Lukas Biewald
137 episodes listed below
PRESS PLAY
Join Lukas Biewald on Gradient Dissent, an AI-focused podcast brought to you by Weights & Biases. Dive into fascinating conversations with industry giants from NVIDIA, Meta, Google, Lyft, OpenAI, and more. Explore the cutting-edge of AI and learn the intricacies of bringing models into production.
Based on 5 episodes we listened to · September 2026
Detailed interviews connect AI research claims to the costs, failures, measurements, and infrastructure involved in making models work.
Lukas Biewald keeps Gradient Dissent: Conversations on AI close to the machinery of building and deploying models. These interviews move from neural-network training and pruning to open-source LLM development, production monitoring, evaluation, and fine-tuning.
A production feature goes bad; Lukas Biewald asks how it happens, which is probably your question too.
Press play if you want
Short reviews from the PlayNext desk, based on the episodes we processed.
A focused ClickHouse conversation about open-source infrastructure, technical selling, and AI-era adoption.
Aaron Katz frames enterprise infrastructure as a bet on timing, tracing ClickHouse from a company launched around a $50 million seed round, before product, customers, revenue, or pitch deck, into a managed service for an open-source database.
A grounded look at open-source model training, where research ambition runs into hardware, checkpoints, and full hard drives.
Open-source language models are treated here less as a headline topic than as a practical research problem. Stella Biedermann frames Aluthor.ai as a grassroots collective built to support open-source AI research, starting from a Discord server for people interested in G2P3.
A grounded look at turning base models into usable systems and judging whether the gains are real.
Scaling language models comes across as a set of practical system choices, not a simple race toward a larger base model. Aidan Gomez, Cohere's CEO and co-founder and one of the authors of Attention Is All You Need, explains how instruction data can turn a base model into a more…
Jonathan Frankle connects pruning research to the practical discipline of training models efficiently.
Jonathan Frankle treats neural network training less like magic and more like a disciplined engineering problem. As chief scientist at Mosaic ML and a soon-to-be Harvard professor, he connects his Lottery Ticket Hypothesis work on pruning to a broader question: how can teams…
A grounded look at why production ML gets harder as models, features, and teams change.
Production machine learning often gets messier before a model reaches its final production check, and this episode gives that mess a practical shape.
Daily TED and TEDx talks hosted by Elise Hu, spanning science, psychology, and creativity.
A wound-down internet-culture show now feeding listeners into Alex Goldman's and Emmanuel Dzotsi's new podcasts.
This interview explores the creation and significance of the Helm benchmark for evaluating large language models, featuring Together AI co-founder Percy Liang.
This episode features an interview with Carina Hong, CEO of Axiom, discussing her company's $64M-funded AI system that excels at mathematical reasoning and verification, including its performance on challenging tests like the Putnam.
This episode features an interview with Corey Sanders, SVP of product at CoreWeave, discussing how the company's specialized AI cloud infrastructure challenges big tech by optimizing GPU performance for machine learning workloads.
This episode features an interview with Edwin Chen, CEO of Surge, discussing his company's rapid growth in the human data collection business and its critical role in building AI systems.
This episode features a conversation with AI researcher and entrepreneur Mike Knoop discussing the R1 and R10 models from China, their similarities to OpenAI's O1, and insights into the ARC-AGI benchmark.
This episode features an interview with Sequoia partner David Cahn discussing AI industry topics including his article on AI's $600 billion question, recent news on DeepSeek and Stargate, and how his religious perspective influences his views on AI.
Clara Shih, CEO of Salesforce AI, discusses her role in unifying AI initiatives at Salesforce, including building integrated systems like the Einstein trust layer and data cloud to enhance business applications.
This episode features an interview with Jake Heller, CEO of Casetext, discussing how his company harnessed AI and GPT technology to innovate legal practice, leading to its $650 million acquisition by Thomson Reuters.
This episode features an interview with Sergey Nastrenko, CEO of Quilter AI, discussing how his company uses reinforcement learning to automate PCB design in electronics.
This interview explores the intersection of AI and healthcare with John Halamka of Mayo Clinic Platform, delving into his background and perspectives on technology's evolution in medicine.
