AI Frontier

A deep-dive AI podcast by Chester Roh, Seungjoon Choi, Jonghyun Park, and Seonghyun Kim, exploring the latest in AI technology, industry, and philosophy.

EP 114

No Chance to Escape

In this episode, we explore how large-scale multi-agent systems and massive compute are transforming mathematical research through the case of an attempted solution to the Navier–Stokes equations. We then discuss GPT Image 2.5’s image consistency and editing performance, along with Astra’s computer use and agent applications. We also explain how Google DeepMind’s AlphaGenome Atlas could affect genetic-variant and disease research and drug discovery, examining the present and future of AI for Science. Finally, we discuss the questions rapidly advancing AI technology raises for research ecosystems, safety, and the role of humans.

9/10/2026 · 1:34:35 Hosts: Chester Roh, Seungjoon Choi, Jonghyun Park

EP 113

After GPT-6 Astra: How Much of AI’s Thinking Can We See?

New models, rethinking the harness. Drawing on hands-on experience with GPT-6 Astra, we discuss speed and cost, the possibilities of No-CoT and latent reasoning, and the limits of observing a model’s reasoning. We explore harnesses in ARC-AGI-3, Computer Use, and the value of trial and error in building good products.

9/9/2026

EP 112

What Happens Inside an AI Chip: From KV Cache to Roofline

This deep-dive lecture with HyperAccel CTO Jinwon Lee explores the core principles of LLM inference from end to end. We break down the difference between prefill and decode, the capacity challenges of the KV cache, why decode is inevitably memory-bound through arithmetic intensity and roofline analysis, and how to interpret the interactivity–throughput Pareto curve in the SemiAnalysis InferenceX benchmark. We then examine production optimization techniques including continuous batching, PagedAttention, chunked prefill, speculative decoding, and prefill-decode disaggregation, as well as why KV cache management is becoming a new bottleneck in the age of agentic AI. We also discuss the future of heterogeneous computing through Bertha, HyperAccel’s low-power, low-cost LPDDR-based AI accelerator; OpenAI’s Jalapeño; and what NVIDIA’s acquisition of Groq could mean for the industry.

9/1/2026

EP 111

Benchmark Leader Eliminated? Inside the Four-Way Dokpamo Race

We unpack Korea's sovereign foundation-model (Dokpamo) selection and why benchmark leader Motif was eliminated, then explore Sutton's continual-learning vision, experiments that make design verifiable, and agent workflows.

8/23/2026

EP 110

AI and Alignment

We map the summer's model release rush and the OpenAI agent intrusion disclosed at Black Hat, then dig into reasoning-trace theft, distillation, and why misalignment became the keyword of 2026.

8/18/2026

EP 109

This Is the Inference Era: Inside Sionic's Token Factory and Model Inference

We tour Sionic AI's GPU data center to see how inference servers and token factories run, from prefill-decode separation to what ultra-fast token generation unlocks.

8/16/2026

EP 108

Connecting Chips with Light: Silicon Photonics and AI Data Centers' Next Bottleneck

Explore silicon photonics from the limits of copper interconnects and the rise of CPO and Optical I/O to AI-driven photonic device design.

8/8/2026

EP 107

The Unpredictable Future: Market Roller Coasters, Kimi K3, and Recursive Self-Improvement

We begin with the roller coaster in AI-related stocks and Leopold Aschenbrenner’s portfolio liquidation, then examine Kimi K3’s architecture, post-training and RL design, the open-weights debate, and the dawn of recursive self-improvement.

8/2/2026

EP 106

The Intelligence Supply War: Opus 5, Kimi K3, GPT-5.6 Sol, and the Power Bill

Starting from the agentic attacker that breached Hugging Face, we walk through a week when Opus 5, Kimi K3, and GPT-5.6 Sol all landed at once. We then trace competition shifting from price per token to price per task, the usage migration toward Codex, and the supply-side bottleneck that ends at electricity.

7/26/2026

EP 105

Can AI Do Research?

Can AI conduct research? We examine cracks in academia’s evaluation system—from a rejected PPO paper to the explosion of papers and reviews at ICML, and the debate over peer review on arXiv.

7/25/2026

EP 104

From the GPT-5.6 Launch to ICML

This episode dives into the surprise launch of GPT-5.6 amid the excitement of ICML week in July 2026. We explore the latest frontier model developments, including the performance and pricing of GPT-5.6—hailed as the new daily driver—the arrival of GPT-Live, signs of recursive self-improvement (RSI), and even rumors about GPT-6. We also take you inside ICML at COEX in Seoul, covering the booths of Big Tech companies, neoclouds, and data providers, the fierce competition for talent, and the sharp divide between scaling advocates and domain experts over AI for Science. Through conversations with frontier lab researchers at side events, we offer a vivid look at the accelerating pace of AI progress and the industry’s rapidly shifting landscape.

