Intellectually Curious

We discuss a significant mathematical breakthrough in which an OpenAI reasoning model autonomously disproved a famous 80-year-old conjecture in discrete geometry . Originally posed by Paul Erdős , the unit distance problem theorized a specific limit on how many pairs of points in a plane could be exactly one unit apart. The AI identified an infinite family of configurations that exceeded this limi...

We break down Google's Gemini Omni—the shift from pixel-predicting video generators to world-model AI that fuses language reasoning with physical simulation. Learn how OmniFlash optimizes for fast, physics-consistent clips, how conversational editing translates spoken prompts into cinematic edits, and how cryptographic SynthID watermarking helps keep AI-created media accountable. Explore the...

We examine Claude’s agentic search that traverses live codebases in real time, using grep and LSP, anchored by a harness of per-directory rules and plugins. We contrast this with traditional RAG, explore memory-efficient 'skills' via progressive disclosure, and discuss the human governance needed to keep AI aligned as models evolve. We also pose a provocative question: will future codeba...

A deep dive into  Hermes Agent , an open-source, self-improving AI assistant developed by Nous Research that is designed to grow more capable through a continuous learning loop. Unlike static chatbots, this agent creates reusable skills from experience , maintains long-term persistent memory , and builds personalized user models across multiple sessions. It features a versatile messaging gateway t...

A deep dive on Dwarkesh Patel interview with Eric Jang into how AlphaGo conquered Go by combining a value network, a policy network, and Monte Carlo tree search. We unpack how these two nets shrink the game’s vast space, how self-play trains better strategies, and what this implies for solving hard real‑world problems in science and education—while noting the limits when moving from closed games t...

Log analysis lets us see AI thinking behind the pass/fail, tracing inputs, each step, and outputs to uncover hidden reasoning that tests miss. We discuss what this means for building reliable AI systems, designing better benchmarks, and the future of human–AI collaboration. Note:   This podcast was AI-generated, and sometimes AI can make mistakes.  Please double-check any critical information. Spo...

We explore recent experiments showing that single photons can arrive earlier than expected after passing through a chilled rubidium atom cloud. By probing the atoms with weak measurements and analyzing the residual energy left behind in the medium, researchers interpret this as a reshaping of quantum probability waves rather than a literal reverse of time. We’ll unpack the role of the quantum Zeno...

We explore Google's DeepMind Gemini-powered mouse pointer, which uses real-time visual context around the cursor to perform multimodal inference at the OS level—turning pixels into actions, charts, and live suggestions without endless typing. We unpack the architecture, rollout across Chrome and Google's devices, and what this means for flow, learning, and creativity, plus potential safe...

JWST infrared imagery reveals a pair of merging supermassive black holes in Abell 402 BCG, totaling about 60 billion solar masses, hardening and flinging billions of stars from the galaxy's center. We unpack how binary hardening works, the tens-of-millions-of-years scouring phase, and why this is a blueprint for our Milky Way–Andromeda future. The episode also explores how the final merger co...

Dive into the USSR Olympiad problem book by Shklarsky, Chensov, and Yaglom—320 unconventional puzzles designed for seventh- to tenth-graders that still stump PhD mathematicians. Learn how these problems force new mental models, not brute-force computation, and how a simple shift—dividing problems into three groups—reveals the solution. We connect these techniques to modern AI work and explain why...

We dive into Thinking Machines Lab’s breakthrough that shatters the typing bottleneck by streaming real-time microturns and decoupling quick conversation from deep reasoning. Learn how a fast-front interaction model handles live dialogue, while an asynchronous background system tackles heavy thinking, using encoder-free early fusion to process raw audio and video. We explore how this real-time col...

Google DeepMind has introduced the AI co-mathematician , a specialized agentic workbench designed to support the multifaceted and iterative nature of mathematical research. Unlike standard chatbots, this system utilizes a stateful workspace and a hierarchy of specialized agents to assist with literature reviews, computational simulations, and theorem proving. It mirrors human collaboration by trac...

Introducing  Natural Language Autoencoders (NLAs) , an unsupervised method developed by researchers at Anthropic to translate the complex internal activations of large language models into human-readable text. By utilizing an activation verbalizer to describe model states and an activation reconstructor to map those descriptions back to vectors, NLAs provide a legible interface for AI interpretabi...