This episode features an interview with Codeium CEO Varun Mohan discussing the future of AI in coding, including how current models generate meaningful code, infrastructure challenges, and lessons from the autonomous vehicle industry.
This episode explores Imbue's vision for AI agents that can execute arbitrary code to give software engineering capabilities to everyone, discussing practical research challenges like the gap between benchmark and real-world LLM performance.
This episode features an interview with Paul Copplestone, co-founder and CEO of Supabase, discussing the company's open-source platform for builders, its database tools including Postgres and PGVector, and its role in the evolving AI and LLM technology stack.
This interview with Perplexity AI's CTO Denis Yarats explores how the company builds high-quality generative AI applications, focusing on product success, team structure, and combining search engines with large language models for fast, accurate answers.
This episode features an interview with Soumith Chintala, co-creator of PyTorch, discussing the framework's origins, its role in machine learning, and open-source community management.
This episode features an interview with Chris Van Pelt, co-founder of Weeds and Bices, discussing how to identify early users and avoid over-customizing products in enterprise machine learning.
Host Lucas B.
This episode features an interview with Jerry Liu, CEO of LlamaIndex, discussing how their framework helps connect large language models to organizational data for building practical AI applications.
This episode features an interview with Joe Spisak, Director of Product Management of Generative AI at Meta, discussing his career journey and current work in generative AI and its applications in science and material innovation.
This episode explores how the U.S.
Jean Marc explains why machine learning isn't always the solution, using the hammer-and-nail analogy to caution against automation hype.
This episode features an interview with Atlassian CEO Mike Cannon-Brooks discussing how AI tools are transforming business processes and developer productivity at his $42B software company.
This episode features an interview with Joe Mora, CEO of Crew AI, discussing how AI agents are transforming enterprise operations and the platform's role in managing thousands of agents for tasks like automation, coding, and support.
This episode features an interview with Stephen Balaban, CEO of Lambda Labs, discussing his company's journey from startup to a $1.2B business, their role in the AI ecosystem selling GPUs and cloud services, and the impact of GPUs on data centers and electricity grids.
This episode features a conversation with AI investors Sarah Guo and Elad Gil about how they stay informed on LLM research and use it to guide their investment decisions in AI startups.
This episode explores why physical AI applications require a completely new data stack, featuring insights from Nico West, founder of Rerun.ai, on robotics, visualization challenges, and the current state of AI in real-world applications.
This episode features an interview with the CEO of Runway ML discussing how their latest video AI model achieved top performance on industry leaderboards through advanced simulation capabilities and training on observational data.
This episode explores the importance of notebooks in AI development through an interview with Mareo CEO Akshay Agrawal, focusing on how interactive computing environments enhance data work and machine learning.
This episode features an interview with Tuhin Srivastava, CEO of Base10, discussing the challenges and opportunities in AI inference infrastructure, model deployment strategies, and his company's journey to becoming a fast-growing player in the machine learning space.
This episode features an interview with Brandon Duderstadt, co-founder and CEO of Nomic AI, discussing the origins and mission of GPT for All as an open-source project aimed at increasing transparency in large language models.
This episode features an interview with AI researcher Joseph E.
This episode features an interview with Arvind Jain, CEO of Glean, discussing how his company uses LLMs to create an enterprise AI platform that functions like ChatGPT for internal company data and knowledge.
This episode features an interview with Julian Green, founder of Brightband, discussing his journey as an entrepreneur in deep tech and how AI is being used to improve weather and climate forecasting.
This episode features an interview with Jonathan Siddharth, CEO of Turing, discussing his company's role in providing code as training data for LLMs and its impact on advancing AGI.
This episode features an interview with Snowflake CEO Sridhar Ramaswamy discussing his career journey from Google engineer to executive and Snowflake's AI strategy, including enterprise LLM use cases.
This episode features an interview with the CTO and Chief AI Officer of Isomorphic Labs discussing how AI and deep learning are transforming the drug discovery process.
GitHub CEO Thomas Dohmke discusses GitHub Copilot, the future of software development, and shares advice on successful acquisitions based on his experience.
This episode features an interview with Erik Bernhardsson, CEO of Modal Labs, discussing the company's infrastructure platform for machine learning and AI applications.
This episode features an interview with Rodrigo Liang, CEO of SambaNova Systems, discussing his company's custom AI hardware solutions for enterprise applications and the story behind its founding.
