OpenAI DevDay

For many folks it’s Thanksgiving Week, so it seems like a great time to say thanks to all the sponsors of MLOps Community that support us and help us do what we do.

And we’re pleased to extend that thanks to SAS, who we’re excited to announce are the latest supporters of the community!

So, while we’re doing podcasts, IRL events, newsletters, virtual meetups, learning platforms and more, they’ll be helping you dig into your data, find cool insights, and make sense of it all with their AI and analytics platform.

I wonder if it’ll be able to help me analyze the Black Friday deals…

Building Defensible AI Apps // Gregory Kamradt // MLOps Podcast #191

So, you may have noticed there’s been a few things happening over at OpenAI recently.

Now, I’m not saying it’s because they heard this podcast I did with Greg about their DevDay.

All I’m saying is, this podcast was dropped on Friday and so was Sam.

At the time of writing, he seems to be having the last laugh, joining Microsoft. Though, the number of turns this story’s had, that could have changed three more times since you opened this email.

On a serious note, I hope the situation settles and works out out for everyone as nobody likes uncertainty.

So, to this episode, Greg and the DevDay. I’m sure you’ll know how much there was to talk about from it, but if not it includes:

Context Length Increase
New Modalities
GPT Assistance
OpenAI Marketplace
GPT Four Vision via API

Like everyone else on the internet we shared what we thought was special, but it was really interesting to see where we had different views…
It was also fascinating seeing his product mindset in action too. From having an idea and how he runs with it, to then listening to him talk about how the OpenAI Market place might develop and impact that process.

He also gave a sneak peek into a cool experiment he’s done. I can’t give too much away, but it involves the new context token limit and how valuable each part of the context window is.

It’s always fascinating listening to Greg, but just be warned, you may feel the need to start reading Paul Graham's essays.

Video || Spotify || Apple

Enterprises Using MLOps, the Changing LLM Landscape, MLOps Pipelines // Chris Van Pelt // Podcast #192

It may be my musical side coming out but when Chris mentioned his first company, CrowdFlower, I thought it would make a great name for a band.

And then he talked about demos and security restricting access, and I just visions of him and Lucas out touring on the road!

Alas, we’ll never get to hear their debut album, but you can hear Chris tell me about setting up Weights & Biases, it’s evolution and becoming CISO.

So, the ‘security restricting access’ comes from one of the security tips he shares for building products, as well as building for personas and lessons he learned as a start up.

And the ‘demo’ is Demo Driven Development, a great idea where they go out to customers with a demo of a minimum viable product, see what the interest is with customers, and respond to their feedback.

So, going out to customers, that’s almost like going on tour, isn’t it?

Video || Spotify || Apple

Start exploring Adala, a robust framework for implementing data agents, from the makers of Label Studio. Adala uses LLMs to continuously automate data processing tasks in the MLops pipeline, including data labeling and generation, with a unique focus on reliability and quality.

Adala is a modular framework that can learn various skills like image classification. It supports multiple runtimes (LLMs), with the underlying principle that it’s guided by human feedback, whether from a ground truth dataset created in Label Studio or other inputs in your data pipeline.

While the technology is early, at HumanSignal, we believe agents have incredible potential and represent the next great frontier in software development and AI operationalization. Adala is free and open source. If you’re interested in the project, explore the GitHub repo and join the Discord to share your feedback!

Put LLM Models in Production Faster with Hopsworks // Jim Dowling // IRL Meetup #54 Stockholm

It sometimes feels that ML engineers love a pipeline as much as Mario, but if it starts to get confusing, Jim Dowling might be able to help.

He talks about Hopsworks, and how the unified architecture helps break down the ML pipeline into three components. Think of it like finding the Warp Pipe location at the end of World 1, Level 2 in Super Mario Bros. Having the three areas, feature pipeline, training pipeline, and inference pipeline, makes it easier to collaborate between data teams, ops teams, and data scientists.

He also shares how Hopsworks provides a feature store, offers a model registry, and other features.

I wonder if one of them is a Fire Flower?

Watch it here

Model Blind Spot Discovery for Better Models // Pavol Bielik, CTO of LatticeFlow and David Garnitz, CTO of VectorFlow // MLOps Mini Summit #2

It's not even Black Friday, and we've already got a great deal for you, two CTOs for the price of one!

First up is Pavol Bielik, who talks about how to identify blind spots in deep learning models to avoid systematic mistakes, particularly with vision models. He shares specific examples like medical and satellite imagery, and also discusses the importance of evaluating custom hypotheses to understand issues.

Next up is David Garnitz who discusses the challenges of ingesting large volumes of unstructured data and the absence of a one-size-fits-all approach, which makes experimentation a necessity.

A real bargain to enjoy over the holiday season!

Watch it here

💡Job of the week

Senior MLOps Engineer // Entrupy (New York based)

As a Senior Engineer, you will work on our unified MLOps platform and associated projects to help us build the world's best authentication engine for high value goods.
Some project areas include:

Infrastructure and libraries to define, deploy, run and monitor training and inference jobs and providing interfaces and tooling for ML engineers to work with
Hybrid cloud server provisioning and automation
Platform advocacy, training and mentoring

Skills and Qualifications:

At least five years of software development experience.
Working knowledge of Python, and solid skills in at least one major backend language.
Production experience deploying and maintaining machine learning models.
Experience with data pipelines and job schedulers.
Generalist with experience in web apps, backend services, and databases.

The Case Against Vector Databases

Vectors are in vogue, there’s no doubt.

But, like wearing Crocs or joining in a video singalong to Imagine, it’s wise to remember that just because other people are doing it, doesn’t mean to say you should too.

Helpfully, Dariusz Semba has produced this gem, which gives a great breakdown of the arguments against them. It doesn’t say you shouldn’t ever use them, just gives some reasons why you shouldn’t get swept along in the vector hype.

Many thanks to Médéric Hurier for the contribution.