TWIML has been going since 2016, which in this field is close to archaeological, and that longevity is the most interesting thing about it. Sam Charrington was interviewing machine learning practitioners when the industry was still arguing about whether deep learning was a fad, and the archive now constitutes a fairly complete oral history of how applied ML actually developed. If you want to understand why the tooling and the practices look the way they do, working backwards through that catalogue is genuinely instructive. The editorial focus is on practitioners more than on frontier research, and that is the show's real distinction.
Interviews with people running ML platforms at large companies, MLOps engineers, applied scientists, and researchers whose work is close to deployment. Discussions about feature stores, model monitoring, data pipelines and organisational structure are less glamorous than capability speculation and considerably more relevant to what most people in the field spend their days on. Charrington's interviewing style is patient and clarifying. When a guest uses jargon he asks them to define it, which keeps episodes accessible to listeners who are not specialists in that subfield.
Episodes run forty five to seventy five minutes, which is a sane length that fits a commute, and the show notes with paper links are properly maintained. Topic indexing on the site makes the archive searchable in a way that most podcast back catalogues are not. The criticisms are real. The interviewing is friendly.
Guests are not pushed hard, claims about a product or an approach tend to be accepted at face value, and there is little of the productive friction that makes Machine Learning Street Talk worth the effort. If you want an interviewer who will challenge a premise, this is not it. That gentleness makes the show comfortable and slightly less useful. Quality varies episode to episode more than most shows, largely tracking the guest.
Some are excellent, some are essentially a company describing its platform, and the titles do not always make clear which you are getting. Sponsorship and event promotion are present throughout. TWIML runs conferences and courses and the podcast is part of that business, which is a legitimate arrangement and does mean some episodes sit close to the commercial side of the operation. Nothing egregious, and worth noticing.
The format constrains depth. Machine learning is mathematical and audio is a poor medium for mathematics, so technical discussions stay at the level of concepts and intuitions. That is fine for the show's purpose and it does mean you should not expect to come away with anything you could implement. In a crowded field, recent episodes feel less distinctive than the earlier ones did.
When TWIML started there was very little else. Now there is Latent Space for the AI engineering scene, Machine Learning Street Talk for research depth, and Dwarkesh for the strategic view, and TWIML's applied middle ground is less obviously unique than it was. It remains the best of the applied ML podcasts and it no longer stands alone. My three point nine reflects consistent, useful, well produced content over a long period, marked down for gentle interviewing and for a commercial layer that occasionally shapes the editorial.
Search the archive by whatever topic you are working on and you will usually find something worth an hour.