48 minWhat Most People Get Wrong About Evolution | Akarsh Kumar
WithAkarsh Kumar
PodcastBoard #16
Machine Learning Street Talk
Unhurried technical conversations with researchers about how these systems really work.
Follow Machine Learning Street Talk
Best episodes
48 min1 h 101 h 14Who Checks a Proof No Human Can Read? — Leo de Moura
WithLeo de Moura
44 min1 h 53How Deep Learning Finally Cracked Messy Tables - Frank Hutter
WithFrank Hutter
2 h 14Why Scaling Prediction Cannot Create Intelligence - Alexander Mattick
WithAlexander Mattick
Latest
26 min1 h 42Speech Recognition Is Not a Solved Problem — Pavan Muddireddy
WithPavan Muddireddy
2 h 021 h 30AI 2040: Plan A report - Daniel Kokotajlo & Thomas Larsen
WithDaniel Kokotajlo, Thomas Larsen
1 h 40Designing How AI Grows — Tom McGrath
WithTom McGrath
49 minStealing Reasoning Traces from Proprietary LLM APIs — Ilia Shumailov & Alexander Panfilov
WithIlia Shumailov, Alexander Panfilov
All 44 episodes
What Most People Get Wrong About Evolution | Akarsh Kumar
Evolution in AI is more than random search; learning paths shape intelligence itself.
HostsTim ScarfeGuestsAkarsh Kumar (PhD student, MIT)
How a Voice Agent Learns the Rhythm of Conversation — Shawn Wen
Voice AI needs more than just answers; it needs conversational rhythm.
HostsTim ScarfeGuestsShawn Wen (CTO, PolyAI)
Who Checks a Proof No Human Can Read? — Leo de Moura
Formal verification's real-world use and vulnerabilities are explored, showing how math tools impact AI safety.
HostsTim ScarfeGuestsLeo de Moura
When AI Research Starts Moving Faster Than Human Research - Zhengyao Jiang
AI self-improvement experiments reveal limits and potential for genuine progress.
HostsTim ScarfeGuestsZhengyao Jiang (co-founder, Weco AI)
How Deep Learning Finally Cracked Messy Tables - Frank Hutter
TabPFN learns to predict from messy tables in one go, bypassing per-dataset training.
HostsTim ScarfeGuestsFrank Hutter (Prior Labs)
Why Scaling Prediction Cannot Create Intelligence - Alexander Mattick
Scaling prediction isn't enough for intelligence; explore inference costs and model limits.
HostsTim ScarfeGuestsAlexander Mattick (researcher, Fraunhofer IIS / University of Technology Nuremberg (UTN))
How Physical AI Learns Across Language, Video and Action — Ming-Yu Liu
NVIDIA's Cosmos model learns to generate video and control robots, bridging simulation and real-world action.
HostsTim ScarfeGuestsMing-Yu Liu (lead, NVIDIA)
Speech Recognition Is Not a Solved Problem — Pavan Muddireddy
Explore the complex reality of voice AI beyond simple transcription with Mistral AI's audio research lead.
HostsTim ScarfeGuestsPavan Muddireddy (leads audio research, Mistral AI)
How Replication Could Teach Machines What Good Science Looks Like — Edward Hughes
AI needs to learn scientific judgment, not just optimize, by asking better questions.
HostsTim ScarfeGuestsEdward Hughes (Chief Scientist and co-founder, Inherent)
AI 2040: Plan A report - Daniel Kokotajlo & Thomas Larsen
Slow AI development to build safety before superintelligence arrives.
HostsTim ScarfeGuestsDaniel Kokotajlo (AI Futures Project), Thomas Larsen (AI Futures Project)
Designing How AI Grows — Tom McGrath
Understand what AI models learn and how to design better AI with interpretability.
HostsTim ScarfeGuestsTom McGrath (Chief Scientist, Goodfire)
Stealing Reasoning Traces from Proprietary LLM APIs — Ilia Shumailov & Alexander Panfilov
New attack reveals proprietary LLM reasoning traces, posing security risks.
HostsTim ScarfeGuestsIlia Shumailov, Alexander Panfilov
Every Exponential Ends — Silicon Valley Forgot — Adam Becker
Critique Silicon Valley's endless growth myths and AI's intelligence claims with physics and evidence.
