Apr 13, 2026 · 1h 0m · big-technology
Anthropic’s Mythos Dilemma, Violence Against AI, Tokenmaxxing at Meta
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the Big Technology Podcast, Alex Kantrowitz and Ranjan Roy analyze the reality behind Anthropic's Mythos model, shifting business models across foundational AI labs, and the escalating physical, political, and regulatory backlash confronting the AI industry.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 48.2% of the talking time here. How this is scored →
Alex as informed peer 5.9 Guest teaching 3.4 Guest disagreement 3.9 Alex pushing back 3.4
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 44:49 Dismissal of TBPN marketing impact
Ranjan sharply dismisses Alex's proposal that OpenAI acquiring the TBPN hosts will fix AI's PR problem, arguing they only appeal to existing AI enthusiasts and will never persuade community activists or politicians.
Hardest push from Alex ▶ 8:23 Host challenges guest on safe release protocolsAlex directly confronts Ranjan's cynical dismissal of Anthropic's release strategy, challenging him to define what an alternative responsible rollout would look like if the safety risks were genuine.
Biggest teaching moment ▶ 30:00 Guest details Stanford meta-harness findingsRanjan clearly explains technical findings from Stanford demonstrating a 6x performance increase achieved solely through optimizing workflow harnesses around fixed models.
Alex holds their own ▶ 18:14 Host outlines proprietary super app shiftAlex synthesizes industry capital flows, API economics, and quotes from Martin Casado to construct a cogent thesis on why foundation model labs are reserving elite capabilities for first-party products.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| Anthropic Previews Mythos and the Glasswing Security Initiative | 6 | 3 | 4 | 3 | Alex frames Anthropic's Mythos announcement with reporting from the Wall Street Journal and community reactions from X, citing the prestigious consortium of partners. Ranjan pushes back against the unverified hype, explaining the origin of Project Glasswing's name and citing critical details from Tom's Hardware. | |
| Scrutinizing the Reality of Mythos's Reported Zero-Day Exploits | 7 | 2 | 5 | 5 | Alex directly challenges Ranjan's cynicism by asking how else Anthropic could responsibly preview dangerous capabilities, before detailing specific findings from the Tom's Hardware report on FFMPEG and Linux exploits. Ranjan argues that if the model were truly dangerous to humanity, the company should pause rather than race toward an IPO. | |
| Analyzing the Viral Sandwich Narrative and Coordinated PR Messaging | 5 | 4 | 6 | 4 | Ranjan maps out a precise timeline of tweets and releases to argue the viral park sandwich breakout story was orchestrated PR. Alex counters by probing whether they are both falling into AI derangement syndrome by refusing to consider that real technical progress actually occurred. | |
| The Rise of Proprietary Super Apps and the Race to IPO | 7 | 2 | 3 | 4 | Alex outlines his comprehensive theory that AI labs are shifting away from developer APIs to reserve their best intelligence for proprietary first-party super apps, citing Martin Casado. Ranjan agrees with the business logic while questioning why Anthropic would still share access through Project Glasswing if absolute exclusivity was the goal. | |
| First-Party App Friction and OpenAI's Anticipated Countermove | 6 | 2 | 4 | 3 | Alex highlights the growing channel conflict between first-party tools and third-party API clients like Cursor. Ranjan rejects the traditional super app terminology and playfully speculates on how Sam Altman and OpenAI will counter with their Spud model release. | |
| Stanford's Meta-Harness Study and System-Level AI Performance | 4 | 6 | 3 | 2 | Ranjan breaks down Stanford's meta-harness paper, explaining how surrounding architecture can drive a 6x performance increase on fixed foundation models. Alex concedes the functional premise while vocally expressing disdain for the terminology. | |
| Escalating Physical Violence Against AI Infrastructure and Executives | 6 | 3 | 2 | 2 | Alex reviews alarming reports of gunfire at an Indianapolis councilman's home and a Molotov cocktail thrown at Sam Altman's house. Both discuss how physical infrastructure and tech executives have become concrete targets for broader socioeconomic anxieties. | |
| Legislative Pushback, State Bans, and AI's Public Relations Crisis | 7 | 3 | 5 | 5 | Alex cites Maine's statewide data center ban and polling data showing AI's broad unpopularity, proposing that content marketers could help repair public perception. Ranjan forcefully disagrees with the idea that podcast personalities like the TBPN hosts could sway anti-data center activists or political leaders. | |
| The Medvi Controversy: Examining an AI-Powered Telehealth Scale-Up | 6 | 4 | 4 | 3 | Alex and Ranjan deconstruct the viral New York Times profile of Medvi, exploring how AI-generated ads, fake doctors, and automated drop-shipping allowed a two-person team to rapidly scale GLP-1 sales. Both criticize the uncritical reporting of inflated ARR numbers while recognizing the blueprint for scammy AI-leveraged businesses. | |
| Token Maxing, Claudeonomics, and Internal AI Gamification | 5 | 5 | 3 | 3 | Alex reports on Meta taking down its internal Claudeonomics token-burning leaderboard, questioning the perverse incentives of gamifying token usage. Ranjan shares firsthand experience from Writer, arguing that tracking heavy token usage is a legitimate indicator of genuine workflow experimentation when detached from performance reviews. |
Statements from this episode (17)
Anthropic previews Mythos model to around 50 critical infrastructure organizations
“The company is making a preview model of its new AI model called Mythos, available to about 50 companies and organizations that maintain critical infrastructure, including Amazon, Microsoft, Apple, Alphabet, and the Linux Foundation.”
