1004: Recursive Self-Improvement

Could an AI get good enough at AI research to build its own, more capable successor and kick off a compounding loop? That’s recursive self-improvement (RSI) and it surged into the conversation after Anthropic revealed that, as of May 2026, Claude wrote more than 80% of the code merged into its production codebase. In this Five-Minute Friday, Jon Krohn separates today’s AI-assisted coding from true RSI, walks through the accelerating evidence - METR’s shrinking task “time horizon,” Google DeepMind’s AlphaEvolve, Andrej Karpathy’s overnight training-tuner, weighs Jack Clark’s 60% bet that AI builds its own successor by 2028 against the compute, data and “marketing” skeptics. As ever, Jon lands in the optimistic middle. Additional materials:⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/1004⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

Frank Basso, VP of Infrastructure at Lightning AI, joins Jon Krohn for a rare ground-level tour of the one layer of the AI stack the show had never covered in over a thousand episodes: the physical data center. Frank explains how Lightning AI provisions its 35,000-plus GPUs through hyperscale co-location, why everything new is liquid-to-chip cooled, how GPUs talk to each other over ultra-fast east-west networks, and what it’s actually like to stand inside a 110-decibel AI data hall. He also debunks the most persistent myths about data-center water and electricity use, and makes the case for fuel cells, nuclear power, and 800-volt DC distribution as the path forward. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/1003⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (02:47) What actually makes an AI data center different from a traditional one (06:04) How Lightning AI provisions its 35,000+ GPUs through hyperscale co-location (24:01) Why liquid cooling doesn’t waste water, debunking the biggest data-center myth (29:46) East-west vs. north-south networks, explained (43:47) “Screaming banshees”: why AI data halls run at 105–110 decibels (51:52) Why data centers don’t actually drive up your power bill

Anthropic’s Claude Fable 5 was the most capable AI model ever released to the public and it lasted just three days before the US government forced it offline. Jon Krohn unpacks both halves of the story: what makes Fable 5 special, and why it was pulled. Fable 5 and its locked-down sibling Mythos 5 are the same model separated only by safeguards, in a new “Mythos-class” tier above Opus. Jon covers its state-of-the-art benchmarks, premium $10/$50-per-million-token pricing, conservative safety classifiers, and the federal export-control directive, reportedly sparked by an Amazon-flagged “jailbreak” that took it down. Additional materials:⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/1002⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

For this episode #1001 special, the tables are turned: SuperDataScience founder Kirill Eremenko takes the host’s chair and Jon Krohn is the guest. They trace Jon Krohn’s path from an Oxford neuroscience PhD to a New York hedge fund to founding the AI consulting firm Y Carrot, why he regrets leaving academia and how tools like Claude Code erased his hard-won technical moat and why that makes skilled engineers more valuable than ever. Along the way: whether AI is a bubble, Jevons paradox and the data-center boom, the RICE framework for choosing AI projects, the single biggest reason AI projects fail and how a well-built AI agent could give anyone “Christopher Nolan–like” focus. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/1001⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (03:42) From an Oxford neuroscience PhD to AI consulting (17:25) Defining AGI and why consciousness isn’t required (30:39) Are we in an AI bubble? Why we benefit either way (46:32) Jevons paradox: why cheaper AI means more data centers (01:08:31) The RICE framework for prioritizing AI projects (01:15:08) The number-one reason AI projects fail in production (01:31:50) AI, attention, and protecting your wellbeing

For this landmark 1,000th episode and the show’s 10-year anniversary, host Jon Krohn is joined by SuperDataScience founder Kirill Eremenko, who hosted the podcast for its first 400-plus episodes before handing over the reins. In a first for the show, the episode was recorded live with the audience invited to join on air, alongside surprise appearances from the team, longtime guests, and even Jon’s family. Together, Jon Krohn and Kirill look back on a decade of the podcast and field listener questions on AI’s biggest opportunities, the build-versus-buy dilemma, how to break into the field today, and how to stay grounded amid the relentless pace of AI. Additional materials:⁠ ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/1000⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

