Researchers revisit their PhD theses and trace what changed since.
Researchers go back through their PhD theses with host Sean, then carry the ideas forward into current work on language models, machine translation, formal verification and large-scale training systems. The questions stay technical but leave room for how each guest got there, and several episodes end with advice for new PhD students.
PRESS PLAY
The Thesis Review picks an unusual starting point for a research interview. Instead of a guest's newest paper, the anchor is their PhD thesis. Sean, the host, treats that document as a fixed point and traces the line forward: what the guest set out to do, what the field did next, and what still holds. In some introductions he describes himself as a postdoctoral researcher at the University of Washington working on machine learning methods for generating and reasoning with natural language. That shows in the questions, which are specific and technical but give the guest room to tell how the work came about. The guest list reaches across computer science. Tianqi Chen, introduced as a pioneer where machine learning meets systems, recalls building an open-source framework around computational graphs, automatic differentiation and CUDA acceleration, and the work of scaling training to more devices. Niloofar Mireshghallah, a postdoctoral scholar at the University of Washington, takes up privacy and safety in large language models, including a 2020 paper by Nicholas Carlini and collaborators that extracted full sequences from GPT-2 by sampling heavily and re-ranking the results. Yulia Tsvetkov describes multilingual NLP moving from a few good translation systems for major languages toward dialects, language varieties and translation built for a particular population. Hady Elsahar credits prompting with turning language models from sentence completers into tools that can answer questions and do math. Swarat Chaudhuri argues that a neural network is a low-level program embedded in human-written code, which brings formal methods and property specification into play for systems like self-driving cars. The episode names alone show the spread, from 'Learning to Map Sentences to Logical Form' to 'Linguistic Knowledge in Data-Driven NLP' to 'Logics and Algorithms for Software Model Checking'. Comparing then with now is what holds each conversation together. Luke Zettlemoyer says dialogue drove the final part of his thesis even though, by his account, he wrote few if any papers on it, and that what context means for language understanding is still wide open. Tsvetkov admits she had to refresh her memory of her own dissertation. Mireshghallah says that talking about her work lets her step outside it and see it from a new angle, a remark Sean calls the thesis of the show itself. The mood is friendly. Sean tells Elsahar he is a big fan of his work and hears the same line back. Several episodes end on advice for people starting a PhD. Chen's version is that five years is long enough to try something bold. Zettlemoyer's is to think about how to be different and have an impact others could not, and Sean notes that no two researchers have answered the same way.
Made for: Suits a listener who follows machine learning and NLP research and wants the story behind a researcher's current work, starting from the dissertation. It also fits someone in a PhD, or weighing one, who wants to hear researchers look back on theirs and offer advice.
What sets it apart: Rather than starting from a guest's latest result, it anchors each conversation to the guest's PhD thesis and measures the present against it. Comparing then with now makes each interview a record of which ideas held up and which shifted.
Each episode of The Thesis Review is a conversation centered around a researcher's PhD thesis, giving insight into their history, revisiting older ideas, and providing a valuable perspective on how their research has evolved (or stayed the same) since.
Sean Willick brings a PhD thesis back to its author; you hear which questions survived.
Press play if you want
Short reviews from the PlayNext desk, based on the episodes we processed.
A focused technical history of XGBoost, MXNet, and TVM, grounded by a case for bold PhD work.
Tianqi Chen's PhD work gives this episode a clean path through one of machine learning's less glamorous but central problems: making training and deployment efficient enough to scale.
A thesis-led discussion of multilingual NLP, adaptive translation, and research goals anchored in curiosity.
Yulia Svedkov's PhD thesis gives the episode its through line, as the University of Washington computer science professor revisits the point where linguistics and data-driven NLP meet.
A thesis on logical form opens into a broader conversation about context, language models, and research choices.
Luke Settlemoyer's PhD thesis, Learning to Map Sentences to Logical Form, gives The Thesis Review a clean bridge from semantic parsing to today's language-model questions.
A clear research-led episode that keeps prompting's usefulness and its selection bias in view.
Hady El-Sahar's route from a 2019 PhD on natural language and structured knowledge bases to controlled language generation gives the episode a clear technical backbone.
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.
Tianqi Chen says a six-month CUDA attempt to train an ImageNet model in his master's work pushed him toward machine learning systems, not just modeling.
The Thesis Review episode with Sean Wellick centers on Niloufar Mireshghallah's UC San Diego PhD work on auditing and mitigating safety risks in large language models.
["Yulia Svedkov's 2016 CMU thesis, \"Linguistic Knowledge in Data-Driven Natural Language Processing,\" sits at the center of a Thesis Review interview recorded in late 2022.", "Sean Wellick talks with her about moving from software work in Israel into NLP research and later into the University of Washington Allen School.", "She describes a Swahili machine-translation problem with almost no resources, Arabic influence on Swahili, and lexical borrowing as a bridge when bilingual dictionaries were missing.", "She also discusses polyglot language models, computational ethics, controllable generation, and why static pre-trained models and narrow labels miss important problems."]
Sean Whalick interviews Luke Settlemoyer about his thesis, Learning to Map Sentences to Logical Form, and frames the conversation around how a University of Washington professor and Meta researcher moved from semantic parsing to large language models.
Hady El-Sahar argues that prompting is useful but noisy, and he contrasts it with distributional control for language models.
Sean Willick interviews Swarat Choudhury, an associate professor at the University of Texas, about his 2007 University of Pennsylvania thesis, Logics and Algorithms for Software Model Checking.
The Thesis Review host Sean Willick interviews Charles Sutton about his 2008 UMass Amherst PhD thesis, Efficient Training Methods for Conditional Random Fields.
The Thesis Review's [41] Talia Ringer - Proof Repair examines Ringer's 2021 University of Washington PhD thesis on maintaining proofs when verified software changes.
The Thesis Review’s episode with Lisa Lee examines her 2021 Carnegie Mellon thesis, Learning Embodied Agents with Scalably Supervised Reinforcement Learning.
Sean Willick interviews Burr Settles about his 2008 University of Wisconsin-Madison thesis, Curious Machines: Active Learning with Structured Instances, and his research leadership at Duolingo.
Host Sean Wellick interviews Martin Arjovsky about his PhD thesis on out-of-distribution generalization in machine learning, covering his research journey and key contributions like Wasserstein GAN and invariant risk minimization.
Host Sean Wellick interviews Chelsea Finn about her PhD thesis on meta-learning and the MAML algorithm for robot learning.