913: LLM Pre-Training and Post-Training 101, with Julien Launay

Hugo Dozois-Caouette speaks to Jon Krohn about his startup MaintainX and how he secured $100 million in venture capital. MaintainX manages and maintains computerized maintenance management systems (CMSs), or work-execution software, for the industrial and manufacturing industries. This “digitized version of a clipboard” with the help of web and mobile applications, provide a list of procedures, guidelines and regulations to help increase worker productivity and give a company the data-driven insights it needs to refine its processes. Listen to the episode to hear Hugo’s thoughts on the gaps in the manufacturing industry that technology can fill, the tech stack used by MaintainX, and the discrepancy of information in manufacturing environments. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/922⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

Using Windows for AI development and the bleeding edge of NPUs: Shirish Gupta and Ish Shah from Dell Technologies speak to Jon Krohn about the latest products from Dell, the future of neural-processing units (NPUs), and how AI developers can make sound hardware investments.  This episode is brought to you by the Trainium2, the latest AI chip from AWS, by ODSC, the Open Data Science Conference and by Gurobi. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/921⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (04:18) Why Windows still outranks other operating systems (20:58) The difference between GPUs and NPUs (32:44) How to access and use Dell’s NPUs and GPUs (49:08) Using processing units on the cloud versus locally (57:43) About the Dell Pro Max

This month’s episode of In Case You Missed It gives us reasons to be cautiously optimistic about the future of large language models (LLMs), with guests discussing what to do about recent reports that found AI agents blackmailed human users when threatened, the importance of post-training LLMs, and the training we have available for data and AI engineers to create robust, secure, and useful AI. Jon Krohn includes clips from his interviews with Akshay Agrawal (Episode 911), Julien Launay (Episode 913), Michelle Yi (Episode 915), and Kirill Eremenko (Episode 917).  Additional materials: ⁠⁠⁠⁠⁠www.superdatascience.com/920⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

PyTorch, AGI, and the future of alignment research: Aurélien Géron joins Jon Krohn in this live interview to talk about the fourth edition of his bestselling Hands-On Machine Learning as well as what superintelligence makes him hopeful for, as well as what concerns him about machines surpassing human intelligence. This episode is brought to you by Gurobi and by the Dell AI Factory with NVIDIA⁠ Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/919⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (02:04) Why Aurélien wrote Hands-On Machine Learning        (20:54) How Aurélien came to decide on material for the new edition  (28:53) Aurélien’s predictions for AGI                                       (51:21) How to support alignment research                    (1:13:42) Does superintelligence mean super-capability

In this Five-Minute Friday, Jon Krohn introduces listeners to CrewAI, an open-source Python framework that can create and manage multi-agent teams. The clue is in the title: CrewAI assembles specialized agents into single “crews” that achieve complex goals between them. CrewAI’s agent teams can also learn and iterate, meaning that after the crew has achieved its goals for the first time, they can refine and tailor their approach to  future goals.  Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/918⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

Founder of SuperDataScience, Kirill Eremenko, talks to Jon Krohn about how he found the best tools and approaches to help launch his 8-week AI engineering bootcamp. He breaks down the topics participants cover each week, and he also shares his tips with listeners who might want to start their own tech bootcamp or sign up for SuperDataScience’s September 2025 cohort. This episode is brought to you by the Dell AI Factory with NVIDIA and by ODSC, the Open Data Science Conference Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/917⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (10:58) Weeks 1-4 of the SuperDataScience bootcamp             (37:52) How to use AI to drive the bottom line in business                    (47:50) Weeks 5-8 of the SuperDataScience bootcamp (54:50) How to convert LLMs to agents (1:09:33) Jon’s feedback on the SuperDataSciencebootcamp

GPT-5 has just been released, but with not very much fanfare. In this Five-Minute Friday, Jon Krohn asks if GPT-5 deserves the community’s underwhelmed response to its release. He outlines five features of the model and explains why people might be feeling less than enthusiastic in the broader context of LLM development. Which LLMs are leading the way, and which are still playing the game of catch-up? Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/916⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

Tech leader, investor, and Generationship cofounder Michelle Yi talks to Jon Krohn about finding ways to trust and secure AI systems, the methods that hackers use to jailbreak code, and what users can do to build their own trustworthy AI systems. Learn all about “red teaming” and how tech teams can handle other key technical terms like data poisoning, prompt stealing, jailbreaking and slop squatting.  This episode is brought to you by ⁠Trainium2, the latest AI chip from AWS⁠ and by the ⁠Dell AI Factory with NVIDIA⁠. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/915⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (03:31) What “trustworthy AI” means      (31:15) How to build trustworthy AI systems  (46:55) About Michelle’s “sorry bench”   (48:13) How LLMs help construct causal graphs   (51:45) About Generationship

In this Five-Minute Friday, Cofounder and CTO of lakeFS Oz Katz talks to Jon Krohn about data warehouses, data lakes, and how companies can handle increasingly complex data infrastructures and formats. Hear about lakeFS’s collaboration with Legofest, lakeFS’s approach to helping users collaborate on data lakes, and how to overcome the challenges of working with multimodal data. Additional materials: ⁠www.superdatascience.com/914⁠ This episode is brought to you by the ⁠Dell AI Factory with NVIDIA⁠.

Julien Launay launched Adaptive to give data science teams in business enterprises their “RLOps tooling” to make reinforcement learning easier. Talking to Jon Krohn, Julien says, “Most of our users are data scientists who write Python codes to interface with the system”. Adaptive is also able to work with companies without data science teams, collaborating with partners like Deloitte to add the necessary personnel. Julien is currently working on making his platform more widely available. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/913⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.