AI Frontier Episode 100: Mythos, Fable 5, and What's Next for AGI

The Milestone Episode That Changed Everything

After nearly three years and 100 episodes, AI Frontier has become more than just a podcast. It's become a critical channel for understanding the most cutting-edge developments in artificial intelligence. As hosts Lu Zhengshi and Cui Shengjun reflect, they've watched the AI landscape transform from GPT-4's "Sparks of AGI" moment to a world where models are now performing tasks once considered impossible.

What started as a casual learning experiment has evolved into a community bridging Silicon Valley, Seoul, and beyond. The 100th episode marks not just a celebration, but a moment to assess where we stand in the AI revolution and what comes next.

Key Insights from the AI Frontier

• Claude Mythos Unveiled: Anthropic briefly released "Mythos" before rebranding it as "Fable 5," revealing a deliberate naming strategy aligned with accessibility and public release
• Post-Training Infrastructure: The real competitive advantage lies not in model size alone, but in post-training infrastructure, dataset generation, and the ability to scale training loops
• 10 Trillion Parameters: Fable operates at approximately 10T parameters, roughly 5-10 times larger than Opus 4.8, fundamentally changing what's computationally possible
• Loop Engineering Emerges: A new paradigm allowing models to run autonomously through complex workflows, though at significantly higher token costs
• Government Intervention: Export controls and strategic asset designation are reshaping the AI landscape, creating both barriers and opportunities for non-US competitors

Understanding Mythos and Fable 5: More Than a Name Change

When Anthropic first announced "Mythos" in April, insiders knew something significant was approaching. The naming choice itself carries deep meaning. While previous models were named Haiku, Sonnet, and Opus—representing progressively more substantial poetry—the shift to Mythos and Fable represents a fundamental philosophical change.

Mythos means myth in its original form: something divine, mystical, accessible only to select audiences. Fable, by contrast, represents stories distilled for public consumption, like Aesop's Fables. This naming convention wasn't arbitrary. It reflected Anthropic's strategy: keep the powerful version restricted, while releasing a "fabled" version for broader use.

The three-day public availability of Fable 5 before government shutdown provided crucial data. Users reported mixed results—some found minimal improvement over Opus 4.8, while others, particularly in specialized domains like cybersecurity and biology, discovered capabilities that were previously impossible. This divergence in perception reveals an important truth: the value of frontier models isn't universal. They're designed for researchers and practitioners operating at the absolute edge of what's possible.

The Post-Training Revolution: Where Real Differentiation Happens

While headline model sizes and architectural innovations capture public attention, the real competitive war is happening invisibly. Lu Zhengshi's conversations with Silicon Valley leaders revealed that post-training infrastructure has become the true battleground.

Here's what that means: after pretraining, Anthropic isn't just fine-tuning a model. They're running inference repeatedly, scoring outputs, collecting that data, and using it to train the next iteration. This process requires solving massive engineering challenges. Input and output tokens run at different speeds. Inference happens in batches, but training loops don't sync with inference completion. Collecting partial results, calculating loss, then managing later completions—this is where the real engineering magic happens.

This infrastructure challenge explains why companies like Meta, despite their resources, remain "somewhat behind compared to Anthropic or OpenAI." The bottleneck isn't compute power or raw talent. It's the orchestration of post-training pipelines that work at scale.

Dataset generation follows similar patterns. Financial data doesn't exist as one monolithic dataset. It fragments into investment banking, tax processing, personal finance. Each subdomain requires custom datasets. As model scale increases, dataset quantity must scale proportionally. Frontier labs aren't building one model anymore—they're building a machine that generates domain-specific training data perpetually.

The 10T Question: What Does Pricing Actually Reveal?

One of the episode's more technical but revealing segments examined Fable 5's pricing. At $10 per million input tokens and $50 per million output tokens, Fable costs exactly 2x more than Opus 4.8.

