Transforming Model Calls into Smart Pipelines

DSPy: Transforming Language Model Calls into Smart Pipelines // Omar Khattab // MLOps Podcast #194

So, what is a Smart Pipeline? No, it’s not Mario with an iPhone.

It’s the DSPy framework, essentially a modular programming structure, designed to optimize the process of using large language models for specific tasks. So, task-independent modules that can be reused in a structured and automatically optimized way. And it’s smart because it doesn’t use a single model for all tasks.

It gets pretty full on, and I needed to ask him to slow down to make sure I was following, but it’s a great framework.

We also discuss the shift from handcrafted prompts to more structured approaches, the evolution and future of retrieval-based methods, the complexities of fine-tuning, and the importance of a structured approach in the design and use of language models.

It’s a lot to pack in, but it'll help you level up!

Video || Spotify || Apple

Generative AI has undoubtably been the hot topic of 2023. We started with the basics of “What every CEO should know” and by now, “The economic potential of generative AI” is clear. Many organizations have already launched pilots, and are now asking: what comes next?

Join us for a virtual 60-minute webinar on "Generative AI: Impact, Learnings, and the Path Forward." We will share examples of how real organizations have used AI, including generative AI tools, to transform domains and enable generative AI at scale. We’ll highlight key learnings and provide perspective on what 2024 holds.

Don't miss out on this unique opportunity to explore how organizations are using and scaling generative AI.

You can register here.

Shriyash Upadhyay from Martian // Lars Maaløe from Corti // Pietro Gagliano Transitional Forms// Moderated by Paul van der Boor and Sandeep Bakshi from Prosus

It can sometimes be difficult to get up and out of bed on these deep, dark December mornings.

You might need some motivation to start the day. Well, here it is!
What could be more inspiring than listening to the advice and experience of the founders of three companies?

This Fireside Chat from the LLM in Production III Conference in October is a great place to start. From the importance of sticking with a unique vision to engaging with the venture capital community and seeking competitive advice to gain insights and sanity checks for business ideas it gives a fantastic insight into developing a startup.

Just be sure to remember us when you become a tech billionaire after selling the startup you created after being inspired from this episode.

Watch it here

MLOps Roundtable on 6 December brought to you with Weights and Biases.

Managing models ain't easy.

Add in regulations, it gets harder.

Some regulations may be a while off, like what ever the successor to the Algorithmic Accountability Act may be in the US, or the EU’s AI Act, but they’ll be coming. And when they do, we’ll have to adapt.

One way to be ready is to look at how models have been managed when dealing with external regulations for specific sectors such as finance.

Join Demetrios as he chats and poses questions to Oliver Chipperfield from m-kopa, Mark Huang from gradient.ai, Darek Kłeczek from W&B and Michelle Conway from Lloyds Banking Group.

Be sure to register here.

Fireside Chat // Joe Reis // IRL Meetup #56 Silicon Valley

It sometimes seems like we’re adrift in an ocean of data. There’s so much of it, how do you know what’s useful?

Data, data everywhere, and insights yet to link.

With this in mind, Joe (wearing a great t-shirt...) delves into the realm of modern data models, tackling the formidable task of extracting actionable insights from vast raw datasets. He anticipates a growing reliance on automation in data governance.

It also goes into the incorporation of expansive language models into corporate datasets, shedding light on the disparity between groundbreaking research and the practical reality of these models.

Plus, he shares his concerns about the rise of inadequately modeled datasets and the need for data governance and management.

Hopefully it’ll put the wind in your sails.

Watch it here!

💡Job of the week

Senior Machine Learning Engineer – Graph/NLP // Health Stealth Co. (Remote, US)

We aim to transform healthcare by shifting the emphasis from treating symptoms to addressing the underlying causes of chronic diseases.

Responsibilities will include extracting and utilizing multi-modal data to construct personal knowledge graphs of our users that will integrate with other graphs (e.g., our community and biometrics graphs) to power RAG, multi-modal ML models to support user/content curation and recommendations, and chat bots.

We're looking for:

Proven track record of continuous deployment of graph and NLP models to production.
Strong proficiency in Python programming and related graph and NLP frameworks.
Experience with labeled property graph databases such as Neo4j.
Familiarity with ontologies and taxonomies is a plus.

Why You Don’t Want to Use Your Data Warehouse as a Feature Store

This blog made me think about the old joke about, “It’s not a bug, it’s a feature.”

Only this time it would be, “It’s not a Feature Store, it’s a Data Warehouse.”

Not quite as catchy. And should probably be the other way around. But hopefully it shows that the blog digs in to the misconception that feature stores are just databases for ML.

It goes in to why relying on data warehouses as feature stores may cause limitations, particularly in supporting real-time ML use cases. Plus it details the complexities of real-time feature serving shares how feature platforms can outshine data warehouses to help you.

Who knows, with progress like this, we may go from the Feature Store the Comedy Store.

Doubtful though, and that’s on me, not feature stores.

With thanks to Vince Houdebine at Tecton for the contribution.

How to Configure VS Code for AI, ML and MLOps Development in Python

Ever heard of the mind trick athletes use called 'visualization technique'? It's like mentally picturing the perfect kick or throw to up your game. Well, to grasp at an analogy, for developers there’s Visual Studio Code.

But, just like how nailing the visualization technique needs a coach, getting the most out of Visual Studio Code needs some help too. So, think of this blog as your coach for VS Code, showing you the ropes on how to set it up for your groove.

We've got a cool list of extensions, specific settings, and shortcuts to make VS Code totally yours. Plus, there are some neat tips and tricks to zip through your code like a pro.

Once you’ve read it, you’ll be ready to do the MLOps version of slam-dunking that home-run in the end zone.

With thanks to Médéric Hurier for the contribution.