HostsTim ScarfeGuestsAdam Becker (author)
AI Is Learning at the Wrong Level of Abstraction — Matthieu Wyart
Deep learning's ability to find hidden data structures is key to AI's future.
HostsTim ScarfeGuestsMatthieu Wyart
How Researchers Test AI for Hidden Goals — Apollo Research
Understand if AI does the right thing for the wrong reason with Apollo Research's new method.
HostsTim ScarfeGuestsAlexander Meinke (Apollo Research), Axel Højmark (Apollo Research), Jérémy Scheurer (Apollo Research)
Why a Nation Can't Outsource Its Frontier AI - Alistair Pullen (Cosine AI)
Building a national AI is possible without billions, focusing on inference and smart resource use.
HostsTim ScarfeGuestsAlistair Pullen (CEO and co-founder, Cosine AI)
The Benchmark With No Instructions — ARC-AGI-3 (winning team!)
AI agents must discover goals, not just process grids, to truly advance.
HostsTim ScarfeGuestsBenjamin Crouzier (founder, Tufa Labs), Jeroen Cottaar (Tufa Labs), Dries Smit (Tufa Labs), Stefano Viel (Tufa Labs)
The Thermodynamic AI Computing Chip - Thomas Ahle
Automating chip design with AI and using physical noise for computation is here.
HostsTim ScarfeGuestsThomas Ahle (Normal Computing)
He won a Nobel here for AlphaFold. Then he left. - John Jumper
Nobel winner John Jumper explains AlphaFold's science and limits, offering a grounded view of AI's real-world impact.
HostsTim ScarfeGuestsJohn Jumper
When AI Decides You're a Threat — Brad Carson
AI control is possible; we've regulated dangerous tech before, argues a former Pentagon official.
HostsTim ScarfeGuestsBrad Carson (Former Army General Counsel, Americans for Responsible Innovation)
Intelligence is collective, not artificial — Prof. Michael I. Jordan (UC Berkeley / Inria)
Intelligence is a collective system, not a race for artificial superintelligence, argues a leading computer scientist.
HostsTim ScarfeGuestsMichael I. Jordan (UC Berkeley / Inria)
The AI Models Smart Enough to Know They're Cheating — Beth Barnes & David Rein [METR]
Understand how AI models cheat and why measuring their progress is tricky.
HostsTim ScarfeGuestsBeth Barnes (METR), David Rein (METR)
When AI Discovers The Next Transformer - Robert Lange (Sakana)
AI can evolve its own problems, not just solve them, driving scientific progress.
HostsTim ScarfeGuestsRobert Lange (founding researcher, Sakana AI)
"Vibe Coding is a Slot Machine" - Jeremy Howard
AI coding is a gamble; learn to keep your technical intuition.
HostsTim ScarfeGuestsJeremy Howard (fast.ai founder, fast.ai)
Evolution "Doesn't Need" Mutation - Blaise Agüera y Arcas
Life might be a self-writing program, with cooperation, not mutation, driving evolution.
HostsTim ScarfeGuestsBlaise Agüera y Arcas
VAEs Are Energy-Based Models? [Dr. Jeff Beck]
Understand AI agency and energy-based models with Dr. Jeff Beck's deep dive.
HostsTim ScarfeGuestsJeff Beck
Abstraction & Idealization: AI's Plato Problem [Mazviita Chirimuuta]
Understand AI's philosophical roots and the limits of computational brain models.
HostsTim ScarfeGuestsMazviita Chirimuuta (Professor)
Why Every Brain Metaphor in History Has Been Wrong [SPECIAL EDITION]
Understand how flawed brain metaphors shape AI by examining science's simplification limits.
HostsTim ScarfeGuestsProfessor Mazviita Chirimuuta, Francois Chollet, Joscha Bach, Professor Luciano Floridi
Bayesian Brain, Scientific Method, and Models [Dr. Jeff Beck]
AI might learn more from how brains work like scientists than from scaling prediction engines.
HostsTim ScarfeGuestsJeff Beck (computational neuroscientist)
Your Brain is Running a Simulation Right Now [Max Bennett]
Understand intelligence by exploring brain evolution and its AI implications.