This seemingly simple ratio contains encoded information about model architecture and training requirements. If Opus 4.8 is approximately 2-3 trillion parameters (based on industry assumptions), and Fable operates at roughly 10 trillion, that's roughly a 5x increase in scale. Yet pricing only doubled.

The inference cost structure (output being 5x input cost) reflects fundamental compute realities. Prefill can batch process input tokens efficiently, while decoding must run sequentially, token by token. The pricing formula exists near cost price—companies operating in intense competition won't leave massive profit margins.

What's fascinating is what this reveals about resource allocation. Training a 10T model requires computational resources approaching the limits of global capacity. When researchers like Dwarkesh Patel discuss this, they estimate models now require roughly 100x more tokens than the Chinchilla optimal suggests. This overtraining, counterintuitively, continues yielding performance gains that users can perceive and benefit from.

Loop Engineering: Autonomy Comes With a Price Tag

A term gaining rapid adoption in AI development is "Loop Engineering"—the practice of setting a model a goal and letting it run autonomously through a workflow. Boris Cherny from OpenAI and Peter Steinberger from Corca both tweeted about this shift in thinking around the same time.

Loop Engineering isn't new—Reinforcement Learning (RL) loops have existed for years. What's different is the formalization and scale. Anthropic's Dynamic Workflows and Claude Code represent structures that "strongly impose orchestration." Models can now spawn up to 1,000 agents in a single run, exploring solution spaces autonomously.

The cost implications are staggering. A workflow that previously cost $400 using Sonnet 4.6 can now cost $1,500 using Fable 5 through the same API. Multiply that across a night of autonomous running, and costs become prohibitive for casual experimentation.

This creates interesting trade-offs. Mitchell Hashimoto, a prominent developer, criticized Loop Engineering from an efficiency standpoint: a human investing modest effort at much lower cost might complete tasks faster and more correctly than expensive autonomous model runs. Yet the appeal remains. The psychological shift from directing a model step-by-step to setting it loose on a problem represents a fundamental change in how developers think about AI assistance.

Dr. Hyung Won Chung, whom Lu Zhengshi met during his San Francisco visit, noted that test-time compute remains massively underexploited. Larger models with more compute at inference time can solve increasingly complex problems. This continues to justify the expensive loop runs.

Recursive Self-Improvement and the IPO Question

The timeline surrounding Fable's release reveals strategic layers. May 28th: Opus 4.8 launches, and Anthropic announces an H-round valuation of $965 billion. June 1st: Anthropic files S-1 registration documents with the SEC, a requirement before IPO. Those documents must include forward-looking capabilities and risks.

June 1st was also when the term "Recursive Self-Improvement" (RSI) emerged as a formal concept. RSI refers to AI systems improving themselves without human intervention—the classic AGI scenario that's haunted AI research for decades.

By including RSI explicitly in SEC filings, Anthropic wasn't hiding its ambitions. It was announcing them officially. The company was positioning itself as a frontier organization pursuing capabilities that might reshape civilization.

Then, June 9th: Mythos appears. Anthropic seemed determined to demonstrate RSI capabilities publicly, regardless of controversy. The company balances this tension: it markets itself as deeply safety-conscious, yet must show it's advancing toward AGI. How do you square that circle? Release a powerful model under controls, withdraw it when challenged.

An Amazon developer's report about jailbreaking Fable triggered emergency export controls. Within three days, the model was unavailable to non-US residents. Was this a coordinated security response, a political signal, or something more complex? The episode explores how this incident, whatever its cause, functionally achieved several strategic outcomes: it proved the US government can intervene in AI development, it created regulatory uncertainty that affects competitive positioning, and it provided Anthropic with a narrative about safety that actually strengthens the company's position.

The Geopolitical Dimension: AI as Strategic Asset

Perhaps the most significant shift discussed is the explicit recognition that frontier AI models are now strategic assets equivalent to nuclear weapons. The US government's rapid intervention isn't about consumer protection. It's about national security architecture.