7/18/2026

EP 103

When Understanding Becomes the Bottleneck

Starting with Fable’s problem-solving ability and rapid usage burn, we discuss model selection, subagent roles, and ways to reduce cognitive debt. We also examine the limits of model evaluation and human preference through Artificial Analysis and LMArena.

7/12/2026

EP 102

Lessons from San Francisco: "Everyone's Gone Crazy"

Chester Roh's three-week Silicon Valley field report on frontier labs, startups, VC, AI x Bio, post-train data, test-time compute, and a startup scene where 20-something founders and 400 billion KRW valuations have become routine.

7/5/2026

EP 101

“Mission Is the Moat”: How VIZCOM Built a $50M AI Design Startup

In San Francisco, we sit down with VIZCOM founder Jordan Taylor to explore where industrial design meets generative AI — his path through Hongik, Honda, and NVIDIA, founding VIZCOM off a Reddit reaction, the Silicon Valley funding scene, and the confidence Korean founders and designers need.

6/29/2026

EP 100

AI Frontier Episode 100: Claude Mythos, Fable 5, and What Comes Next?

AI Frontier reaches episode 100 with a discussion of Anthropic Fable 5, Mythos-class models, post-training, dataset engineering, loop engineering, RSI, sovereign AI, and Korea's foundation model opportunity.

6/16/2026

EP 99

Opus 4.8 Launch, Today's AI Competition, and Human Work

Prompted by Google I/O and the Opus 4.8 announcement, we look at how quickly AI model development cycles are accelerating. Through Gemini 3.5 Flash, Anthropic Dynamic Workflows, code as harness, and long-running agents, we discuss why operating structure, cost, and latency are becoming more important than the models themselves.

6/7/2026

EP 98

In the Age When AI Executes, What Remains for Humans Is 'Intent' (Hashed CEO Seojoon Kim)

Hashed CEO Seojoon Kim joins AI Frontier to discuss with Chester Roh how the age of AI and agents may reshape the future. They explore a world where AI handles execution and humans are left with intent, the rise of a new kind of founder, and the unbundling of traditional institutions such as VCs, universities, and companies.

5/26/2026

EP 97

People in the Age of AI Psychosis

From Claude Code and Codex evolving into super apps to T_brain, burnout, and AI psychosis, Roh Jungseok and Choi Seungjun look at the human bottlenecks and new working rhythms emerging in the AI era.

5/21/2026

EP 96

LLM Inference Infrastructure and Token Economics

We study Dwarkesh's chalkboard lecture and unpack the economics of LLM inference infrastructure, from KV cache and prefill/decode to HBM, NVLink, and roofline analysis.

5/8/2026

EP 95

Reading the DeepSeek-V4 paper

A deep dive into DeepSeek-V4 — how the 1.6T scale, Sparse Attention, mHC, and Muon Optimizer reshape long-context performance and KV cache cost, alongside this week's GPT-5.5 and Cloud Next news.

4/28/2026

EP 94

Anthropic and the low-hanging fruit

We take a lightning tour of the flood of AI industry updates over the past two weeks—from Anthropic’s Mythos, Opus 4.7, and Claude Design to shifts across the coding-agent ecosystem. We also examine the acceleration of model release cycles, the felt cost impact of tokenizer changes, and emerging harness/security architecture trends like “brain–hands” decoupling. Finally, we discuss two “escape routes”—unbundled AI app businesses and AI for Science—and share a perspective on what to study next and where to place bets.

4/21/2026

EP 93

After the Claude Code Source Code Leak

On March 31, 2026, Claude Code's entire source code was leaked through npm distribution. With Noah Ko (Sionic CEO), we discuss the timeline, the Claw Code controversy, and how IP value is transforming in the age of harness engineering.

4/7/2026

EP 92

Close the Loop

Centered on podcasts featuring Andrej Karpathy and Sarah Guo, along with an interview between Terence Tao and Dwarkesh Patel, this episode examines the possibilities and limits of 'manifest' in the age of AI.

3/29/2026

EP 91

AI from the Business Perspective — 26.Q1 Update

OpenClaw meetup recap, viewing AI as a search problem, bundle-unbundle structural shifts, 1/10x efficiency vs 10x new business, and honest conclusions from 4 years of AX.

