snowflake and are progressing on their Gen AI journey. #snowflake #cortex #genai #chatbot

4

1 Comment

Snowflake AI Research Introduces Arctic-SnowCoder-1.3B: A New 1.3B Model that is SOTA Among Small Language Models for Code A Three-Phase Pretraining Approach Using 555B Tokens, Outperforming 1T Token Models with 36% Higher Performance in Code Generation The research team from Snowflake AI Research, University of Illinois at Urbana-Champaign, and Seoul National University introduced Arctic-SnowCoder-1.3B, a novel approach to pretraining code models by progressively refining data quality over three distinct phases. This method combined general pretraining, continued pretraining with high-quality data, and final pretraining with synthetic data. The researchers leveraged existing datasets, such as The Stack v1 and GitHub crawls, and artificial data generated using Llama-3.1-70B to build a smaller, more efficient model. This process focused on optimizing the data used in each phase to ensure that the model could outperform its competitors. In the first phase, Arctic-SnowCoder was trained on 500 billion code tokens derived from raw data sources such as The Stack v1 and GitHub. This data underwent basic preprocessing steps, including filtering and deduplication, resulting in approximately 400 billion unique tokens. During this phase, the model was trained without advanced quality filters, and the data was grouped by programming language and repository. This approach ensured a broad code knowledge base but required further refinement. In the second phase, the research team selected 50 billion tokens from this initial dataset, focusing on high-quality data. A BERT-based quality annotator was employed to rank code files, and the top 12.5 billion tokens were repeated four times to train the model further. This phase significantly improved the data quality, as the annotator was specifically trained to select tokens aligned with the model’s downstream applications...... Read our full take on Arctic-SnowCoder-1.3B: https://lnkd.in/gm7kjnrW Paper: https://lnkd.in/grWK_wNu Yuxiang Wei Hojae Han Rajhans Samdani Snowflake Snowflake Developers #smalllanguagemodels #languagedemols

43

3 Comments

Here is a wonderful paper Published last week by MIT, Drexel, Harvard, BU & Lehigh on LLM benchmarks using Game Theory with strategic scenarios & Possibility of Nash equilibrium with LLM reasoning. :) https://lnkd.in/dGKSC6uK #ai #data #datascience #ml #gametheory This is probably the start of multiagent systems/LLMs coming together at the same time for a given use case :)

18

Some thoughts/observations from my recent visit to the Databricks Data AI summit. If you're a small enterprise:- Looking for a data lake/mesh BI AI/ML platform? Just buy Databricks and be done with it. At best, do a POC with another platform e.g. Snowflake and go for the lowest bidder. These are evolving rapidly and feature parity is a constant catchup so don't go only based on what exists today. If you're a query engine/federation tool (e.g. Starburst/Dremio):- Delta (the table format) has been open source for a while but Photon (the engine) isn't (and may never be). So Databricks has been under pressure from the vendor/partner ecosystem as well as large enterprises to support non-vendor-lock-in options. It's acquisition of Tabular makes this space interesting and here's hoping the ambition to support Iceberg as a first-class citizen is genuine. Snowflake also made the Polaris announcement for Iceberg support recently. Freshen up you're price/performance stats, you're going to need it. If you're a large enterprise Governance/Unity Catalog Soo many presentations all claiming Unity as the 'single' point of governance solving all their problems. But is any large customer really all-in on Unity? Nobody I spoke with can claim that 'single' is a reality. Now they all have one more catalog. Every large enterprise that uses Databricks also uses <name-your-fav-alternative>. While Unity helps with governance when data/models are accessed from within Databricks - it doesn't push down to the underlying external platform - e.g. it isn't going to generate Lake Formation grants for you. So you'll need two enforcement points for folks who access that Glue table from LF vs from Databricks/Unity. Immuta - can you wake up quickly to this opportunity pls? ABAC (Attribute Based Access Control) Zeashan Pappa/Kristen Wilder had an interesting talk on what's coming with Unity and ABAC. They're going with tags at the Account level as part of the design. Most Databricks customers will have one account for the entire org. Some very large customers probably have one account per LOB within the org. Either way, curious to know how many such orgs have governance centralized at such a high level instead of being federated deeper. If you are more federated, are we back to CS's fav problem - naming things? Domains/namespaces etc. E.g. two different parts of an org may not interpret 'client' the same way. ML capabilities and integrations Every AutoML vendor will need to compete with code-gen, it will not be sufficient for any auto-generated model to be adopted without a need to tweak the underlying code. The Unity and git integrations around MLFlow make it very compelling. How they continue to evolve Mosaic and better integrate into the platform will be interesting to follow. Still feels like very early days on that - feels a bit square-peg-round-hole still. All in all - tremendous pace of innovation. Lots of hype but much of it is deserved.

61

2 Comments

At their Data Intelligence Day, Databricks recently announced its partnership with Indian customers such as UPL, Air India Limited, Aditya Birla Fashion and Retail Ltd., Freshworks, InMobi, Meesho, Myntra, Parle Agro Pvt Ltd, and many others. One of these is UPL Limited, a multinational company headquartered in Mumbai with a market cap of $4.37 billion and a portfolio that includes crop protection, seed treatment, post-harvest solutions, and other agricultural inputs in more than 138 countries. Having experienced massive growth in recent years, the partnership between UPL and Databricks has become critical for the former. “In 2022, we were at the end of a massive growth period. We grew 2x in about three to four years,” Vijay Balakrishnan, ex-chief data & analytics officer at UPL Group, told AIM on the sidelines of the event in Bengaluru on March 22. This growth triggered the need for UPL to leverage data more effectively to make informed decisions. Hence, the largest manufacturer and distributor of agrochemicals in India with a market share of around 20%, has actively started using Databricks’ Unity Catalog, which democratises data across the organisation. By leveraging Databricks, UPL also aims to enable AI solutions across 20 countries, with the potential to scale up to 50 countries in the future. For instance, they conduct extensive farmer outreach programmes, providing training and advisory services to help farmers adopt best agricultural practices. The company has also found Databricks’ serverless SQL engine to be efficient, especially in instances where the speed of data pooling is critical. PS: Thanks to Suhani Lalwani, Jacintha Ng, and Kedar Kanvinde for enabling the conversation Read the story for more details:

29

We just refreshed our performance benchmark showing that Mixpanel is up to 7x faster than data warehouse funnel queries. People often ask how we compare to traditional BI tools. While SQL will always have its place, using BI tools for analytics is going to feel more and more like continuing to use the fax machine when email came out. Keep using the warehouse for what it’s meant for: storing and processing business data. With Mirror and our new Warehouse Connectors, bring your trusted data into Mixpanel to automate your reporting and experience true interactive self-serve analytics. https://lnkd.in/dnNawYWa

162

14 Comments