A Post-Transformer Architecture Crushes Sudoku (Transformers Solve ~0%)

A game millions of people solve over morning coffee is exposing a fundamental weakness in today’s most powerful AI models. In this Five-Minute Friday, Jon Krohn breaks down Pathway’s new Sudoku Extreme benchmark, roughly 250,000 of the hardest Sudoku puzzles available and why leading LLMs like o3-mini, DeepSeek-R1, and Claude 3.7 Sonnet scored effectively zero percent, while Pathway’s post-transformer BDH architecture achieved 97.4% accuracy at a fraction of the cost. Listen to the episode to find out what BDH is doing differently, why Sudoku performance matters far beyond puzzles, and what this means for the future of AI reasoning. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/978⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

What’s going to be the next big step function that blasts us forward in AI capabilities? To find out, Jon Krohn sits down with Professor Kyunghyun Cho, whose 200,000 citations and co-authorship of the first paper on attention place him among the most influential AI researchers in the world. In this episode, Kyunghyun explains why today’s models have already captured most correlations in passive data, making the real challenge about actively choosing which data to collect. He also weighs in on the open debate around world models, whether AI needs high-fidelity, step-by-step imagination or whether a high-level latent representation that lets it skip ahead is sufficient and shares the surprising discovery that 80% of his 200 computer science students had never installed a coding agent. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/977⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (06:43) The story behind the attention mechanism (28:43) Sample efficiency and active data collection (39:04) World models and latent planning (49:52) Teaching undergrads with coding agents (58:21) Reranking, multi-stage ranking, and the foundations of RAG

NVIDIA just dropped Nemotron 3 Super, a 120-billion-parameter open-weight model that only activates 12 billion parameters at a time and it’s built for the agentic AI era. In this Five-Minute Friday, Jon Krohn breaks down the model’s hybrid Mamba-Transformer architecture, its million-token context window, and why its combination of frontier-class reasoning with blazing-fast throughput matters for anyone building multi-agent systems. Find out how Nemotron 3 Super claimed the #1 spot on the DeepResearch Bench leaderboards, which companies are already adopting it, and where you can start using it today. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/976⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

Zack Kass speaks to Jon Krohn about his bestselling, tech-positive book, The Next Renaissance, that charts the rapid progress of humanity and the benefits that artificial intelligence will bring to us, as well as why a future where intelligence is a cheap and abundant resource will give humanity an edge. Elsewhere in the show, Zack discusses why it’s important to hold parents, teachers and students accountable for their education, why it is incumbent on us to build a healthier relationship with technology, and his 4 principles for thriving in the age of AI. This episode is brought to you by the⁠ ⁠⁠Cisco⁠, by ⁠Acceldata⁠ and by ⁠⁠ODSC, the Open Data Science Conference⁠⁠. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/975⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (03:14) About Zack Kass’ book, The Next Renaissance (20:18) The importance of literacy skills in the age of AI (28:01) AI in education (41:01) Principles for living in the era of AI

In this week’s Five-Minute Friday, Jon Krohn holds the AI bubble up to the light. He points to the deep greyzone found in AI startups like Cluely that are established on dubious ideas (Cluely’s tagline was “cheat on everything”) and funding bluster, as well as the staggering spending by companies on infrastructure and researcher salaries. Listen to the episode to hear about the historical precedents to the AI bubble that go all the way back to the invention of the railway, what to make of current investments in AI, and what you can do about these changes as an AI practitioner. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/974⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

No one should be manually writing code in 2026, thinks Chris Fregly, Jon Krohn’s guest on this week’s episode. In this interview about Chris’ latest book, AI Systems Performance Engineering, he explains why it’s so important to consider memory bandwidth when evaluating GPU performance, that understanding the full hardware software stack is the most valuable skill for anyone working in AI development, and which shortcuts we still shouldn’t ever take when writing code, even though we might be outsourcing a great deal to generative AI. This episode is brought to you by the ⁠⁠Cisco, by Acceldata and by ⁠ODSC, the Open Data Science Conference⁠. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/973⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (03:39) Why Chris wrote AI Systems Performance Engineering  (21:39) Essential coding metrics  (37:24) The importance of inference when coding (42:11) How to manage workflows while using AI agents (51:37) Where and how to invest in the AI market

Jon Krohn recaps the month of February in this episode of In Case You Missed It. Across four interviews with Will Falcon (Episode 965), Tom Griffiths (Episode 969), Antje Barth (Episode 963), and Praveen Murugesan (Episode 967), Jon questions the brains behind some of the AI industry’s most innovative companies about launching a startup, developing a popular product, what artificial intelligence can still learn from human intelligence, and how AI might finally start to think on its own. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/972⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

Lin Qiao, CEO of Fireworks AI, talks to Jon Krohn about how she builds effective models quickly, why coding agents can perform at the level of a junior engineer, and what she attributes to the success of Fireworks AI: True to its name, the company exploded into the AI industry with over $300 million secured in venture capital, as well as netting a further $250 million Series C funding. For Lin, many enterprises miss out by not being familiar with open models. Open models give a lot of control to the user, offering customizability and at a much lower price point. Listen to hear how Fireworks AI helps companies continue to save money through AI. This episode is brought to you by the⁠ ⁠⁠⁠⁠Dell⁠⁠⁠, by⁠ ⁠⁠Intel⁠⁠⁠, by ⁠Cisco⁠ and by ⁠Acceldata⁠. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/971⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (01:19) All about Fireworks AI (24:16) Why companies need to take notice of open models (33:05) The commercial viability of slow-reasoning models (38:51) Fireworks AI’s approach to model performance evaluations

Working with code-gen models and Claude Code: In this Five-Minute Friday, Jon Krohn addresses how AI superstars like Andrej Karpathy are using AI agents in their coding work, the outlook for code-gen in 2026, and how you can get started. Hear about Karpathy’s work as well as the soaring success of Peter Steinberger and how he managed to surpass the GitHub commit rate of teams as an individual working with AI agents. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/970⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information.

Princeton Professor Tom Griffiths talks to Jon Krohn about his new book, The Laws of Thought, which grapples with the mathematical models behind biological and artificial intelligence, and what makes the human brain so fascinating for psychologists and computer scientists to study. In this episode, he details how the mathematical principles governing the external world can also be used to explore cognitive science, or “the internal world.”  This episode is brought to you by the ⁠⁠Dell⁠⁠, by ⁠⁠Intel⁠⁠, by Cisco and by Acceldata. Additional materials: ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠www.superdatascience.com/969⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠ Interested in sponsoring a SuperDataScience Podcast episode? Email [email protected] for sponsorship information. In this episode you will learn: (01:18) Tom Griffiths’ current research (21:23) On mathematical inference in LLMs (35:19) How to engineer inductive bias (52:00) How to model curiosity into AI systems