Your co-founder Dillon spoke to us last year. I see a lot has happened at Paperspace since then. Can you tell us about that?
The past year has been an incredibly exciting time for our company. We’ve made some major announcements over the past two months alone. In September, we enabled our Gradient Machine Learning (ML) platform as a service (PaaS) to run in enterprise data centers, hybrid could, and multi-cloud environments. It was already one of the first ML PaaS offerings to allow end-to-end lifecycle management of ML models, and includes everything from developing, training, tuning, and deploying those models. The first version ran within our infrastructure and Google Cloud Platform (GCP). This latest iteration launched in September now gives customers the choice of cloud and on-prem environments.
More recently, we launched a free Cloud GPU service for Machine/Deep Learning development on our Cloud Computing and Deep Learning platform. We designed it for students and professionals who need to learn how to build, train, and deploy ML models. We also added an ML Showcase which is a curated list of interactive ML projects that can be forked, similar to a GitHub repo. We consider it a solid ML/DL starter kit that can help anyone develop and expand their skills, and try out new ideas without any financial barrier. This new offering has exploded in popularity since its launch.
MLOps seems to be gaining momentum. I see you’re at the forefront of this movement. What’s prompting this shift?
There are many reasons, but what it really boils down to is the fact that ML in the enterprise is slow, tough to scale, lacks automation, collaboration is difficult, and the actual operationalized models delivering business value are few and far between. That’s why we need good MLOps that are designed to standardize and streamline the lifecycle of ML in production.
If you think about it, MLOps is the logical reaction to the current challenges enterprises are facing when putting ML into production. In software engineering we have DevOps, so why not MLOps? Good DevOps ensures that the software development lifecycle is efficient, well documented, and easy to troubleshoot. It’s about time we developed a similar set of standards for ML.
What is your role in the MLOps movement?
Entreprises are maneuvering to become more agile, nimble and responsive to remain competitive in today’s fast-changing marketplace. Shockingly, only 25 percent of a Data Scientist’s time is spent training models — the rest of their time is spent managing infrastructure and tooling. This has created a need for Machine Learning developers to move faster than ever and for enterprises to decrease time to market.
Simultaneously, enterprises are increasingly opting for ‘Multi-Cloud’ alternatives. They require enhanced development capabilities with greater flexibility across multiple cloud and on-prem environments to suit their specific workflow needs. We responded to that demand with our new iteration of Gradient which offers agile Machine Learning (ML) tooling and methodology across multi-cloud, on-premise, and hybrid environments — without the need for DevOps or any manual configuration.
Who in the organization owns MLOps?
If given a platform that fully abstracts infrastructure and provides robust reproducibility and determinism, we believe ML Engineers can operate without a dependence on SREs or software teams to handle tasks like managing a Kubernetes cluster, ingesting data, pipelining, etc.
When ML teams can operate with full autonomy and own the entire stack, they are much more efficient and agile. So our position is that the ML Org should own MLOps. Often that is now led by a Head of ML, of sorts.
How would you describe the ML/AI developer toolset at this point? What are the challenges/opportunities?
I think it’s fair to say that there are some leaders in the Deep Learning technologies space, but a true stack hasn’t emerged in the traditional sense. There isn’t a LAMP stack equivalent or go-to tooling. In terms of frameworks, TensorFlow may have jumped out to an early lead, but recently PyTorch has come out of nowhere and is now right there with them.
Choosing which tools to use is still being defined — and that’s definitely one of the biggest challengers in our space. Looking back a bit, PyTorch was not even on the radar when we first got into the ML universe, and then out of nowhere it blew up.
Another tool that has exploded in popularity is Jupyter. We made an early bet that Jupyter would become the de facto IDE for ML and that bet has really paid off. I’ve said previously there’s a Cambrian explosion of tools right now and I still believe that’s true.
What are your top predictions for AI in 2020?
I believe MLOps will see big growth in 2020, and will become just as established as DevOps has as a movement. Developer tools will also continue to become easier to use with better front-ends and more infrastructure abstraction. We will also experience continuous delivery, and I’m confident deployment will gain more ground, especially in AI development. In terms of AI chips and hardware, I believe both will continue to de-commoditize, and there will be more and more specialized chips. My final prediction is that in 2020, the Chief AI Officer position will start to gain ground as a needed role in the enterprise.
Thank you, Daniel! That was fun and hope to see you back on AiThority soon.
Daniel is a Co-Founder of Paperspace, building the future of Deep Learning in the cloud.