Episodes

  • The Self-Play Engine: Can AI Teach Itself What Humans Never Showed It?
    11 Nov 2026

    Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

    Most AI begins by learning from humans. Our writing. Our code. Our images. Our games. Our scientific knowledge.

    But what happens when the machine starts generating its own problems, experiments, opponents, and training data?

    In Episode 18 of Zero to Singularity, we investigate self-play and self-generated learning one possible route beyond the limits of human-created data.

    We explore AlphaZero-style self-play, reinforcement learning, synthetic problems, automated curricula, verifiers, simulated environments, adversarial training, and systems that create increasingly difficult challenges for themselves.

    Could future AI improve by repeatedly competing against previous versions of itself? Can it discover strategies humans never demonstrated? And where does self-learning fail when there is no reliable way to determine whether an answer is actually correct?

    The deeper question is:

    Can intelligence keep improving once humans are no longer capable of providing all the lessons?

    Zero to Singularity — Understanding AI from the fundamentals to the frontier.

    Show more Show less
    58 min
  • The Infinite Context Illusion: Can AI Actually Remember Everything You Give It?
    9 Nov 2026

    Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

    Modern AI systems can accept books, codebases, conversations, research papers, and enormous collections of documents inside a single prompt.

    Some models can process astonishing amounts of information at once.

    But a bigger context window does not necessarily mean a better memory.

    In Episode 17 of Zero to Singularity, we investigate what a context window really is—and what happens when AI is asked to reason across hundreds of thousands or even millions of tokens.

    We explore attention, tokens, positional encoding, KV caches, retrieval, “lost in the middle,” long-context benchmarks, memory compression, RAG, recurrent memory, and the enormous computational cost of keeping more information available.

    Can AI genuinely reason across an entire library? Does it remember information equally well wherever it appears? And will the future belong to gigantic context windows—or systems that learn what to remember, retrieve, and forget?

    The deeper question is:

    Does giving an AI access to more information actually make it understand more or merely give it a larger place to search?

    Zero to Singularity — Understanding AI from the fundamentals to the frontier.

    Show more Show less
    57 min
  • When Intelligence Becomes Infrastructure: What Does the Future of AI Actually Look Like?
    6 Nov 2026

    Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

    For decades, artificial intelligence was something humans used occasionally. The next phase may be very different.

    AI could become a persistent layer running through software, workplaces, laboratories, robots, vehicles, education, medicine, and everyday devices—less like a single product and more like infrastructure.

    In Episode 16 of Zero to Singularity, we step back from individual breakthroughs and ask what the next decade of AI may actually look like.

    We examine reasoning systems, autonomous agents, persistent memory, multimodal models, robotics, scientific discovery, smaller local models, inference-time scaling, synthetic data, specialized experts, compute, energy, chips, regulation, and the economics of increasingly abundant machine intelligence.

    Which trends are already visible? Which require major breakthroughs? What bottlenecks could slow progress? And what would change if useful cognitive work became dramatically cheaper?

    The deeper question is:

    What happens to technology and society when intelligence stops being a scarce capability and starts becoming infrastructure?

    Zero to Singularity — Understanding AI from the fundamentals to the frontier.

    Show more Show less
    59 min
  • The Compressed Mind: Can a Smaller AI Learn What a Larger One Knows?
    4 Nov 2026

    Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

    he most capable AI models can be enormous, expensive, and difficult to run everywhere.

    But what if a much smaller model could inherit part of what a larger model has learned?

    In Episode 15 of Zero to Singularity, we investigate knowledge distillation the family of techniques used to transfer behavior, predictions, reasoning patterns, or task ability from a powerful teacher model into a smaller student.

    We explore teacher-student training, soft targets, synthetic data, model compression, reasoning distillation, self-distillation, quantization, pruning, and the trade-off between capability and efficiency.

    Can a small model reproduce the useful behavior of a much larger one? What gets lost during compression? Can an AI learn from another AI without seeing the original training data? And if increasingly capable models generate the training material for the next generation, where does the knowledge really come from?

    The deeper question is:

    How much intelligence is contained in the size of a model and how much can be compressed into a smaller system without losing what made the larger model capable?

    Zero to Singularity — Understanding AI from the fundamentals to the frontier.