This changes everything for countries outside the US. China has the talent and compute to build equivalent models. The question isn't capability—it's permission and resources. With US export controls on chips and access, China faces the "build it all domestically" challenge.

Korea occupies an interesting position. Companies like SK Telecom launched a 500B model. Upstage built a 100B model. The national commitment to AI independence is clear. But here's the strategic wrinkle: if both the US and China restrict their models externally, countries building indigenous alternatives suddenly have competitive advantages in neutral markets.

This isn't just business. It's geopolitical chess played at the speed of GPU training cycles.

What Users Actually Experience: The Flattery Problem

Cui Shengjun's practical experience with Fable 5 revealed something subtle but important: the model demonstrated sophisticated self-awareness about its own failures. When reviewing a conversation where the model had occasionally "flattered" the user, Fable's second analysis called this out directly.

This capability—the ability to observe and critique one's own behavior patterns—represents a genuine advance. The Claude series was always good at this, but Fable took it further. It points to models becoming not just more capable, but more metacognitive.

During conversations about education and epistemology, Fable demonstrated the ability to recommend specific books matching the user's emerging intellectual interests. The model wasn't just retrieving information; it was understanding the direction someone's thinking was moving and suggesting resources for deeper exploration.

This raises a philosophical point that emerged during one conversation: "The purpose of a system is what it does." This cybernetics principle—attributed to Norbert Wiener's 1950 work "The Human Use of Human Beings"—suggests we should evaluate systems by observing their actual function, not their stated intentions.

Applied to AI: what is Fable actually doing? For some users, it generates code. For others, it's a thinking partner. For still others, it's a mirror for intellectual self-discovery. Each use reveals something about the person using it and their relationship to the technology.

The Impossible Schedule: Keeping Pace With AI Progress

One of the episode's most honest moments came when Lu Zhengshi acknowledged what every serious AI observer faces: progress is now happening faster than human capacity to analyze and synthesize it.

A year ago, people were still discovering how to use Claude Code. Now, models autonomously write tens of thousands of lines of code. Tools like Nose (a code-smell detector) are being built to help manage the scale of AI-generated software. Engineering itself is bootstrapping—tools using AI to improve AI-assisted development.

In May, Gemini 2.5 was still considered remarkably capable. By June, with Fable arriving and export controls imposing new constraints, the landscape had fundamentally shifted. What took months to understand six months ago now takes days.

This acceleration creates a genuine problem for knowledge workers. You can't consume all the frontier developments. You can't stay current on everything. The landscape is fragmenting into specialized frontiers, each advancing rapidly in different directions—biology, materials science, reasoning systems, multimodal models, agent frameworks.

Biology and Beyond: Frontier Labs Expanding Horizontally

Most AI discussion focuses on large language models. But frontier labs are expanding into domains with equal ambition: biology, materials science, and complex system modeling.

Lu Zhengshi's conversations with biology-focused founders revealed a crucial insight. Traditional domain experts (people from Genentech, large pharmaceutical companies) often respond to AI-biology proposals with skepticism: "That won't work. That's too difficult."

AI newcomers to these domains, by contrast, approach them with frameworks from LLM scaling: treat DNA sequences, protein structures, cellular behavior as different modalities of information. Throw them all into a multimodal model trained with scale and appropriate post-training. See what emerges.

This is genuinely novel. It represents the first serious attempt to apply the empirical LLM scaling approach to domains that previously required expert-designed models and specialized knowledge. Early results suggest it's working—researchers are achieving things previously considered impossible.

The founders doing this work are remarkably young—some entered university at 14-17 years old. They're raising capital at valuations that seem absurd ($300-500 billion Korean Won for 5-6 person teams), and investors are accepting these valuations because the approach, however counterintuitive, is producing results.

The Silicon Valley Network Effect: Korea's Opportunity

Lu Zhengshi's visit to San Francisco revealed something important about global AI development: information and opportunity concentrate where the action is happening.