Host Lucas B.
This episode features an interview with Pinecone founder and CEO Edo Liberty discussing vector databases, their role in machine learning and large language models, and how they function as a new type of database for AI applications.
The host interviews Cerebras Systems CEO Andrew Feldman about their announcement of the world's fastest, most accurate, and cheapest AI inference solution using their wafer-sized chips for machine learning workloads.
This episode features an interview with Martin Shkreli discussing his journey from pharmaceutical controversies to developing AI applications in finance and drug discovery, along with insights on his jail experience and entrepreneurial strategies.
This interview explores the origins and development of Cursor, an AI-powered coding tool, through a conversation with co-founder Sualeh Asif about its inspiration from scaling laws and language models.
This episode features an interview with Christopher Ahlberg, CEO of Recorded Future, discussing how AI is applied to cybersecurity and threat intelligence by analyzing internet data to detect global threats.
This episode features an interview with Stanford professor Chelsea Finn discussing the current capabilities and limitations of robotics, focusing on how robots learn through data collection and play.
This episode features an interview with DeepL CEO Jarek Kutylowski about how his company uses AI to disrupt the translation industry, focusing on business applications and competing with major players like OpenAI.
How are AI chips made like cookies from dough, starting as big wafers?
Harrison Chase reveals why retrieval failures are the biggest mistake in building LLM-powered applications.
["Aaron Katz says ClickHouse raised a $50 million seed round in August 2021 with no pitch deck, no product, no customers, and no revenue.", "He traces the company back to Alexey Milovidov at Yandex, where ClickHouse was built for petabyte-scale streaming data and later open-sourced in 2016.", "Katz says early pull came from Uber, eBay, Disney, Comcast, Deutsche Bank, Anthropic, OpenAI, and Weights & Biases, and that ClickHouse chose an engineer-led motion instead of enterprise-heavy selling.", "The conversation also covers Salesforce lessons, the Datadog migration, AI-assisted coding, and the $100 million Silicon Valley Bank wire that cleared before SVB froze."]
["Gradient Dissent focuses on EleutherAI's move from a July 2020 Discord server to releasing 1.3B, 2.7B, 6B, and 20B parameter language models after OpenAI kept GPT-3 closed.", "Stella Biederman says a CoreWeave GPU deal enabled the 20B run, and she links that release to a broader shift toward public models from Facebook, BigScience, NVIDIA, and Chinese groups.", "She explains how to choose among open models, from local 6B-to-20B runs on A6000 or A40 hardware to instruction-tuned systems like T0, MT0, and NovelAI fine-tunes.", "The episode also goes into The Pile, 96 A100 GPUs, 400 checkpoints, and the scaling-law result that later matched a DeepMind estimate of about 20 data points per parameter."]
Aidan Gomez tells Lucas Biewald that fine-tuning is a "mature system feature" you use after deployment, optimization, and every other lever is exhausted.
Jonathan Frankel tells Lucas Biewald why the lottery ticket hypothesis says a sparse subnetwork can be reset to its original weights and still train well.
["Lucas Biewald talks with Shreya Shankar about her paper Operationalizing Machine Learning in an Interview Study and why production ML is a different job from model research.", "Shreya says the study interviewed around 20 practitioners whose models were already serving outputs in production and grouped the work into experimentation, evaluation and deployment, monitoring and response, and data collection.", "She says velocity, early validation, and versioning matter because teams do not want to discover a bad model only after final A/B testing or a production alert.", "The conversation turns to retraining cadence, data corruption, feature maintenance, Jupyter notebooks, and a more realistic view of reproducibility in ML systems."]
An interview with Jasper AI's CEO and head of AI about how marketers use their text generation product and the technical details of large language models.
An interview with venture capitalist Sarah Catanzaro discussing the hype and reality of AI, drawing on her technical background and investment experience.
An interview with Chris Valenzuela, CEO of Runway, discussing AI-powered video editing and the role of machine learning in creative work.
Lukas Biewald interviews Jordan Fisher, CEO of Standard AI, about building computer vision-powered autonomous checkout for retail stores.
Lukas Biewald interviews Stefan Fabel about NVIDIA's Base Command platform for scheduling AI jobs on supercomputers.