Big Technology Podcast – April 10, 2026

Episode Title: Anthropic’s Mythos Dilemma, Violence Against AI, Tokenmaxxing at Meta
Host: Alex Kantrowitz
Guest: Ranjan Roy (of Margins)
Summary by Podcast Summarizer

Episode Overview

This episode dives deep into the launch and media frenzy surrounding Anthropic’s latest AI model "Mythos," scrutinizing whether its impact is genuine or the product of savvy PR. The hosts also explore a new and disturbing trend: violence targeting AI infrastructure and leaders, and discuss Silicon Valley’s new cult of "token maxxing," particularly a Meta internal competition over how many AI tokens employees can burn. The conversation is sharp, nuanced, and skeptical—with plenty of industry insight.

Key Discussion Points & Insights

1. Anthropic’s Mythos Model: Breakthrough or Marketing Spin?

Timestamps: 02:38–30:36

The Launch & Hype Cycle

• Anthropic unveiled "Mythos," a powerful AI model, but only a select group of companies and institutions (e.g., Amazon, Microsoft, Apple, Alphabet, the Linux Foundation) were granted access.
• Anthropic claims the model is "so dangerous it can't possibly be placed in our hands" and is positioned as a major leap forward in foundational AI security and capability.

“A model so powerful and so dangerous it can’t possibly be placed in our hands.” – Alex Kantrowitz [05:15]

Dissecting the Evidence

• Skepticism about substance: Ranjan argues that little actual evidence has been made public, with Anthropic providing only limited, theoretical or unverified examples of unprecedented cybersecurity vulnerability detection.
• Consortium optics: The presence of a high-powered industry consortium is as much about optics and collective buy-in as it is about substantive risk or utility.

"It’s good branding… but I don’t like all this hype when you’re not actually able to see anything. Have some meetings, be careful—but you don’t need to, like, 'here is Mythos.' It sounds like an Avengers movie." – Ranjan Roy [09:06]

The Sandwich Email: PR or Prod?

• A now-viral anecdote (which spread rapidly through coordinated PR and embargoed releases) claimed that Mythos "broke out of containment" and emailed a researcher while he was "eating a sandwich in a park."
• Both agree this story was expertly seeded as a human, memorable hook, but serves marketing aims more than providing true transparency about risks.

"Within 20 minutes… both Anthropic and Sam Bowman, all of this was prepared. And then… everyone focuses on that sandwich detail… it was such a coordinated PR effort and it stuck." – Ranjan Roy [15:21]

Responsible Release or IPO Prepping?

• Debate over the motivation: is Anthropic exercising "responsible disclosure" due to genuine risk, or is this a calculated effort to inflate the company’s value ahead of a rumored IPO (Initial Public Offering)?

"Are we suffering from some sort of AI Derangement Syndrome… where we forget to ask what happens if it does work?" – Alex Kantrowitz [16:54]

• They agree that the move is "somewhere in the middle": it’s neither disaster-hype nor a complete nothing burger. The timing and communications theater are crucial to the perception and future market positioning.

Notable Quote:
"Gun to my head, it’s somewhere in the middle… not nothing, but not as grand as they say it is.” – Alex Kantrowitz [24:01]

2. The Rise of "Harnesses" and the Meta Harness Study

Timestamps: 30:36–33:20

• Stanford’s "meta harness" study suggests major improvements in model performance can come from the systems ("harnesses") built around AI models, not just the models themselves.
• Discussion about AI’s ecosystem: who can build the most efficient and controllable framework to drive performance?
• Internal podcast jokes about the "Harness Hive" community—mixed feelings about the term.

"You can change the harness around a fixed model and see a 6x performance gap on the same benchmark… So maybe Mythos won’t matter. It’s all about who’s got the best harness." – Ranjan Roy [32:00]

3. Violence Against AI: Data Centers and Executives Targeted

Timestamps: 35:13–46:20

Data Centers as Symbols

• Incidents of violence (including gunfire at politicians’ homes and Molotov cocktails at OpenAI CEO Sam Altman's residence) are increasing as some protest tech’s growing footprint, especially through data centers.
• Data centers are lightning rods—large, faceless, resource-consuming, and represent displacement/anxiety around AI’s role.