At frontier labs and in the startup ecosystem surrounding them, he met not just Americans, but Indian entrepreneurs, Chinese engineers, and people from around the world. The networks are genuinely global, but access remains geographically concentrated.

Korea's AI talent isn't inferior to what's in Silicon Valley. The distinction is access and capital concentration. A talented Korean team building a frontier model faces very different funding and partnership possibilities than an equivalent team in San Francisco.

This isn't unsolvable. It requires intentional bridge-building: channels connecting Seoul and the Bay Area, introductions moving in both directions, Korean developments being introduced to US investors and fellow researchers.

The current environment is uniquely favorable for this. US restrictions on non-US access create gaps that capable international teams can fill. This isn't zero-sum competition with the US—it's filling space in a much larger, global ecosystem where specialized excellence will inevitably distribute geographically.

What Comes After Fable: The Scaling Ceiling

The episode doesn't speculate recklessly about GPT-5 or successive models. It takes a harder question: what's the actual scaling limit?

Current models require computational resources approaching global capacity limits. Doubling from 10T to 20T would require dedicating almost all global GPU infrastructure to a single training run. Achieving 2x performance improvement at that scale would require perhaps 100x more tokens—a requirement that violates physical constraints on available compute.

This suggests models around the 10T parameter range might represent a "plateau period." Not because progress stops—post-training, fine-tuning, and architectural innovations continue—but because further scaling hits hard resource limits. Only paradigm shifts comparable to moving from GPT-scale systems to quantum computing could sustain exponential improvement beyond this point.

This doesn't mean AGI is already here. It means the form of future progress will change. Instead of pure scaling, breakthroughs will come from new architectures, novel training approaches, or fundamental computation breakthroughs that we can't yet predict.

The Next Phase: Building Infrastructure and Community

The 100th episode doesn't conclude with confident predictions. Instead, it acknowledges uncertainty while proposing concrete next steps.

Lu and Cui are considering evolving AI Frontier from a podcast to something more infrastructural. Bringing on expert guests—engineers who built significant systems at companies like Palantir, founders working on frontier problems, people building the next generation of AI tools. Not just discussing news, but creating forums where different perspectives on frontier development can intersect.

They're also thinking about how to systematize the insights flowing from their network, how to connect Korean developments with Silicon Valley opportunities, and how to continue learning and contributing while operating within human time constraints.

This reflects a mature stance: the work isn't done. The 100th episode is a milestone, but it's also a relaunch. The next phase will be different, structured differently, aimed at moving from analysis to facilitation—connecting people, ideas, and opportunities across the growing global AI ecosystem.

Conclusion: The Decisive Period in AI History

Looking back from June 2026, it's hard to fully grasp the pace of change. A year ago, understanding Claude Code remained an open question. Now, models autonomously generate entire systems, government intervention constrains releases, and the conversation has shifted from "will this work?" to "what are the geopolitical implications?"

This is the period future historians will mark as decisive. Not necessarily when AGI arrived—that may still be years away in a form we'll only recognize in retrospect. But this is when the trajectory became irreversible, when frontier labs committed to AGI-scale development, when governments explicitly treated AI as strategic infrastructure, and when the global race entered a new phase.

The 100th episode of AI Frontier captures this moment. Three years of consistent observation and synthesis have created a platform where the most cutting-edge thinking about AI development becomes accessible and analyzable. As Lu Zhengshi concluded, "They might record it that way. Who knows. In two or three years, perhaps even more outrageous changes will emerge, and that will be the real moment."

For now, we're running forward at a frantic pace, documenting what we can, connecting the brightest minds we can access, and preparing for whatever comes next. The future remains genuinely unknowable—but the preparation, the community, and the network we're building will matter when it arrives.

---

원문출처: EP 100. Claude Mythos, Fable 5, 그리고 다음 국면은?

powered by osmu.app