How to think different: community-fuelled innovation

MLOps.WTF Edition #3

Ahoy there 🚢​,

Matt Squire here, CTO and co-founder of Fuzzy Labs, and this is the 3rd edition of MLOps.WTF, a fortnightly newsletter where I discuss topics in Machine Learning, AI, and MLOps.

On how to think different

In 1975 two renegades put out an invitation asking similarly-minded enthusiasts to spend an evening together in a garage in Menlo park, California. Nothing weird, mind you, just a meetup for people who enjoyed building their own computers. Although come to think of it, in 1975 that likely was pretty weird.

The idea, which came from Fred Moore and Gordon French (a political activist, and computer engineer, respectively), was simple: a meeting where enthusiasts could exchange ideas, share knowledge, and talk about their hobby projects. It would be strictly non-commercial.

What motivated the Homebrew originators was nothing more than passion, curiosity, a love for technology, and a desire to see it democratised. These are the values which we can imagine drew the attention of one enthusiast in particular, Steve Wozniak, who attended the first meeting after receiving the flyer from a friend (Allen Baum).

This is where Apple was born. On that first night at the Homebrew Computer Club, in Gordon’s garage, Steve Wozniak got the inspiration for a computer design that would ultimately become the Apple I. As Wozniak built his prototype, Steve Jobs began joining him at the club where together they demonstrated, iterated, and refined the concept.

We can only imagine the conversations that must have taken place. People openly sharing hard technical problems and solutions, debating the best way to get such-and-such microchips to work well together, angry arguments about whether it's possible to have "too many Steves"?

It was a time where most things were still to be figured out. Perhaps where we are today with production AI is similar – particularly the generative flavour.

Last week our team at Fuzzy Labs hosted a hackathon for large language models. We gathered 10 teams, each with their own data and their own idea, and we spent the day building things. We did it to get some feedback on tooling that we've been using internally. But we also wanted to create an environment for shared learning and discussion about tools more broadly.

And in that respect, the hackathon was a huge success! Not only did we get some invaluable feedback on the tooling, we were blown away by the innovation and ingenuity in the ideas of the attendees. It really felt like we had kicked off a positive feedback loop from which some incredible things will emerge, ultimately helping to make AI easier to productionise.

There’s something puzzling about the MLOps tooling landscape: on the one hand there are frankly far too many tools and products competing for attention. But on the other hand, it doesn’t feel like an ecosystem yet. In a software ecosystem, there are standards, or at least widely accepted conventions, and generally each tool fills a specific need, complementing the others.

Suppose you want to write a data pipeline. You’ll discover a spectrum from tools that purely concern themselves with the definition of your pipeline, but leave how to run the pipeline to something else (my preferred approach!), to tools that blend the two, right through to ones that will also train your model, deploy it, and make your coffee too for some reason.

In DevOps, things look simpler: we have a small number of basic non-overlapping concepts that everybody agrees on: like CI/CD, infrastructure as code, containerisation, immutable deployments. And for each of these we have a small number of very mature tools that tend to work nicely together (and I can’t help noting that open source dominates every category here).

It’s not a surprise that things look the way they do right now; DevOps looked a bit like this too, at one time. So how do we move forward?

Put tools to one side: what are the fundamental challenges that we face today when we build software applications around machine learning? (After all, that’s why we productionise ML, because we want to build something on top of it). What can the various people involved in building these systems — data scientists, software engineers, infrastructure engineers, etc — learn from one another?

The path from the Homebrew Computer Club, through ubiquitous personal computing, to cloud computing and now to the nascent MLOps revolution has been long, and the tech scene today is unrecognisably different. But the ethos that allowed that small group of enthusiasts to change the world is just as powerful today. Getting like-minded people together to swap ideas and engage in passionate discussion remains vital in generating out-sized impact. 

Let’s take our cue from them, and place community and collaborative thinking first.

And finally

Assorted things of interest

A great thing about the internet is how often you stumble across little fun facts that really stick in your head. One such fact I encountered this week was from Max Fagin (https://x.com/MaxFagin): “It takes 2 hours to orbit at the surface of any object made of rock”.

Whether it’s the size of an asteroid, the moon or a planet, if it’s a chunk of rock, the length of an orbit is two hours. Its size doesn’t matter, only its density. 

We can use this fact to finally answer the question: is the moon made of cheese? To test it out for yourself, just jump in a rock and get into lunar orbit. If orbit takes you two hours, the moon is made of rock, if it takes 3 hours 15 minutes, it’s cheese (specifically gouda).

Thanks for reading!

Matt

About Matt

Matt Squire is a human being, programmer, and tech nerd who likes AI and MLOps. Matt enjoys unusual programming languages, dabbling with hardware, and computer science esoterica. Matt doesn’t have a garage but if he did it would be full of whiskey drinking nerds brandishing soldering irons. He’s the CTO and co-founder of Fuzzy Labs, an MLOps company based in the UK.

Each edition of the MLOps.WTF newsletter is a deep dive into a certain topic relating to productionising machine learning. If you’d like to suggest a topic, drop us an email!