Chip Huyen joins host Jon Krohn for this milestone episode 999 to talk about her record-breaking book "AI Engineering" the most-read title on the O'Reilly platform last year and how the AI landscape has shifted since her last appearance. Chip breaks down what separates AI engineering from machine learning engineering, makes the case for a "start simple" workflow, gets candid about the real costs of running LLMs in production, and shares why she's now fascinated by physical AI, robotics, and world models and why the durable problems worth solving are increasingly human ones. Jon Krohn guides the conversation from the practical content of the book through to where the field is heading next. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/999⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (06:48) What separates AI engineering from machine learning engineering (14:44) The “start simple” approach: prompting, then RAG, then fine-tuning (18:19) Why web search is so painfully expensive in production (35:11) Is the “ChatGPT moment” for physical AI really here? (52:21) Why the durable problems left to solve are people problems

In this month’s episode of ICYMI, Jon Krohn explores how AI agents are simultaneously creating new risks and unlocking powerful new ways of working with data. Hear from Anneka Gupta, Cal Al-Dhubaib, Trevor Manz, Jazmia Henry, Jeremy Mumford, and Jacob Miller, discussing why the old cybersecurity playbook breaks down in the age of Claude Mythos, how the notebook became an AI agent’s working memory, what it really takes to build a foundation model from scratch, and why failing slowly is the most expensive mistake an AI team can make. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/998⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (00:40) Why Claude Mythos Changes Everything About Cybersecurity (08:11) Why Your Notebook Should Be Your Agent’s Working Memory (13:19) What It Actually Takes to Build a Foundation Model From Scratch (20:46) Failing Slowly Is the Most Expensive AI Mistake

Dr. Andrey Kurenkov returns to the show to talk about Astrocade's astronomical growth from pre-alpha to over 20 million engaged users, what it actually takes to build a vibe-coding platform that scales, and how the broader AI landscape has shifted since his last appearance. Andrey shares behind-the-scenes lessons from building B2C user-generated content products, why the real moat is community rather than tech, and his current thinking on humanoid robotics, AGI, and the AI risks people actually overlook. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/997⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (02:11) The Astrocade elevator pitch and how it grew to 20M users (16:19) Why there's no secret sauce behind the platform (24:56) UGC as the real moat, not the AI (46:57) Why household humanoid robots are now 2–3 years away (58:33) What AGI actually means, and why Andrey is an ASI skeptic

TrueFoundry co-founder and CEO Nikunj Bajaj speaks to Jon Krohn about how enterprises like Nvidia and Siemens are realizing returns of over $100 million from single agent deployments, the AI gateway architecture that makes it possible to connect, observe, and govern agents at scale, and why the familiar advice to “start small” is the wrong way to roll out AI agents inside a large organization. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/996 Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.⁠⁠⁠ In this episode you will learn: (01:21) What TrueFoundry does and why agents in production need a control plane (06:32) Breaking down the AI gateway: the model, MCP, and agent gateways (16:47) Taming tool sprawl with scoped, read-only MCP access (19:10) Why the agent gateway is the hard part and the kill switch most teams lack (22:24) The five-workflow framework behind $100M agent deployments

Jazmia Henry joins Jon Krohn to break down what it actually takes to build end-to-end foundation models for the energy industry. From wrangling decades of handwritten oil-and-gas documents into usable training data, to bespoke tokenizers, reinforcement learning, and inference at scale, Jazmia walks through every stage of the stack. Along the way she explains why reinforcement learning models are "bursty," what reward hacking is and how her Grounded Continuous Evaluation framework fixes it, and revisits the 2023 NeurIPS paper that argued, to widespread skepticism at the time, that scaling bad data degrades model performance. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.superdatascience.com/995⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (10:06) The User Agnosticism Tenet (20:02) The Zillow Offers parable (23:25) Why workflows should come before agents (29:57) Why data engineering is the bedrock of AI (52:41) Why velocity is the only durable moat