A discussion with Venky Ganti, CEO of Numbers Station, and Laurel Orr about the merits and limitations of 'one-size fits all' LLMs, the contrast between general LLMs and verticalized models, the significance of ownership structures, and the future of LLMs. They also explore the challenges of hosting private LLMs, AI's role in data wrangling, customizing LLMs for complex data scenarios, the evolution of LLMs and the need for customization, and evaluation methods for text-to-sequel models.
Read more
AI Summary
Highlights
AI Chapters
Episode notes
auto_awesome
Podcast summary created with Snipd AI
Quick takeaways
The future of LLMs involves customization and personalization, tailoring models to specific use cases and exploring the trade-offs between larger and smaller models.
Democratization of data recipes aims to make high-quality training data sets more accessible, enabling open-source models to compete with proprietary models.
Enterprise customers seek private LLMs to protect their intellectual property, highlighting the need for secure and efficient hosting options.
Evaluating LLMs requires focusing on actionable outputs aligned with user needs, collaboration with customers, and understanding desired outcomes.
Deep dives
Advancements in LLMs and Customization
The future of LLMs is likely to see advancements in customization and personalization. Rather than relying on a single large model like GPT-4, the trend will be towards tailoring models to specific use cases and fine-tuning them to achieve higher accuracy. This will involve exploring different sizes of models, context lengths, and understanding the trade-offs between larger models and smaller, more focused models. The availability of more affordable hardware will allow for hosting and training local models, enabling individuals and organizations to have more control and customization over their models.
Democratization of Data Recipes
The democratization of data recipes is another trend that may come to fruition in the next six to nine months. This involves making the process of curating high-quality training data sets more accessible so that open-source models can perform as well as proprietary models. Contributions from the open-source community and advancements in tools and techniques for data curation and labeling will drive this trend.
Demand for Private LLMs
There is a growing demand for private LLMs, particularly among enterprise customers. These customers value their data and want to deploy models in-house without sharing their intellectual property. The ability to host private LLMs in a secure and efficient manner will be a key focus in the coming months.
Evaluating LLMs
Evaluating LLMs can be challenging, especially when dealing with free-form text outputs. While standard metrics exist for certain tasks like classification, they may not be suitable for evaluating outputs involving free-form text. Instead, focusing on actionable outputs that align with user needs and objectives can provide a more effective evaluation approach. Collaboration with customers and understanding their desired outcomes is crucial for defining evaluation criteria.
Implications of Long Context Models
The emergence of long context models, such as those with larger context windows like 100,000 tokens, brings new possibilities. These models allow for increased customization and may alleviate the need for extensive fine-tuning. The potential for reduced training requirements and dynamic adaptation through longer context lengths offers a fascinating avenue for exploration.
Gateway Approach and Orchestration Layers
The use of AI gateways and orchestration layers may become more prevalent in the future. These tools can help manage different workloads and route tasks to specific models based on their requirements. This approach provides flexibility in choosing the appropriate model and context window for specific use cases.
Challenges of Evaluation in Text-to-Sequel Models
Evaluating text-to-sequel models involves dealing with both free-form text outputs and the need for accurate SQL conversions. While classification tasks can have clearer evaluation metrics, measuring the quality of free-form text can be more challenging, and comparing SQL outputs requires specific context and considerations. Evaluations may vary depending on the desired outputs and actions required from the models.
Laurel Orr is a Principal Engineer at Numbers Station, a startup that applies Foundation Model technology to the enterprise data stack. Venky Orr is SVP of Product & Engineering.
MLOps podcast #196 with Numbers Station's Venky Ganti SVP, Product & Engineering and Principal Engineer, Laurel Orr, LLMs in Focus: From One-Size Fits All to Verticalized Solutions.
// Abstract
Dive into the realm of large language models (LLMs) as we explore the merits and limitations of 'one-size fits all' LLMs, and their role in data analytics. Through customer stories, we showcase real-world applications and contrast general LLMs with verticalized, enterprise-centric models. We address the significance of ownership structures, with a focus on open-source vs proprietary impacts on transparency and trustworthiness. Delving into the NSQL foundation models, we emphasize the importance of diverse, quality training data, especially with enterprise challenges. Lastly, we speculate on the future of LLMs, highlighting hosting solutions and the evolution towards specialized challenges.
// Bio
Laurel Orr
Laurel Orr is a Principal Engineer at Numbers Station, a startup that applies Foundation Model technology to the enterprise data stack. Her research interests include how to use FMs to solve classically hard data-wrangling tasks and how to put FM technology into deployment. Before Numbers Station, Laurel was a postdoc at Stanford advised by Chris Re as part of the Hazy Research Labs working in the intersection of AI and data management. She graduated with a PhD in database systems from the University of Washington.
Venky Orr
Venky brings over two decades of experience in software engineering and technical leadership to Numbers Station as SVP of Product & Engineering. Most recently, he served as General Manager leading several initiatives on query understanding and commerce in the ads product area at Google. Before that, he was CEO and co-founder of Mesh Dynamics, the API test automation company, which was acquired by Google in 2021. Prior to Mesh Dynamics, Venky was CTO and co-founder of Alation, the enterprise data catalog company, where he led technology and helped create the new data catalog product category.
// MLOps Jobs board
https://mlops.pallet.xyz/jobs
// MLOps Swag/Merch
https://mlops-community.myshopify.com/
// Related Links
Website: https://www.numbersstation.ai/
--------------- ✌️Connect With Us ✌️ -------------
Join our slack community: https://go.mlops.community/slack
Follow us on Twitter: @mlopscommunity
Sign up for the next meetup: https://go.mlops.community/register
Catch all episodes, blogs, newsletters, and more: https://mlops.community/
Connect with Demetrios on LinkedIn: https://www.linkedin.com/in/dpbrinkm/
Connect with Laurel on LinkedIn: https://www.linkedin.com/in/laurel-orr/
Connect with Venky on LinkedIn: https://www.linkedin.com/in/venky-ganti-2679a2/
Timestamps:
[00:00] Venky's and Laurel's preferred coffee
[00:36] Takeaways
[03:15] Please like, share, and subscribe to our MLOps channels!
[04:38] Venky's background
[07:47] Laurel's at background
[09:38] Data wrangling
[13:45] Sequel query
[19:25] One size-fits-all LLMs vs Verticalized and Specific LLMs
[23:42] Model Choice Trade-offs
[30:18] NSQL Foundational Models
[37:26] LLM Trends in 12 Months
[40:09] Data recipes being democratized
[45:16] Claude and 100,000 Context
[48:02] Exploring Varieties of LLMs
[50:02] AI Gateway
[51:07] Text-to-SQL Model Evaluation
[54:00] Wrap up
Get the Snipd podcast app
Unlock the knowledge in podcasts with the podcast player of the future.
AI-powered podcast player
Listen to all your favourite podcasts with AI-powered features
Discover highlights
Listen to the best highlights from the podcasts you love and dive into the full episode
Save any moment
Hear something you like? Tap your headphones to save it with AI-generated key takeaways
Share & Export
Send highlights to Twitter, WhatsApp or export them to Notion, Readwise & more
AI-powered podcast player
Listen to all your favourite podcasts with AI-powered features
Discover highlights
Listen to the best highlights from the podcasts you love and dive into the full episode