Will AI Kill Programming?
Press enter or click to view image in full size
Scale By the Bay is famous for its no holds barred, provocative, deep tech debate panels.
The tradition started in 2017 with the Functional Programming for Machine Learning panel (above), and continued at the 2019 Who Needs Serverless panel.
This year, all three panels at Scale By the Bay 2020 are debates!
AI changed our industry already in many ways, big and small. As Carlos Guestrin said, every application is a Machine Learning application. The end-to-end data pipelines we had that terminated in some customer action are more and more ending in an AI model making a decision about the customer interaction. Data Science is on the rise in every company, and even though 90% of it is not AI but rather Data Wrangling, or perspiration, it comes with its own set of attitudes and tools.
Generally speaking, data scientist are not software engineers first. But they have to work with the software engineers to feed them data. And if the company is “webscale”, you need big data, and increasingly streaming, realtime data to operate your AI. Data scientists use Python while the rest of the stack is usually Java or Go, or even C++, Rust, Scala, Haskell, and so on. There’s a significant impedance mismatch between what data scientists do and what software engineers do. Just some of the higlights:
Software Engineers
- Strongly typed, compiled programming languages
- Distributed systems, clusters, reactive streams, actors, message buses, immutable logs
- APIs, REST, GraphQL
- Test-Driven Development
- IDEs, cloud builds, docker, Kubernetes, cloud deployments
- DevOps and MLOps clearly understood
- Service Mesh and other ways to do the cloud right are welcome
- Oncall
Data Scientists
- Python or R for modeling
- Model deployment as a third-party service
- Laptop-bound, a variety of scale-up options without a standard
- Notebooks for sharing findings and collaboration, not in production
- Using APIs or files for data access in training
- Hard to handshake for data in production, translation errors for model invocation by software engineers
- Software engineers are across the fence, DevOps and MLOps are underground
- Oncall when the model loses money
Since data science drives the business metrics and has a way to analyze them, the question often becomes, how much programming and devops can you eliminate in a cloud setup to run your AI business as AI-first? Will programmers be further commoditized, as glorified plumbers? Will Nocode allow data scientists to replace them with autogenerated ETL? Will AI evolve to write the boilerplate code to feed itself data and actuate its decisions? Come to our panel to find out!
