Anthropic’s Mythos Dilemma, Violence Against AI, Tokenmaxxing at Meta

Episode Summary

Executive Summary: The episode debates whether Anthropic’s “Mythos” model is a genuine leap in capability or a carefully coordinated marketing campaign, then pivots to the growing backlash against AI infrastructure and the questionable but real dynamics of AI-enabled startup scaling. The hosts also discuss “token maxing” inside Meta, arguing that internal token-usage leaderboards may reflect genuine AI adoption but can quickly become performative and distort incentives.

Main Topics: Anthropic’s Mythos model: breakthrough or branding? (Priority: 5/5): The hosts scrutinize Anthropic’s claim that Mythos is so capable at cybersecurity exploitation that it cannot be released publicly. They argue the rollout, consortium partners, and selective evidence suggest a strong PR component, though they stop short of calling it a nothingburger. Security claims and the Glasswing Project (Priority: 5/5): They examine Anthropic’s Glasswing consortium and system-card reporting, noting that many alleged vulnerabilities were manually reviewed and some were theoretical or non-exploitable. The hosts question whether the model’s danger is overstated to generate headlines and institutional buy-in. AI backlash turns physical: data centers and attacks (Priority: 5/5): The episode covers bullets fired at an Indianapolis city council member’s house with a note saying 'no data centers,' plus broader hostility toward data centers, delivery robots, and AI leaders. The hosts see this as evidence that AI’s unpopularity is becoming materially consequential. Medvi and the rise of AI-enabled one- or two-person companies (Priority: 4/5): The New York Times story on Medvi sparks debate about whether AI is creating unprecedented leverage or just accelerating sketchy business practices. They conclude the company is real and notable, but the headline revenue framing is inflated and the business has legal/regulatory issues. Token maxing at Meta (Priority: 4/5): A Meta internal leaderboard ranking employees by AI token usage becomes a discussion of how companies are incentivizing AI adoption. The hosts argue this can be a useful signal of who is experimenting most with AI, but becomes dangerous if tied to performance evaluation or status games. First-party AI apps vs third-party APIs (Priority: 4/5): The hosts discuss a structural shift in which model makers are increasingly using their best models in first-party products rather than giving full access via APIs. This raises concerns for companies like Cursor and for developers who rely on third-party model access.

Key Arguments: Anthropic’s Mythos rollout looks highly coordinated: a preview-only release, a dramatic sandwich anecdote, and synchronized PR timing suggest message amplification as much as technical advancement. The cybersecurity evidence behind Mythos is weaker than the hype implies; claims of thousands of severe vulnerabilities rest on a small number of manual reviews and some issues appear non-exploitable. Even if Mythos is real, Anthropic may have strategic reasons to keep the strongest model inside its own products rather than broadly exposing it through APIs. Public resistance to AI is shifting from abstract worry to concrete action, including violence, protests, and political efforts to restrict data center development. Medvi demonstrates that AI can dramatically reduce the labor needed to launch and operate a business, but the story also shows that AI can scale deceptive or predatory practices. Leaderboard-style token metrics can identify heavy AI users and encourage experimentation, but they risk becoming performative 'token maxing' if tied to status or compensation. Model companies may be moving toward first-party super-app experiences, which could squeeze API-dependent businesses and change the AI ecosystem’s economics.

Data Points: Mythos preview availability: about 50 companies and organizations - Anthropic is reportedly limiting access to critical-infrastructure and security partners Organizations named in Glasswing: Amazon, Microsoft, Apple, Alphabet, Linux Foundation, Cisco, CrowdStrike, NVIDIA, JP Morgan Chase, Broadcom, Palo Alto Networks - Used to argue the consortium gives the release legitimacy and breadth Manual vulnerability reviews: 198 - Tom’s Hardware reported Anthropic’s severe-zero-day claims relied on 198 manual reviews Reported vulnerabilities: thousands of high-severity vulnerabilities - Anthropic’s claim about Mythos’s cybersecurity-finding abilities FFMPEG exploit age: 16 years - One cited vulnerability had existed for 16 years and was not deemed critical System card/tweet coordination window: ~20 minutes - Hosts argued the sandwich-story PR timing was tightly coordinated Anthropic valuation round: $380 billion - Referenced as evidence of incentives to amplify model hype before an IPO Medvi startup setup: 2 months and $20,000 - The founder used AI tools to launch the business from home Medvi employees: 2 - The NYT profile framed Medvi as a two-person company Medvi first-month customers: 300 - Initial customer acquisition after launch Medvi second-month new customers: 1,000 more - Growth after the first month Medvi 2025 sales: $401 million - Reported current-year sales at the time of the profile Medvi projected sales: $1.8 billion - The article said the company was on track for this annualized figure Meta leaderboard scope: 85,000 employees - Internal leaderboard reportedly aggregated AI usage across the company Meta leaderboard ranking: top 250 power users - The leaderboard highlighted the most active token users Indianapolis shooting: more than a dozen bullets - Council member Ron Gibson’s house was shot at, with a 'no data centers' note left behind Council member family safety: 8-year-old son was present - The child was in the home and unharmed during the shooting Maine data center moratorium: until November 2027 - The state moved toward a bill pausing data center construction

Pivotal Quotes: "I think this is marketing in just a moment." — Ranjan Roy: Opening skepticism about Anthropic’s Mythos announcement "It sounds like an Avengers movie." — Ranjan Roy: Critique of the branding and narrative around Mythos and Project Glasswing "I legitimately think there's a chance that this is what's happening." — Ranjan Roy: Theory that model labs are keeping best models for first-party products and super-apps

Implications: AI competition is increasingly about PR, product control, and public trust—not just model quality. Expect more backlash against data centers, tighter model access, and more scrutiny of AI-driven business claims and internal usage metrics.