Building CLIP

Adventures in Building CLIP & Other (Largeish) LMs // MLOps Podcast #180 sponsored by Prem AI

A picture paints a thousand words, so they say.

I think they’re going to have to do a re-count if they’re using CLIP.

And leading up that count will be this week’s guest, Sachin Abeywardana from Canva. As well as his love of math, we talk about the challenges building CLIP. Some were for the whole team, but what struck me the most was how honest he was about the challenges he faced and how he’s developing.

Speaking of which, something I’d like to see CLIP process would be a picture of the lecture’s face when they hear Sachin say his PhD was a waste of time!

Video || Spotify || Apple

The Centralization of Power in AI // MLOps Podcast # 181

Open-mic nights. Open-top cars. Open-club sandwiches.

Somethings are just better open, right? Of course! Especially Open AI, right?

Well, Kyle Harrison from Contrary gets in to this and more when we talk about his report "The Openness of AI”. It’s a fascinating read that focuses on the dangers of centralization of power in AI, specifically in relation to OpenAI, including privacy concerns and trustworthiness with handling data.

Open source might not be so open and shut after all.

Video || Spotify || Apple

MLOps@GetYourGuide // MLOps Podcast #182

It can be a confusing and, the machine learning world.
What you need, is a guide.

I do my best to be your sherpa on this journey through the abstract, but sometimes you need something practical and real-world.

Well, this is it. Jean, Meghana, Olivia, and Theodore share their experience building a machine learning platform at, ironically, Get Your Guide.

This realistic view of what it takes to build and maintain such a system can be your guide when you feel you’re battling through your build.

Video || Spotify || Apple

Job of the week

Senior Machine Learning Engineer // Too Good To Go (Paris, France)

At Too Good To Go, we want to inspire and empower everyone to fight food waste together. We are looking for a Senior Machine Learning Engineer to be a part of our team that defines, builds and delivers the AI features of the Too Good To Go Platform.

Responsibilities Include:

Exploratory data analysis, machine learning model training and evaluation
ML models deployment, monitoring and management
Improvement of our ML tools and workflows in collaboration with the rest of the team

How does your AI model stack up in production?

Did you know that 87% of AI models never move from lab to production?

In fact, this is one of the biggest challenges faced by machine learning teams today. Just because a model excels in a test environment doesn't ensure its success in the real-world. Furthermore, as you deploy AI models to production, they often degrade over time, especially with incoming new data. So, how do you know what is causing your AI models to fail? And how do you fix these issues to improve model performance?

Join us on 8th November where we will delve deep into these issues and provide insights on how you and your ML teams can systematically identify and fix model and data issues. Empowered by intelligent workflows, our end-to-end AI platform is built by ML engineers to enhance your model's performance across the entire AI lifecycle, all while unleashing the full potential of robust and trustworthy AI at scale.

Happening November 8 - Click here to register for the event now.

How LinkedIn Is Using Embeddings to Up Its Match Game for Job Seekers

Co-Authors: Jacob Mannix and Shaobo Zhang

As embedding-based retrieval continues to challenge the tech industry, Linkedin managed to build a comprehensive set of new infrastructure components designed to help their teams deliver more relevant search results and recommendations. These new components will enable:

Authoring Composite and Multi-Task Learning Models:

The new infrastructure will ease the development of two tower models and enable multi-task learning for the embeddings.

Feature Cloud for Offline and Streaming Embedding Generation:

The infrastructure includes fully hosted platform called "Feature Cloud" that combines offline and streaming embedding generation, as well as preparing precomputed embedding vectors into feature stores and EBR indexes.

Automated Embedding Version Management:

Versioning for embeddings is no walk in the park, specially if numerous models or systems depend on a specific version of the trained embeddings. To tackle this challenge, the platform ensures the embeddings and all the dependencies are version controlled.

Model Cloud for Inference Graph Orchestration:

Linkedin model inference stack, "Model Cloud," has been extended to run inference graphs. This allows easy execution of inference graphs in a more serverless fashion.

If you're curious about how these tools work, give the article a read to drill down on each component.

Thanks to Mohamed Sadek for the contribution.