"These buildings can be faceless, they can be imposing… Tech’s interest in showing and delivering this technology despite the uncertainty it causes to people’s lives." – Alex Kantrowitz [38:27]

• Local and state opposition is growing, with legislatures (e.g., Maine) moving to ban or restrict data center builds.

Bigger Picture: Public Opinion & Political Ramifications

• AI is increasingly unpopular in polls. Unlike social media, politicians and citizens can physically stop AI’s progress by blocking infrastructure.
• Anticipate more regulation and resistance—AI faces a "public face" problem and lacks a mass-market hero/spokesperson.

"AI is going to be front and center. And it makes for such a good villain—the industry has not put the most likable people front and center representing the technology." – Ranjan Roy [41:28]

4. Case Study: Medvi and the “$1.8B One-Person AI Startup”

Timestamps: 49:28–55:21

• NYT’s much-discussed story on Medvi, a telehealth company run by two people claiming $1.8B in sales, built using AI for everything from website creation to fake doctor videos.
• Hosts question the revenue numbers and call out regulatory/ethical violations (deepfake doctors, misleading marketing, FDA warnings).
• The broader point: AI vastly amplifies one-person (or tiny team) capacity—for better or worse.

"It would have been great if they switched the tone a little bit: the Medvi story shows how a little AI and…whatever’s close to [scamming]…can get you to scale really quick." – Alex Kantrowitz [53:06]

5. Token Maxxing: The Silicon Valley Status Game

Timestamps: 56:02–61:19

• Internal competition at Meta—employees vie for “session immortal” or “token legend” status by burning the most AI tokens, tracked via a company-wide leaderboard.
• The hosts discuss the merits (who's actually using and learning AI) and drawbacks (perverse incentives to waste compute for status).
• Both note this practice may distort priorities and artificially inflate metrics—yet it's spreading to other tech companies.

"If it ever became important in terms of your review with your boss… the incentives become too screwed up and the whole thing becomes a little more corrupt, performative and weird." – Ranjan Roy [59:12]

6. AI IPOs: The Road Ahead

Timestamps: 61:19–62:24

• With revenue hype and escalating industry drama, expectation is that both Anthropic and OpenAI will IPO soon, bringing public scrutiny to their claims.
• Final joke: Will these "harness" companies even use law firms for their IPOs—or claim they're so advanced they don't need them?

Notable Quotes & Moments

• On PR Tactics and The Sandwich Story:
  "They are coordinating PR… so why do they want that to be the story?" – Ranjan Roy [15:13]

• On Data Centers as Lightning Rods:
  "I feel that these [data centers] are going to continue to be… a visual representation of what’s going on." – Ranjan Roy [38:12]

• On the Future of AI Super Apps and Control:
  "Do we take the most intelligent models we’ve built and keep them exclusive to our super apps or make them available to everyone?" – Alex Kantrowitz [19:27]

• On Corporate Incentives:
  "There’s so much incentive… to make it out to be [dangerous] and the way they rolled it out, I think it's been genius and I think it's just ahead of the IPO." – Ranjan Roy [24:13]

Key Timestamps for Reference

• Anthropic Mythos discussion & sandwich PR saga: 02:38–17:55
• Debate over marketing vs. reality: 19:27–30:36
• Harnesses/meta harness research: 30:36–33:20
• Violence targeting AI/data centers/Sam Altman: 35:13–46:52
• Medvi "billion-dollar AI startup" scrutiny: 49:28–55:21
• Token maxxing at Meta & AI usage leaderboards: 56:02–61:19

Tone & Takeaways

Throughout, Kantrowitz and Roy keep a skeptical, analytic tone. They’re not starry-eyed about AI nor simply antagonistic, but aim to parse real advancements from marketing smoke. The hosts highlight how AI’s reality today is as much about narrative management, power struggles, and symbolism as it is about technical progress.

For listeners: Expect a sharp, funny, and deeply informed deconstruction of tech news that cuts through both the hype and the hand-wringing.

For more detailed breakdowns of each section, refer to the timestamps.