A clear case for replaying real events instead of relying on neat train-test splits.
The episode makes a strong case for online learning by setting it against batch train-test splits, which it treats as a weak stand-in for the real world.

Patient conversations about Python in practice.
The Python Podcast.__init__ treats Python as a working ecosystem, not just a language. Its conversations move from project origins into architecture, production tradeoffs, data workflows, and the friction teams meet when tools leave the notebook.
PRESS PLAY
The Python Podcast.__init__ approaches Python as a practical engineering world made of libraries, maintainers, workflows, and hard production choices. The format is interview-centered and patient. Instead of racing through announcements, it gives guests enough space to explain how a tool came together and why its design matters. Recent episodes show a strong focus on machine learning and data systems. River becomes a way to talk about models that learn from live event streams, replay user interactions, and evaluate new model versions against the order of real history. Predibase opens a discussion about declarative machine learning, data-to-deployment platforms, serverless training, REST APIs, batch prediction, explainability, monitoring, and community integrations. Deepchecks leads into a careful question about validation: what correctness means when outcomes are probabilistic and the acceptable boundary depends on the domain. Baseten brings the conversation closer to application work, with user-provided Python code, isolated execution environments, local developer workflows, and drag-and-drop interfaces for data scientists. The show also keeps room for classic open source tooling. PyScaffold is discussed through its command line interface, internal APIs, nested project structure, action pipeline, extensions, and changing Python packaging practices. Polyastro broadens the scope again, connecting Python to astrodynamics, orbital mechanics, simulations, visualization, and a developer's path from Fortran 90 to scientific Python. Across these conversations, the show repeatedly asks how abstractions behave when real teams use them. It cares about the implementation, but also about the user experience wrapped around that implementation. The tone is technical without being theatrical. It is measured, specific, and often more interested in the second-order consequence than the headline feature. That makes it useful for engineers, data practitioners, and technically curious Python users who want to understand why a tool exists, how it is shaped internally, and where it fits in the wider ecosystem. The result is less a news feed than a working map of Python's applied software culture.
Made for: For Python developers, data engineers, machine learning practitioners, and technically curious software builders who like architecture-level conversations. It also suits listeners who want open source context before adopting a library or workflow.
What sets it apart: The show stands out by pressing past feature descriptions into implementation choices, developer experience, and production consequences. It treats Python tools as systems shaped by maintainers, users, and real organizational constraints.
The podcast about Python and the people who make it great
Based on 6 episodes we listened to · September 2026
Tobias Macy makes Python tooling concrete by following each project's architecture, workflow, and tradeoffs from initial idea to working system.
Tobias Macy turns conversations about Python projects into careful tours of the engineering decisions underneath them. His questions follow software from its governing idea into implementation: River's replayable event logs for streaming machine learning, Predibase's path from…
A click log becomes a time machine for your next model; you get to follow the wiring.
Press play if you want
Short reviews from the PlayNext desk, based on the episodes we processed.
A clear case for replaying real events instead of relying on neat train-test splits.
The episode makes a strong case for online learning by setting it against batch train-test splits, which it treats as a weak stand-in for the real world.
A practical look at how ML platforms try to cover the path from data to deployment.
The episode presents the platform as more than a model tool, with Tobias Macy and Travis Adair focused on the practical work of making machine learning usable inside an organization, from data intake through deployment.
A grounded look at making model validation more trustworthy.
The episode looks at how to validate a model well enough to trust it. Shir Kharev and Philip Tanner use DeepChecks as a starting point, but the real point is broader: a good score is not the same as doing the job well.
A practical look at Base10 and the path from local Python code to a running ML app.
Tobias Macy and Tuhin Srivastav keep the focus on Base10 and the practical path it offers for data scientists building ML applications. The conversation works best when it traces how Python code moves from a local workflow into a running app, with isolated environments,…
A practical look at how PyScaffold makes Python project setup repeatable and extensible.
PyScaffold gets a grounded, practical showcase as a Python project template generator built to spare developers the repeated grind of setup work.
A focused review of Flora's checkpointed replay approach, its side effect model, and the overhead limits that shape faster ML iteration.
Machine-learning teams often pay for the same work twice when training runs have to be repeated, and this episode follows one attempt to avoid that waste.
A Python interview about Beware, black swans, PyScript, and everyday utility.
Russell Keith-Magee gives the episode its center of gravity. The conversation moves through the Beware project, Python's black swans, and PyScript's place in making the language useful in new settings.






Power Platform news with practical static.
Tobias Macy hosts Russell Keith-Magee to catch up on the Beware project, which aims to make Python viable for mobile and GUI development.
Tobias Macy interviews Rolando Garcia, a UC Berkeley PhD candidate, about FLOR, a Python tool for automatic checkpointing in machine learning training.
An interview with Angela Lin and Jeremy Shee about EvalML, an AutoML library for building and optimizing machine learning pipelines in Python.
Interview with Michael Shalonka about Gephira and developing applications for Kubernetes environments.
Interview with Ronak Rahman about automating the path to production to maintain development velocity.
Interview with Natan Yellen about Robusta, a tool for debugging applications on Kubernetes.
Interview with the creators of MushroomRL, a Python library for reinforcement learning.
Max Halford describes River, a Python toolkit for streaming and online machine learning, as the counterpart to scikit-learn for batch workflows.
Travis Adair says Predibase grew out of Uber’s ML work, Ludwig, and Horovod, where teams needed a declarative layer for production models.
Tobias Macy interviews Philip Tanner and Shir Kharev about DeepChecks, a Python package for validating machine-learning models and data with minimal effort.
["Tobias Macy interviews Tuhin Srivastav, Base10's CEO and co-founder, on podcast.init's Python machine-learning series.", "Srivastav says Base10 exists because his Gumroad team could build fraud-detection and content-moderation models, but still needed 3 to 4 weeks and full-stack help to productionize them.", "He argues that ML teams often get stuck after model serving because they still need APIs, backend logic, and UI layers for human-in-the-loop workflows.", "The conversation covers Python SDK deployment, Kubernetes-backed execution, sandboxed user code, templates for common use cases, and why Base10 chose to lean into code rather than hide it."]
Florian Wilhelm says PyScaffold began in 2014 after a Redmine wiki about `setup.py` grew to 20-30 pages and stopped being usable.
Interview with Juan Luis Cano Rodriguez about Polyastro, an open source Python library for astrodynamics and orbital mechanics.
An interview with Travis Adair about building and training machine learning models using Ludwig and Horovod, with a focus on Python.
Interview with Ben Stancil about data frustrations and improving the data experience, hosted by Tobias Macy.