    Show more Show less
    38 min
  • The Hidden Committee: How Mixture-of-Experts AI Decides Which Brain to Use
    30 Oct 2026

    Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

    Most people imagine an AI model as one enormous neural network using all of its intelligence for every question.

    But some of the most powerful modern systems work differently.

    Instead of activating every part of the model for every token, mixture-of-experts architectures can route different inputs through different specialized groups of parameters.

    One part may become useful for code. Another may respond more strongly to mathematics, language, or other patterns. A small routing system decides which experts should participate.

    In Episode 14 of Zero to Singularity, we investigate why AI researchers are building models that can contain enormous numbers of parameters without activating all of them at once.

    How does an AI decide which experts to use? Do experts genuinely specialize? Why can sparse models be cheaper to run than equally large dense models? What happens when the router makes a bad decision? And could future AI systems become less like one giant brain and more like a dynamically assembled society of specialists?

    The deeper question is:

    Does the path to larger intelligence require making every part of an AI bigger or teaching it which parts of itself are worth activating?

    Zero to Singularity — Understanding AI from the fundamentals to the frontier.

    Show more Show less
    54 min
  • The Inference Revolution: Can AI Get Smarter Just by Thinking Longer?
    25 Oct 2026

    Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

    For years, the dominant recipe for improving AI was straightforward: train larger models on more data using more compute.

    But another path is becoming increasingly important.

    Instead of spending all the computation before deployment, what if an AI can spend more computation after you ask the question?

    In Episode 13 of Zero to Singularity, we investigate inference-time and test-time compute the idea that AI performance can improve by searching, checking, revising, using tools, exploring multiple solutions, and allocating more computation to difficult problems.

    Why can the same underlying model sometimes produce a much better answer simply by being allowed to work longer? Is this genuinely better reasoning, sophisticated search, or both? And could future AI systems dynamically decide how much computation each problem deserves?

    The deeper question is:

    What if the next leap in intelligence comes not from building a bigger brainbut from giving the brain more time to think?

    Zero to Singularity — Understanding AI from the fundamentals to the frontier.

    Show more Show less
    43 min
  • The Autonomous Scientist: Can AI Discover What Humans Haven’t?
    19 Oct 2026

    Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode!

    For most of AI’s history, machines have learned from discoveries humans already made.

    But what happens when AI begins helping create the next discoveries?

    In Episode 12 of Zero to Singularity, we investigate the rise of AI systems that can generate hypotheses, design experiments, analyze results, operate automated laboratories, and decide what to test next.

    We explore autonomous labs, scientific agents, causal reasoning, drug and materials discovery, simulation, robotics, verification, and the difference between predicting something correctly and actually understanding why it happens.

    Can AI discover something genuinely new? Can it choose which scientific questions are worth asking? And if an AI proposes an experiment, robots perform it, and independent laboratories confirm the result who made the discovery?

    The deeper question is:

    What happens when machines stop merely learning from humanity’s scientific knowledge and begin helping produce knowledge that did not exist before?

    Zero to Singularity — Understanding AI from the fundamentals to the frontier.

    Show more Show less
    51 min
  • The Verification Problem: How Do We Know When AI Is Actually Right?
    12 Oct 2026

    Hey! I'd love to hear your thoughts, send me a voice note.

    AI can now write code, solve math problems, analyze research, plan multi-step tasks, and produce answers that sound remarkably confident.

    But intelligence creates a strange problem:

    What happens when producing an answer becomes easier than verifying whether the answer is correct?

    A model can generate thousands of lines of code in minutes.

    A human may need hours to inspect them.

    An AI can propose a mathematical proof almost instantly.

    Checking every step may be harder than creating the proof.

    An autonomous agent might perform hundreds of actions before a person ever sees the final result.

    So the next frontier of AI may not simply be building systems capable of generating better answers.

    It may be building systems capable of proving when those answers deserve to be trusted.

    Central question

    If AI eventually becomes better than humans at solving difficult problems, who or what will be capable of checking its work?

    That gives us a fantastic episode spanning AI reasoning, hallucinations, uncertainty, reinforcement learning, formal verification, code execution, scientific replication, AI judges, and the emerging idea of scalable oversight.

    Show more Show less
    55 min