1026: OpenAI’s GPT-6 Astra

In Episode #1026, Jon Krohn breaks down GPT-6 Astra, OpenAI’s new flagship that its president has floated as a possible marker of AGI. Jon covers what the model is, what it costs, its state-of-the-art results across computer use, coding, abstract reasoning and science and the safety story, which for this release is unusually intertwined with capability. He weighs the AGI claim against Anthropic’s Fable 5.1 and lands, as ever, in a measured middle. Additional materials:⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/1026⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (00:14) What GPT-6 Astra is, and what it costs (03:58) The capability highlights that matter most (10:23) The safety story and the AGI question

In Episode #1025, Dr. Luis Serrano (Founder of Serrano Academy) joins Jon Krohn to explain the paper he co-authored on the curved spacetime of transformer architectures, in which attention stops being a lookup table and becomes something closer to gravity: words bend the space around them, and the embedding of "bank" visibly curves toward "river" as it travels through the layers of the network. In this episode, he recreates Eddington’s 1919 eclipse experiment inside a transformer, draws the line between an LLM workflow and an actual agent, explains why agent evaluation is a step harder than evaluating an essay, and gives the cleanest account of GRPO you will hear. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/1025⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (00:10:53) What changed, and what survived, between the two editions of Grokking Machine Learning (00:27:48) Word gravity: how attention pulls "bank" toward "river" (00:42:12) Why RAG is an LLM workflow rather than an agent (00:51:23) The two-by-two that explains why GRPO powers reasoning models

In ICYMI Episode #1024, Jon Krohn tracks the gap between AI investment and AI return, from the technology side to the people side. Hear from Pete Johnson, Jerry Yurchisin, Priyanka Vergadia and Tristan Handy, discussing why four out of five organizations have the structures for AI success in place while only one in five sees the returns, which decisions should never be handed to a language model however confident it sounds, how to structure Claude skills so that your output stops being slop and why the semantic layer matters more, not less, now that analytics agents are the ones asking the questions. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/1024⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (00:56) Vector Search, Agentic Memory and Effective RAG (09:20) Mathematical Optimization in the Agentic AI Era (17:30) Anyone Can Write Code Now, So What Gets You Hired? (27:14) How dbt Won Analytics Engineering

In Episode #1023, Aishwarya Srinivasan (Co-Founder of The Gen Academy) joins Jon Krohn to work out where a competitive moat comes from once anything you can build in ten minutes, somebody else can build in ten minutes too. Ash came to teaching through Illuminate AI, the mentorship community she started in 2020, and now trains senior engineers and leaders to ship agentic AI in production; she is blunt that vibe coding lowers the floor without touching the engineering judgment that production demands. In this episode, she explains what a whole-system eval covers that a model eval misses, traces reinforcement learning from the algorithm she patented at IBM to its resurgence in agentic fine tuning and lays out the MIND framework from her TED Talk for living with AI. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/1023⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (00:10:10) Why cheap code shifts the software engineering job rather than ending it (00:15:50) What a whole-system eval covers that a model eval misses (00:36:11) Why reinforcement learning came roaring back for agentic AI (00:41:23) The one skill Ash says matters more than any hard skill

In Episode #1022, Jon Krohn tackles the art of steering AI agents, deciding where your instructions should live so they get followed reliably without bloating every request. A sequel to Episode #1020 (where model size and effort set an agent’s horsepower), this one is about direction: the seven ways to deliver instructions, why a hook beats a prompt, the industry-wide agents.md standard, and three practical takeaways you can apply whatever your stack. Additional materials:⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/1022⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (02:52) The seven ways to deliver instructions to an agent (06:42) Why a hook is a guarantee and an instruction is only a probability (13:00) Three takeaways for organizing your instructions

