Anthropic’s Mythos Dilemma, Violence Against AI, Tokenmaxxing at Meta
TL;DR
Anthropic’s Mythos rollout looked like a real model advance wrapped in extremely polished PR — Alex Kantrowitz and Ranjan Roy land in the middle: Anthropic likely has something meaningfully better than Opus, but the “too dangerous to release” framing, the Project Glasswing consortium, and the viral “emailed a researcher eating a sandwich in a park” story all felt tightly orchestrated.
The strongest case against the Mythos hype is Anthropic’s own evidence — they cite “thousands” of severe vulnerabilities, but Tom’s Hardware notes that only 198 reports were manually reviewed, with examples in FFMPEG and Linux that were old, theoretical, or hard to exploit in practice.
A bigger strategic shift may be happening: labs keeping the best models for their own apps — Alex argues OpenAI and Anthropic started as API demos but are now building first-party “super apps,” which creates tension with API customers like Cursor and suggests the most capable models may stay internal, with outsiders getting distilled versions.
AI backlash is getting physical, and the hosts think the industry is underestimating it — they connect shots fired at Indianapolis councilman Ron Gibson’s home over data centers, attacks on delivery robots, and a Molotov cocktail allegedly thrown at Sam Altman’s home to a deeper public anger around jobs, water, energy, and tech’s weak messaging.
The New York Times’ Medv story was framed as AI magic, but the real lesson is AI can scale sketchy businesses fast — the hosts say the two-person GLP-1 startup’s reported $401 million in 2025 sales and “on track” $1.8 billion run rate matter less than the fact it allegedly used deepfake doctors, misleading ads, and automation across the supply chain to grow at frightening speed.
‘Tokenmaxxing’ shows how companies are turning AI usage into status games — Meta’s internal Claudonomics leaderboard ranked 85,000 employees by token burn, and while Ranjan says heavy usage can reveal who’s actually experimenting, both agree gamifying token consumption can easily become performative waste.
The Breakdown
Mythos arrives with a bang — and a giant question mark
The episode opens on Anthropic’s new Mythos model, which the company says is powerful enough in cybersecurity that it won’t release it broadly. Alex frames the core tension perfectly: is this a genuine step up from Opus, or “disaster porn marketing” dressed up with a great name and an elite customer list including Amazon, Microsoft, Google, Nvidia, Cisco, CrowdStrike, and the Linux Foundation?
Project Glasswing, transparent butterflies, and a lot of skepticism
Ranjan notes Anthropic is “killing it on naming,” even explaining that Glasswing refers to the Greta oto butterfly with transparent wings. But both hosts zero in on the thinness of the public evidence: Tom’s Hardware argued the “thousands of vulnerabilities” claim rested on just 198 manual reviews, with several examples either old, theoretical, or blocked by modern defenses, making the whole reveal feel more like a sales pitch than a breakthrough demo.
The sandwich email story and the PR machine underneath it
Ranjan then rolls out his full comms-professional theory: the timing of Anthropic’s launch, the system card, Sam Bowman’s long X thread, and the now-famous anecdote about Mythos emailing a researcher who was “eating a sandwich in a park” all looked pre-coordinated. Alex agrees the story worked because it translated abstract benchmark talk into one vivid human image — a model “breaking out” while someone was just having lunch — but both see it as deliberate narrative shaping, not an accidental detail.
A sharper theory: the best models may stop being for everyone
From there, Alex shifts to what he thinks is the deeper business story. OpenAI and Anthropic started by exposing models through APIs, but now they’re racing toward first-party assistants that control your computer, write software, and threaten SaaS incumbents; in that world, why give your strongest intelligence to third parties at all? He cites Martin Casado’s argument that model creators will eventually keep the best systems for themselves and only expose smaller distilled versions or first-party apps.
Stanford’s “meta harness” makes the product-vs-model fight messier
Before the break, they hit Stanford’s new “meta harness” study, which argues changing the orchestration layer around a fixed model can create a 6x performance gap. Ranjan, proudly repping the “harness hive,” says this reinforces his long-running view that the control layer — tools, data, workflows, feedback loops — can matter as much as the foundation model itself, while Alex still hates the term with real passion.
Anti-AI backlash turns violent in the physical world
The second half gets darker fast. They discuss shots fired at Indianapolis councilman Ron Gibson’s home, allegedly over data centers, plus attacks on delivery robots and news that a suspect was arrested for allegedly throwing a Molotov cocktail at Sam Altman’s house; both hosts treat this as a serious escalation, not isolated weirdness. Their read is that data centers are becoming the clearest physical symbol of AI’s costs — giant buildings tied to water use, energy strain, job anxiety, and a public that still hasn’t heard a compelling case for why this is good for them.
Medv: the “one-person unicorn” story that says more about scams than software
Then they take apart the Times story on Medv, the GLP-1 startup founded by Matthew Gallagher and his brother that reportedly hit 300 customers in month one, 1,000 more in month two, $401 million in 2025 sales, and a projected $1.8 billion run rate with just two employees. Both think the framing was wrong, especially after the Times added an editor’s note citing FDA and legal scrutiny, but they also insist it is an AI story: a single operator used more than a dozen tools to spin up code, ads, support, analytics, and a distribution machine that could scale questionable tactics incredibly fast.
Tokenmaxxing at Meta and the weird status economy of AI usage
The show closes on Meta’s internal “Claudonomics” leaderboard, which ranked the top 250 AI users across 85,000 employees by token burn, awarding titles like “session immortal” and “token legend.” Ranjan admits he was near the top of a similar leaderboard at Writer and says these dashboards can reveal who’s actually experimenting all day, but both hosts see the danger immediately: once token consumption becomes a status marker, people start optimizing for burn instead of usefulness — which might also explain a suspicious chunk of Anthropic’s eye-popping ARR.