3/21/2026

EP 90

Ten Years Since AlphaGo (feat. Jinwon Lee, CTO of HyperAccel)

On March 14, 2026, ten years after AlphaGo shocked the world, we look back on how AI has accelerated since then. Joined by Jinwon Lee, CTO of HyperAccel, we discuss the growth of Korea's deep learning community, the cost problem that inference-focused AI semiconductors are trying to solve, and the technical arc from AlphaGo to today's reasoning models. We also cover DeepMind's AlphaGo 10th anniversary post, the direction the next decade symbolizes through "Move 37 (Platform 37)," and the meaning of self-improving loops like Autoresearch and RLVR. In a fast-changing era, look to the timeline of the past for clues about what to build and how to prepare.

3/15/2026

EP 89

One Click and Bumps

With the release of GPT-5.4 and Claude's new feature announcements ushering in the 'one-click' era of AI coding tools, we explore how to effectively leverage AI agents through harness engineering and scaffolding.

3/15/2026

EP 88

There Is No Secret

In March 2026, we welcome back Seonghyun Kim to discuss the changes in the AI industry over the past two months and the outlook ahead. Through the GLM 5 report, we confirm that RL remains the core methodology driving model advancement, and share our analysis that environment scaling will be the key bottleneck determining the trajectory of future progress.

3/4/2026

EP 87

The Age of One Click — Between Grief and Joy

On a Saturday morning, February 21, 2026, Chester Roh and Seungjoon Choi discuss the speed, FOMO, and so-called 'AI depression' brought on by recent AI model and agent trends.

2/24/2026

EP 86

Agentic Workflow for Your Real Work

Together with Lablup CEO Shin Jeong-gyu, we directly showcase `Backend.AI:GO` (the productization of the Continuum router), completed at roughly 1 million lines of code in just 40 days, discussing why it was built and what's different about it. We explore how features that "grew because we needed them"—such as routing by connecting local and cloud models, circuit breaking for failure response, benchmarking/statistics, translation, and image generation—were organized under a coherent philosophy. We also share frontline insights on the agentic coding era: how token usage and bottlenecks are changing, methods to reduce thinking budgets, and why high-speed inference is becoming increasingly important. Ultimately, we walk through practical workflows for how to design a "harness that reduces your workload" and extend automation to everyday tasks across development, finance, marketing, and content.

2/18/2026

EP 85

OpenClaw and the Signals of February 2026

We explore the rapidly emerging AI agent/harness trends (OpenClaw, Pi, Moltbook, etc.) and interpret together what it means when "the paradigm shifts." We compare approaches: the Ralph Loop style of iterating until it works, the Human-in-the-Loop method of steering and calibrating, and the expansion into multi-agent systems — while also discussing security and sandboxing risks through real-world examples. As the time gaps between breakthroughs grow ever shorter, we reflect on what kinds of tacit knowledge and data/context layers we need to hold onto in order to survive without being swept up in big tech's gravitational pull. In a future that is "changing insanely fast" and unevenly at that — what attitudes and choices can we take?

2/11/2026

EP 84

Let's Explore Physical AI: with Jong Hyun Park (sudoremove)

Explore why Physical AI and Vision-Language-Action (VLA) models are at the forefront of robotics today with Jong Hyun Park from sudoremove, examining the latest demonstrations from Boston Dynamics' Atlas to Figure's Helix and tactile-enabled systems. Discover what's driving the shift from "impossible to possible" and dive deep into the fundamental challenge of scaling robot action data—and the strategies like teleoperation, simulation, and data augmentation being used to solve it. We discuss how far robot foundation models might go and what opportunities lie ahead in this emerging market.

2/11/2026

EP 83

Transformers: The Pilgrimage of the Reincarnated Token

This session intuitively explores why Transformers work the way they do, focusing on the journey a token undergoes when it's created. Connecting elements like KV cache, attention, RoPE, and MoE with metaphors like the "memory palace" and the "pilgrimage of reincarnated tokens," we explore how token creation ultimately creates a landscape of meaning. Furthermore, in an era where 10x productivity is becoming the new normal, the discussion expands to encompass the sense that "humans are the bottleneck," the explosion of tools and harnesses, and how to build Minimum Viable Knowledge (MVK). Finally, we'll also be discussing Prompt and AI in Munrae-dong on February 6-7, 2026.

1/28/2026

EP 82

Prompting with Transformer Mechanics in Mind

With the arrival of Claude Opus 4.5, even Andrej Karpathy admitted to feeling FOMO. In this episode, we revisit the principles behind prompting. Why do certain tokens unlock hidden spaces within the model? From Claude's skill system to CoT Faithfulness, and how RL combines skills—now, as AI accelerates rapidly, is the perfect time to revisit the fundamentals. In the next episode, we'll dive deep into Transformer MoE architecture and explore how understanding principles transforms your prompting.