In Episode #1021, Tristan Handy (Founder and CEO of dbt Labs) joins Jon Krohn to explain how a study of about a hundred companies in 2016 became analytics engineering, and then became a tool that over a hundred thousand data teams rely on. Tristan coined the term, chose SQL when Spark was the fashionable answer, and spent a decade turning down acquisition offers because none of them were good for the people using dbt. He is now merging dbt Labs with Fivetran and taking on the presidency of the combined company, the first deal he says cleared that bar. In this episode, Tristan walks through what dbt does to your raw data, argues that the semantic layer matters more once analytics agents are asking the questions, explains the type safety behind the Fusion engine, and details how a 12-kilobyte skill file collapses a million-dollar migration into six weeks. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/1021⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (00:07:22) Why Tristan chose SQL over Spark, and what progressive complexity means (00:10:44) How a dbt project turns raw data into modeled tables (00:17:50) Why a decade of acquisition offers kept failing his one test (00:40:47) How 12-kilobyte skill files cut year-long migrations to six weeks

In Episode #1020, Jon Krohn unpacks the two dials that increasingly decide what you get out of a large language model: which model size you pick and how much effort you tell it to spend. Using a July Anthropic blog post by Claude Code’s Lydia Holly as a jumping-off point, with guidance that generalizes to any model family, Jon explains what each setting actually does under the hood. Model size swaps which frozen weights handle your request (roughly, how capable), while effort sets how thorough and certain the model must be before calling a task done, not a simple “thinking-time slider.” He offers a clean diagnostic for when to raise effort versus move to a bigger model, shows why cheaper-per-token isn’t always cheaper-per-task and surveys how OpenAI, Google and open-weight labs have all converged on these same two dials. Additional materials:⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/1020⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (00:56) What the model-size dial actually does (05:29) Why effort isn’t a thinking-time slider (13:25) Three practical takeaways for using both dials

In Episode #1019, Priyanka Vergadia (founder of The Cloud Girl, former Senior Director of AI Transformation at Microsoft and Head of North America Developer Relations at Google) joins Jon Krohn to explain why almost every company has bought AI tools and almost none of them are seeing a return. Her fix is a budget split that will make any CFO wince: seven dollars on training employees for every dollar spent on the tools themselves. Having spent a decade turning dense cloud and AI concepts into sketches that a quarter-million developers actually remember, and having carried GitHub Copilot into Fortune 100 boardrooms, she has watched the gap between tool purchase and real production use up close. In this episode, Priyanka defines the elusive quality she calls taste, walks through how she structures Claude skills so her output stops being slop, unpacks her 10-20-70 framework, and shares breaking news about what she is building next. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/1019⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (00:10:39) What “taste” actually means and why Priyanka now interviews for it (00:31:39) How to build a Claude skill by breaking a task into explicit sub-tasks (00:36:11) The 10-20-70 framework for AI budgets (00:47:52) The weekend exercise for finding what makes you different

In Episode #1018, Jon Krohn breaks down Qwen3.8-Max, Alibaba’s enormous new flagship, a 2.4-trillion-parameter mixture-of-experts model that, if its promised weights ship, becomes the largest open-weight release in history. Landing just weeks after Moonshot’s Kimi K3, it extends the price war and the open-weight surge Jon covered in Episode #1012. Alibaba positions it as second only to Anthropic’s Claude Fable 5 / Mythos 5 and independent signals land in a similar neighborhood. Jon walks through its capabilities and multi-day agentic demos, its aggressive pricing ($2 in / $6 out per million tokens, with cached input eight times cheaper), and the question he gets asked most: are Chinese models safe to use? His answer hinges far less on the model than on how your data reach it. Additional materials:⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/1018⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

In Episode #1017, Pete Johnson (Field CTO of AI at MongoDB) joins Jon Krohn to explain why four out of five organizations have AI steering committees and success metrics, yet only one in five sees a return on the investment. Having made nineteen stops across six countries this year advising more than a hundred companies on their AI strategies, Pete has an unusually wide view of what is actually working in production. In this episode, he traces the history of SQL and denormalization, unpacks why the embedding model is the most underrated choice in a RAG pipeline, explains Matryoshka embeddings and lays out what better agentic memory looks like. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/1017⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (00:06:34) Why the AI ROI gap happens and what to do differently (00:18:21) Jevons paradox, bank tellers and toll booth workers (00:24:05) From Codd’s 1970 paper to denormalization (00:32:31) Why the embedding model is not a commodity (00:40:20) What better agentic memory looks like