Roy: OpenAI and Anthropic are in a death race to IPO first
“Right now, open AI and Anthropic are in like a death race to who can get out first in terms of their IPO.”
Kantrowitz: Anthropic extrapolated thousands of Mythos bugs from 198 manual reviews
“Ananthropic states it can't actually confirm all the thousands of bugs that Mythos claims to have found are actually critical security vulnerabilities. It's just extrapolated, extrapolated that number from having found, found it in around 90% of these 198 manu…”
Kantrowitz: Skeptics focus on tech failure and forget to ask what if it works
“Oftentimes skeptics can ask, like, what happens if it doesn't work? But sometimes you ask that so often you forget to ask what happens if it does work.”
Kantrowitz: OpenAI and Anthropic Are Shifting to Proprietary Super Apps
“Over the past three or four months, both of them have gravitated toward building a super app. Something that uses the most advanced intelligence to control your computer, that will, to help you get things done, to in some cases even build new software for you”
Kantrowitz: Anthropic Is Giving Microsoft Mythos Only to Test, Not Resell
“Anthropic isn't giving Microsoft mythos to sell through Azure. It's giving Microsoft mythos to test.”
Kantrowitz: AI labs will eventually have to choose between apps and APIs
“I think you will eventually have to make a bet on what your business is. It's very tough to sustain both for a while, and who do you want to have the best models in that case?”
Roy: No company will build a WeChat-style super app in the US
“I hate super app. I don't, it, no one's gonna be WeChat in the U.S.”
Roy: Anthropic and OpenAI model codenames are Mythos and Spud
“While Anthropik's codename for their kind of like incredible model is Mythos, OpenAI's codename internally for their next model is Spud.”
Roy: Stanford Meta-Harness Research Shows 6x Performance Gap on Fixed Models
“Now, Stanford came up with the idea of a meta harness. It's a harness over other harnesses. It's the idea that like the that you can change the harness around a fixed model and see a six x performance gap on the same benchmark model.”
Roy: Indianapolis has added 40 new data centers over recent years
“Apparently Indianapolis is, there's like a number of state tax incentives. They've like grown 40 new data centers over the last few years.”
Kantrowitz: Federal data center moratorium bill has no chance of passing
“And of course, there's this bill that Bernie and AOC introduced about a data center moratorium to, you know, be national and, you know, there's no chance of that passing”
Maine Lawmakers Advance Bill Halting Data Center Construction Until November
“Lawmakers in Maine greenlit the text of a bill this week to block data centers from being built in the state until November,”
Kantrowitz: Anthropic's Super Bowl Ad Was a Miss for the AI Category
“I think that, you know, in retrospect, their Superbowl ad, even though they were praised for, it was kind of a miss because it ended up bringing down the category as opposed to making people excited about AI.”
Roy: Medvi has not actually made $1.8 billion in sales
“Regular listeners will know of my hatred of ARR as a term. We have no idea what that means. They have not made 1.8 billion dollars.”
Meta Intranet Leaderboard 'Claudeonomics' Ranked Token Usage for 85,000 Employees
“The ranking set up by Meta, by a Meta employee on its intranet, use company data, measure how many tokens employees are burning through, dubbed Claudeonomics after the flagship product from Anthropic, the leaderboards, leaderboard aggregates AI usage from 85,0…”