1/28/2026

EP 81

Everything DeepSeek Changed: MoE and RLVR, 2025 AI Year in Review

Saturday morning, December 27, 2025 — Chester Roh, Seonghyun Kim, and Seungjoon Choi look back on 2025 in AI and forecast 2026. They discuss how MoE and RLVR (agent post-training) became mainstream after DeepSeek, and trace how China-led open frontier models have driven the ecosystem. From 'recipes' under limited compute to the evolution of RL infrastructure and the importance of data, they offer grounded perspectives. Together, they imagine what changes scale-up, continual learning, self-play, and more autonomous agents might bring next year.

1/28/2026

EP 80

Will 2026 Be the Year of Science? AI and Science

2025 will be remembered as the year of coding with the launch of Claude Code, and the upcoming 2026 is projected to be a year where AI shifts the paradigm of science. The advancements of OpenAI's GPT-5.2 and Google DeepMind demonstrate the potential for AI application in fundamental scientific fields like mathematics and biology, which will dramatically accelerate research speeds in laboratories. Furthermore, the U.S. government's announcement of the ‘Genesis Mission’ positions AI as a national-level scientific and technological challenge, signaling that AI will play a pivotal role in solving humanity's greatest problems. This video discusses the new future of science that AI will unlock and the changes coming our way.

1/28/2026

EP 79

AI Frontier: The Runaways' Alliance Retrospective & GPT 5.2

In this episode, we unpack what the surge in OpenAI's GDPVal benchmark after GPT-5.2 really means—how costs, speed, and labor are being reshaped. As AI models now handle tasks worth "hours of human work" (per METR/Epoch AI estimates), we ask candidly: what's actually becoming scarce? We also reflect on the Generative Conference (Runaways' Alliance), where 140+ people gathered to share anxieties, ask questions, and self-organize sessions. Our takeaway: embrace symbiosis with AI and live as entrepreneurs who define problems and own the outcomes.

1/28/2026

EP 78

Ilya Sutskever Explains

Why did Ilya Sutskever's single remark send the AI community into an uproar? In this episode, we dissect the core points from Ilya Sutskever's appearance on the Dwarkesh Patel podcast and examine what Noam Brown outlined as the “researcher consensus.” We dive deep with AI researcher Kim Sunghyun into scaling limits, the meaning of emotions as value functions, and why continuous learning is the key to AGI. The latter half quickly covers practical topics too, from the Opus 4.5 vibe check to the moment a black hole scientist got “pilled” by AI.

1/28/2026

EP 77

Gemini 3 & Antigravity: The Insanely Steep Curve of Innovation

This week, the long-awaited Gemini 3 was released, making a big splash in the AI industry. With the advent of Gemini 3, we deeply discuss the advancements in pre-training and post-training, the validity of AI scaling laws, and the impact of accelerated AI development on business and individual capabilities. Through practical application cases such as UI generation, music, and interactive visualization using Gemini 3, we re-examine the importance of flexible approaches like 'unlearn-learn' in the rapidly changing AI era.

1/28/2026

EP 76

Education and AI: Thoughts and Practices of Choi Seung-jun, Founder of Hanmi Kindergarten

Hanmi Kindergarten Website: https://www.hanmiu.cc/ This video explores AI Frontier co-host Seungjun Choi's educational philosophy and his thoughts and practices on the future of education. As a media artist and educator, Seungjun Choi supports the adoption of the Reggio Emilia approach, which shifts focus from programmatic education to fostering children's curiosity and spontaneous exploration. This episode explores why kindergarten education is crucial in the AI era and how technology is unlocking new possibilities for education through concrete examples. This episode provides an opportunity to rethink the essence of education as we prepare for an uncertain future.

1/28/2026

EP 75

Reinforcement Learning (Without Math Formulas)

China's Moonshot Kimi K2 Thinking model has achieved benchmark scores surpassing GPT-5 and Sonnet 4.5, demonstrating rapid model advancements in the post-training era. This video explains the core concepts of Reinforcement Learning (RL), the differences between on-policy and off-policy learning, how the capabilities of models formed during pre-training are strengthened into generalizable patterns through RL, and why accurate feedback is crucial.

1/30/2026

EP 45

AI Business Survival Strategy: Where Should You Build an AI Business?

How can startups survive the assault of frontier models? We discuss the fork in the road between building services on top of AGI infrastructure and creating vertically integrated AI services, and share the insight that the non-verifiable data domain is the true moat that frontier models cannot invade.

3/29/2025