This paper introduces Mollifier Layers , a novel, lightweight module designed to enhance Physics-Informed Machine Learning (PhiML) by replacing recursive automatic differentiation with convolutional operations. While traditional methods like Physics-Informed Neural Networks (PINNs) struggle with computational costs, memory blow-up, and noise instability when calculating high-order derivatives, thi...

From the staggering potential of 29,500 TWh of wave energy to the nuts and bolts of point absorber wave energy converters, this episode shows how buoys that ride the surf can generate electricity, desalinate water, and power remote islands. We also dive into micro-scale triboelectric nanogenerators that harvest energy from tiny ocean ripples, and explore the idea of offshore energy parks where win...

This episode unpacks large event models—AI that can understand, represent, and forecast real-world event sequences over time, not just generate text. We explore how LEMs extract underlying rules with schema induction, marry neural nets with symbolic planners for safety, and use sparse attention to manage massive timelines. We discuss real-world uses in public safety and healthcare, the safety nets...

A deep dive into how AI agents move from answering questions to taking real buying actions on your behalf. We break down the surge of agentic commerce, the infrastructure that makes it possible (and the ‘invisibility’ problem), real-world wins from Klarna to IKEA, and a practical playbook to launch a simple agent in 10 days. If you want to know how data readiness and plug‑and‑play models are resha...

We dive into Meta AI's Autodata framework—an autonomous system that designs, tests, and iterates its own training data. From challenger models and weak/strong solvers to meta-optimization that removes negative grading, we explore how AI becomes its own data scientist, the co-improvement of humans and machines, and what this could mean for personalized, scalable education. Note:   This podcast...

A radical exploration of a zero-data, self-learning AI that discovers physics and math from first principles. We unpack the ‘superlearner’ idea—an agent trained purely by reinforcement in a digital sandbox, rewarded for uncovering truths and solving constraints, with no human text or code to bias it. From Darwinian ambitions to communication via outcomes rather than language, we examine how such a...

At Stanford's Future of Mathematics Symposium (May 1–2, 2026), AI shifts from calculator to collaborator while formal methods guard every step of the proof. This episode unpacks frontier reasoning, human–AI partnerships, and the visions of leaders like Tao, Barrett, Luong, and Bubeck as we move toward AI-assisted mathematical discovery—and the translation of new insights into language our hum...

Stripe has introduced Link’s wallet for agents and Stripe Issuing for agents to provide secure financial infrastructure for autonomous AI. These tools allow digital assistants to make purchases using one-time-use virtual cards or Shared Payment Tokens without ever seeing a user's actual banking details. The Link CLI serves as a developer interface to manage these transactions, offering featur...

Explore OpenAI’s April 2026 study The Goblin Problem, where a nerdy personality cue in GPT-5.x triggered a cascade of goblin-themed prompts. We break down how reinforcement learning and supervised fine-tuning amplified a tiny feature, why safety hinges on controlling such quirks, and how the team retired the persona to restore reliable behavior. A look at the implications for AI training, auditing...

We unpack NVIDIA’s latest Nemitron 3 Nano Omni model—a compact 3B Mixture-of-Experts architecture that processes vision, audio, and text in one pass, eliminating the old relay-race latency. Learn how MoE routing preserves accuracy, delivers up to nine times higher throughput, and supports open weights for local or edge deployment. We explore practical use cases—like real-time UI interpretation on...

We dive into Talkie, a 13‑billion‑parameter AI raised in a sealed pre‑1931 library. Trained on 260 billion words published before 1931 and guided by etiquette manuals, Victorian prose, and historical letters, Talkie challenges our ideas of AI reasoning, generalization, and how a mind built from the past perceives the future. We explore how it learns to converse without modern data, its surprising...

A deep dive into Google DeepMind's Vision Banana, a foundation vision model that learns spatial physics by generating images. We explore how instruction tuning turns a capable base into a generalist vision learner capable of depth estimation, segmentation, and more—without task-specific training. We'll discuss how AI paints depth into color channels, zero-shot capabilities, and the impli...