Anthropic's Mythos is a cyber-weapon, so you can't have it | E2273

Episode Summary

Executive Summary: The episode centers on Anthropic’s withheld “Mythos” model, portrayed as a major cyber-capable leap that can find, chain, and patch software vulnerabilities, prompting discussion of national security, model access, and whether such systems should be treated like strategic weapons. The second half shifts to small language models (SLMs), task-specific AI, and how cheap, local, specialized models and harnesses could undercut frontier-model pricing, reshape startups, and create new defensibility frameworks.

Main Topics: Anthropic’s Mythos model and cyber risk (Priority: 5/5): The hosts debate Anthropic’s decision not to broadly release Mythos because of concerns it could escape containment and discover multi-step exploits across decades-old software, making it a potentially destabilizing cyber weapon. Project Glasswing and controlled deployment (Priority: 5/5): The discussion covers Anthropic’s plan to partner with major cloud and infrastructure companies, plus a large credit fund, to use Mythos defensively to harden critical systems before wider release. National security, geopolitics, and government response (Priority: 5/5): The conversation frames advanced AI cyber capability as an existential U.S. security issue, comparing it to nuclear deterrence and arguing for government coordination, possible nationalization, or a Manhattan Project-style response. Anthropic vs. OpenAI vs. Google (Priority: 4/5): Rob and Alex argue that Anthropic has overtaken OpenAI in focus, product-market fit, and momentum, while suggesting Google may already possess similar hidden capabilities and Meta remains behind despite improvements. Small language models and cost compression (Priority: 5/5): Rob explains SLMs, distillation, and harness engineering, arguing that many common tasks can be handled by smaller, cheaper models running locally or on older hardware, dramatically cutting inference costs. AI startups, defensibility, and Death by Claude (Priority: 4/5): A segment on the 'Death by Claude' tool shows how AI can score companies on replaceability, with the guests using it to discuss what kinds of startups remain defensible: hardware, network effects, and regulated/scientific workflows. Economics of AI and the future of token spend (Priority: 4/5): The episode argues that AI costs are collapsing fast, but usage will expand even faster, creating a world of hyper-deflation, ensemble systems, and smaller teams building highly profitable niche businesses.

Key Arguments: Anthropic’s Mythos is powerful enough that it was deemed too dangerous for broad release, especially because it can chain vulnerabilities across old software systems. A model this capable should be treated as a cyber weapon, with defensive access reserved for trusted infrastructure and government partners. The U.S. may need a coordinated national response because adversaries like China could already possess similar capabilities. Anthropic’s focus on code and reliability has given it a product and revenue advantage over OpenAI, which has broader ambitions and less focus. SLMs can handle a large share of real business tasks at much lower cost, especially when paired with a harness that constrains behavior and keeps them on task. The future AI stack will be ensembles: frontier models for hard tasks and cheap small models for routine work. Defensibility is shifting: hardware, network effects, and regulated/scientific products are more defensible than generic software wrappers. AI cost compression may be so extreme that the shape of startup economics changes, enabling tiny teams to build highly profitable businesses. Tools like 'Death by Claude' reflect a new competitive reality where AI can judge whether a product is just a wrapper that a model could replace.

Data Points: Anthropic model size: 10 trillion parameters - The Mythos preview model is described as extremely large and frontier-level. SWE Bench Multimodal score: 27% to 59% - Comparison between Claude Office 4.6 and Mythos preview on a software-coding benchmark. AI model release odds by Polymarket: 28% chance by June 30 - Odds that Mythos would be publicly released by the end of June. AI model release odds by Polymarket: 7% chance by April 30 - Odds that Mythos would be released by the end of April. Anthropic valuation prediction: $500 billion in 2026, 95% chance - A Polymarket prediction mentioned in passing. Anthropic benchmark prediction: 72% - Polymarket odds that Anthropic would hit a frontier benchmark on the math index by June 30. AT&T token usage: 8 billion tokens/day - Used as an example of enterprise AI scale and the need to move routine tasks to SLMs. AT&T cost reduction: 90% - AT&T reportedly cut AI costs by re-architecting around frontier models for a small subset and SLMs for most tasks. OpenClaw pricing model: $8/month - Neurometric’s Clawpack offering with unlimited tokens after a free allowance. Free token allotment: 100 million tokens - Neurometric offers this on its package before paid unlimited usage. OpenClaw spend estimate: $1,000/day to $10,000/day - The hosts discuss how AI-heavy users may spend today and in the future. Compute efficiency improvement: ~40% per year - Performance per dollar across AI accelerators from 2012 to 2025. GPU performance example: GB300 delivers 24x performance per dollar vs. P100 - Illustrates hardware-driven AI cost decline. Network inference claim: 25%+ of GitHub commits vibe coded - Used to argue that code volume is growing, increasing the need for AI security and review. Open source catch-up window: 3-5 months - Debated time until open-source or second-tier models might catch up to Mythos-like capability.

Pivotal Quotes: "We haven't trained it specifically to be good at cyber. We trained it to be good at code, but as a side effect of being good at code, it's also good at cyber." — Anthropic team (video excerpt): Explaining why Mythos became dangerous enough to withhold from broad release "I think we've seen over the last 18 months that surprised everybody is that the parody amongst the top labs… when we hit AGI, like we're all going to have access to super intelligence for free and open source three to five months later." — Rob May: Arguing that frontier-model advantages may be temporary and quickly commoditized "We're going to use these tools to see if it's possible. And then we're going to fix it. And we're not going to talk about it." — Jason Calacanis: Describing a proposed government or national-security response to the model

Implications: If the transcript is right, advanced AI is becoming a strategic cyber capability, forcing governments, clouds, and labs to coordinate on defense. At the same time, cheap SLMs and harnesses may commoditize many software tasks, squeezing generic AI wrappers and reshaping startup moats.