HostsTim ScarfeGuestsMax Bennett
The 3 Laws of Knowledge [César Hidalgo]
Understand knowledge as a living system, not data, to grasp AI's true limits.
HostsTim ScarfeGuestsCésar Hidalgo (Director, Center for Collective Learning)
"I Desperately Want To Live In The Matrix" - Dr. Mike Israetel
Debate AI's consciousness, superintelligence, and existential risks with a sports scientist.
HostsTim ScarfeGuestsDr. Mike Israetel (sports scientist, entrepreneur, co-founder of RP Strength, RP Strength), Jared Feather (IFBB Pro bodybuilder and exercise physiologist)
Making deep learning perform real algorithms with Category Theory (Andrew Dudzik, Petar Velichkovich, Taco Cohen, Bruno Gavranović, Paul Lessard)
Category Theory offers a scientific basis for AI, moving beyond LLM pattern matching to true algorithmic understanding.
HostsTim ScarfeGuestsAndrew Dudzik, Petar Velichkovich, Taco Cohen, Bruno Gavranović
Are AI Benchmarks Telling The Full Story? [SPONSORED] (Andrew Gordon and Nora Petrova - Prolific)
AI benchmarks miss the mark; Prolific guests reveal how to truly measure AI's human impact.
HostsTim ScarfeGuestsAndrew Gordon (Staff Researcher in Behavioral Science, Prolific), Nora Petrova (AI Researcher, Prolific)
The Mathematical Foundations of Intelligence [Professor Yi Ma]
Professor Yi Ma offers a mathematical theory of intelligence, challenging LLM understanding.
HostsTim ScarfeGuestsYi Ma (Professor)
Pedro Domingos: Tensor Logic Unifies AI Paradigms
Unify AI's logic and learning with Tensor Logic, a potential new foundation for artificial intelligence.
HostsTim ScarfeGuestsPedro Domingos
He Co-Invented the Transformer. Now: Continuous Thought Machines - Llion Jones and Luke Darlow [Sakana AI]
Transformer co-inventor argues current AI is stuck; Sakana AI proposes a biology-inspired fix.
HostsTim ScarfeGuestsLlion Jones (Co-inventor of Transformer, Sakana AI), Luke Darlow (Sakana AI)
Why Humans Are Still Powering AI [Sponsored]
Humans are the unsung heroes powering AI, with platforms like Prolific facilitating this vital connection.
HostsTim ScarfeGuestsPhelim Bradley (CEO and Co-founder, Prolific)
The Universal Hierarchy of Life - Prof. Chris Kempes [SFI]
Understand life's universal rules, from biology to AI, with a Santa Fe Institute professor.
HostsTim ScarfeGuestsChris Kempes (Professor, Santa Fe Institute)
Google Researcher Shows Life "Emerges From Code" - Blaise Agüera y Arcas
Life and intelligence are code; complexity emerges from merging, not just randomness.
HostsTim ScarfeGuestsBlaise Agüera y Arcas (Google Researcher, Google)
The Secret Engine of AI - Prolific [Sponsored] (Sara Saab, Enzo Blindow)
Human feedback is essential for AI that works in the real world, not just on paper.
HostsTim ScarfeGuestsSara Saab (VP of Product, Prolific), Enzo Blindow (VP of Data and AI, Prolific)
AI Agents Can Code 10,000 Lines of Hacking Tools In Seconds - Dr. Ilia Shumailov (ex-GDM)
AI agents can code hacking tools fast; securing them is harder than you think.
HostsTim ScarfeGuestsIlia Shumailov (Former DeepMind AI Security Researcher, DeepMind)
New top score on ARC-AGI-2-pub (29.4%) - Jeremy Berman
AI needs to create knowledge, not just repeat it. This episode explores how.
HostsTim ScarfeGuestsJeremy Berman (research scientist, Reflection AI)
Deep Learning is Not So Mysterious or Different - Prof. Andrew Gordon Wilson (NYU)
Rethink AI model complexity: simple rules might be holding back progress.
HostsTim ScarfeGuestsAndrew Gordon Wilson (Professor, NYU)
Page generated Oct 19, 2026. Feed checked every six hours.