All the MLOps podcasts the last year were special.
I want to call out a few of them as I reflect on the year. It may be useful for you in case you were looking for something to do with all the free time you’ll have over the holidays.
2021 MLOps Podcast/Meetup Awards
TL;DR And the winners are
Best MLOps On Prem Convo – Chris Albon
Best Fundamentals Convo– Svet Penkov which led to #testing-ml
Most Underrated Convo – Alex Patry
Best Non Financial Advice – Sarah Catanzaro
Most Legendary Guest – D. Scully
Best Case Study – Leonard Aukea
Biggest Surprise – Ben Wilson
Easiest Interview of the Year – Erik Bernhardsson & Mike Del Balso
Most Entertaining – Jacopo Tagliabue
Chris Albon has become a bit legendary in the MLOps space. Mainly for his in-depth ML and software knowledge, sharings on his website, and shitposting on Twitter.
I found him incredibly easy to talk to. The level of transparency with which he talked to Neal and me about MLOps in Wikimedia was top-notch. Many of you may already know this, but it was news to me that Wikimedia does not use any software that is not open source. They also have a strict on-premise policy for all the storage and compute they consume. Inevitably this comes with an extra layer of added complexity when dealing with ML & data products.
Chris spoke to us candidly about what can go wrong when everything is DIY. He told us a story about the time he ordered some GPUs and then got a picture from the Wikimedia data center workers. The picture came with a note telling him it was the wrong size and couldn’t fit on the rack.
As for tools, Albon also explained what he would like to see as far as oss tooling for ML goes.
Svet is passionate about changing the narrative when it comes to ML failing. Why should we accept that software will fail?
In this conversation, we go back to the basics. Svet details the importance of testing ML and how no matter what the use case, or what the model, we cannot have the attitude of releasing something and then see what happens.
He draws a parallel to the aviation industry and talks about how there is no culture of releasing a new airplane and expecting things to break when it flys. On the contrary, there are heavy checks and regulations that need to be met so we can breathe easy next time we want to jet off to Bali for a yoga class.
For some reason, this one slipped under the radar of many. Alex Pantry came on the pod to talk with me and guest host Skylar Paine about the LinkedIn jobs recommended system. He explained the subtleties necessary for creating a model around such a touchy subject, and also dove into the many iterations of the project and how it evolved for the better over the past couple years.
Alex had more quotable moments in this episode than I have had all year. It was hard to choose just one as the best so ill leave you with a couple of my favorites.
- “AI engineers are like water, they go where there’s no friction”
- “Innovation happens when you empower people”
- “Every good MLE should be able to tell a story with the data”
Sarah is an investor at Amplify Partners, a VC firm that has been heavily investing in the ML tooling space. She also has the advantage of having a data science background so she is able to understand the landscape at a level that most VCs can’t articulate.
In this conversation, she walked us through how she sees the MLOps field and the needs that can be filled by tools. Sarah also makes the point there isa disconnect between the research ML circles and the industry ML circles. Neither camp is helping build bridges. The researchers’ arent working on real-world problems, yet the researchers can’t work on real-world problems if they don’t have access to industry data. Quite the conundrum.
You may know D. without ever realizing you know him. If you’ve read the ML test score or the high-interest credit card debt of machine learning you have read his papers. Yep, he is one of the authors of both of those legendary papers. (Word on the street has it he is working on the reliable machine learning book at this very moment).
In this conversation, David, Vishnu, and I were humbled by his ability to talk about what has changed since putting out those classic papers. We delve into a bunch of shit that I didn’t understand but I bet was cool if do understand.
I first came into contact with the MLOps work at Volvo Cars from Leonard’s “Zero to MLOps talk slides“. It caught my eye because most talks I have seen about MLOps don’t go into that level of detail. When he came on our meetup, he went two levels deeper. Not only does he explain the Volvo ML architecture, but he also gives the why behind it. Leonard takes us through the pros and cons of each part and gives us his vision for what he would like to see the platform evolve into.
I appreciate the transparency on this one. Not many companies are willing to be this candid and vocal about how they have set up their stack. Let’s start the trend!
Watch the full presentation on youtube.
Why was this a surprise? Especially if you know Ben Wilson, this should be no surprise. Well, I was approached by the publishing company of his book Machine Learning in Action asking if I would have him on the podcast. At first, I was apprehensive, but after reading a bit of his book I figured why not give it a shot. Boy am I glad I did.
I went into the conversation with pretty low expectations. I guess a part of me thought Ben would just shill his book and not offer much value. I could not have been further from the truth. This was one of the most technically in-depth conversations that I think we ever had without a full presentation equipped with architecture diagrams. Ben was able to explain the method behind the madness of MLOps in a way that was simultaneously in-depth and digestible.
I walked away from the conversation a huge fan and as you probably guessed by now, I bought his book. Ben has also since become one of the most insightful members of the community slack. I really appreciate this conversation and feel bad for the alternate universe where I said no to haveing him on the meetup.
Wait a minute, these two guests are heavy hitters in their own right, how could this have been the easiest interview? Well, they practically interviewed themselves. I got to grab some popcorn and watch. Mike and Erik dive into data vs ML platforms. Should they be separate things? The pair also go deep on the flaws that centralized and decentralized data platforms have while explaining the signs of each.
The award for most entertaining interview has got to go to Jacobo. I am not sure if it’s the Italian in him that gets him so passionate about the MLOps or if he is like this with everything he does. Whatever the case, I had a hard time not smiling when Vishnu and I chatted with him a few weeks ago.
Jacobo wanted to make one thing clear during the conversation, we are still very early days of ML and MLOps. He is immensely optimistic about where the field is going. It was a breath of fresh air to hear the hope for what this space can become and see the work he is doing to try and help us get there!
Nial Murphy – A legend in his own right. Can’t get on a call with him and not learn something.
Eugene Yan – RecSys deep dive in a way that only Eugene could do it. I love seeing the way he looks at the bigger picture and can extrapolate out design patterns.
Jeremy Howard – I mean, how could I not mention him? Fun fact: while we were interviewing him, I felt like absolute shiiiiiiiit. It was midnight and all I wanted to do was go to bed. I have to give all the credit to Vishnu for picking up the slack and delivering a great episode.