PRESS PLAY
Note: The TDS podcast's current run has ended. Researchers and business leaders at the forefront of the field unpack the most pressing questions around data science and AI.
Based on 6 episodes we listened to · September 2026
Careful interviews connect AI research methods to their practical uses, unresolved limits, and wider consequences.
Towards Data Science turns ambitious AI questions into careful interviews that keep moving between underlying research and practical consequences.
An H-shaped maze gives your AI power-seeking worry a surprisingly precise place to stand.
Press play if you want
Short reviews from the PlayNext desk, based on the episodes we processed.
A brisk finale connecting AI scaling, broader data capabilities, and the safety questions that follow greater capability.
The finale follows AI's widening ambitions through the logic of scaling, from GPT-3's recipe to DeepMind's Chinchilla revision and the belief that sustained investment can keep producing capability gains with no known ceiling.
A maze experiment gives AI power-seeking a concrete shape, with the largest real-world leap still left open.
AI scaling has made broadly human-level and superhuman systems feel less speculative, and this single-episode review follows the worry that greater intelligence may carry power-seeking pressure with it.
A resume generator reveals both the craft and the verification burden of working with GPT-3.
GPT-3 feels less like a spectacle here than a workable material for small software. Through a data scientist's resume and cover-letter generator, the conversation frames language models as tools that everyday users can shape when the problem is narrow enough and the prompts are…
A measured look at tiny ML sensors that process sensitive image data before sharing only a narrow result.
Towards Data Science turns to a quieter corner of AI: tiny ML systems that move machine learning onto edge devices. In conversation with Matt Stewart, a deep learning and tiny ML researcher at Harvard University, the review follows a proposal to bring sensing and computation…
A measured look at where deep networks resemble brains, and where the comparison still gives way.
Towards Data Science brings J.R. King into a clean but slippery question: if deep learning models and transformers can work with language, images, planning, and games, how close are they to biological intelligence?
A clear look at how data trust can turn into a security problem.
The episode stays anchored in a practical question: if someone buys data, how do they know the rows are real? That framing keeps the focus on trust, provenance, and the danger of depending on information that may be fabricated.
AI research, explained by the people doing it.
Talking shop with the people behind data products
Specialists walking through software's tradeoffs, one system at a time.
Founders and engineers on AI coding agents like Claude Code, and supply chain security.
David Hirko describes fuzzing data pipelines to find vulnerabilities that could be exploited for intentional data poisoning.
Jeremie Harris explains why he and Ed left Sharpest Minds after GPT-3 and founded an AI safety company with a senior leader from the U.S.
Towards Data Science interviews Ed Harris about AI safety research suggesting that advanced systems may seek power by default through convergent instrumental goals.
Towards Data Science examines how Amber Tang used GPT-3 and OpenAI Playground to build a practical cover-letter generator.
Towards Data Science host interviews Matt Stewart of Harvard University about ML sensors that perform machine learning beside the raw sensor instead of sending raw data to the cloud.
Towards Data Science host interviews J.R.
Ryan Fedasiuk says automated nuclear launch is the clearest line the United States and China should not cross.
Alex Watson says differential privacy lets you remove one record and get “the exact same answer” or a negligible difference.
Ala Shaabana and Jacob Steeves describe BitTensor as a way to put AI on the blockchain, where "digital intelligence" becomes the commodity instead of hashing power.
"Wait 15 more minutes" becomes the lesson after one woman arrives late and brings two others.
Alexey Baevski of Meta AI Research describes Data2Vec as a continuation of Wave2Vec, BERT, and self-supervised learning across speech, text, and vision.
Towards Data Science examines Andy Jones's independent AI safety work and the Scaling Hypothesis linking larger models, more data, and more compute to predictable performance gains.
Stanford professor Margot Gerritsen argues that deep neural networks are difficult to interrogate because their millions of learned manipulations do not expose clear reasons for decisions.
Ben Goertzel, an early AGI researcher, contrasts fixed reward optimization with open-ended intelligence and evolving goals.