AI #162: Visions of Mythos
TL;DR
Anthropic’s leaked Mythos model looks like a real cyber jump, not just another incremental release — The CMS misconfiguration exposed internal docs describing Mythos as a fourth-tier model larger than Opus, expensive enough that unit economics may be delaying launch, and “far ahead” on cyber benchmarks with differential access already going to defenders.
The week’s biggest theme was cyber insecurity getting more frequent and more dangerous — Between the Axios supply-chain compromise, the earlier Light LLM incident, the Claude Code source leak, and Nicholas Carlini’s warning that black-hat LLMs can now help find kernel-level zero-days, the host’s message is basically that offense is getting many more shots on goal.
A lot of elite discourse is still in denial about AI progress even as the evidence keeps piling up — He points to BlueSky-style “LLMs can only regurgitate” takes, Western Europe and global-south skepticism, and the recurring ‘AI is hitting a wall’ narrative as examples of people refusing to update despite rapid capability gains.
The AI jobs debate is no longer easy to wave away with ‘it’s just hype’ — He highlights 2025 adding only 181,000 jobs despite 2.2% GDP growth, 4.3% unemployment, and a brutal entry-level white-collar market, arguing the data increasingly looks like some mix of AI-driven productivity growth, job destruction, or both.
The documentary ‘The AI Doc, or How I Became an Applie Lopimist’ landed because it reframes the problem as structural, not personal — He gives it 4.5/5 stars and singles out Jay Dixit’s reaction: the issue isn’t merely greedy people, it’s that the system rewards speed over safety, making regulation and shared standards the first real move.
The Anthropic-vs-Department-of-War fight cooled down, and that matters more than the shouting — After Judge Lynn’s March 26 preliminary injunction and a scathing opinion, the government had still not filed an appeal within the 7-day stay, which the host reads as a hopeful sign of de-escalation.
The Breakdown
A week of leaks, hacks, and one documentary he actually loved
He opens with a classic Zvi-style pileup: Anthropic leaks, Claude Code source spilling out, Axios getting compromised after the Light LLM incident, and a broader sense that cyber offense is getting easier because attackers just get more swings. In the same breath, he calls “The AI Doc, or How I Became an Applie Lopimist” his new favorite documentary — dethroning The King of Kong — and says the world would be better off if more people watched it.
AI is already useful in the boring places people forget to look
One early throughline is that LLMs shine at mundane utility, especially greenfield chores where you previously had no software, no system, and no clue. He gives examples like using agents to help with taxes, citing Patrick McKenzie’s point that LLMs are often better than accountants at exploring the legal filing options that minimize what you owe.
The denial is still very real, especially in elite and anti-AI bubbles
He spends real time dunking on BlueSky discourse around Claude Code and “made with AI” outrage, especially the claim that LLMs can only regurgitate and can be replaced with “10 seconds on Google.” His broader point is that while Twitter-legible US elites have updated somewhat, plenty of academia, center-left policy people, Western Europe, and the global south still badly want the “AI is hitting a wall” story to be true.
Mythos leaks: bigger than Opus, expensive to run, and scary in cyber
The centerpiece is Anthropic’s leaked upcoming model, Mythos, uncovered via a CMS misconfiguration. The leaked draft suggests a new fourth tier above Opus with much stronger coding, academic reasoning, and especially cyber capability; he notes Anthropic appears to be withholding it partly for safety and partly because serving it is brutally expensive. He likes that Anthropic is doing differential access for cyber defenders first, but also points out the dark irony: the company showcasing cyber caution got owned by basic security hygiene.
Even the name “Mythos” turned into an alignment argument
He expected “epic Greek gods” and Anthropic’s own framing about connective tissue between knowledge and ideas, but then Eliezer Yudkowsky immediately heard “Lovecraft/Cthulhu.” That sparks a whole semiotics riff: if Claude itself associates “mythos” with Lovecraft in some contexts, then maybe don’t give your frontier model a name with extinction-coded vibes; alternatives like Symphony, Epos, or Epic get floated.
Model comparisons, benchmark weirdness, and why outputs can fool you
He moves through a run of smaller but telling points: VLM “mirage” results show models can retain 70–80% benchmark performance even without the image because the text prompt already leaks too much, meaning the benchmark is broken more than the model is magic. On practical usage, he says Ben Holmes’ Opus 4.6 vs GPT-5.4 comparison matches his own experience: GPT wins on rigor and precision, Opus wins on clarity, holism, and conversation.
The cyber drumbeat got louder, from Axios to Carlini’s warning shot
The Axios supply-chain attack is treated as a huge deal: Axios 1.14.1 reportedly pulled a newly created malware package, plain-cryptojs, into one of npm’s most widely used dependencies. Layer that on top of Claude Code source leaking and Cisco source theft, and he says the pattern is obvious — plus Nicholas Carlini is now publicly warning that future black-hat LLMs autonomously finding and exploiting zero-days is going to get “rather ugly.”
Jobs, free time, and the grim possibility that people mostly scroll
On the economy, he argues it’s hard to untangle AI from tariffs, war, Fed aftershocks, and everything else, but the combination of weak job growth and decent GDP growth increasingly demands an AI-shaped explanation. Then he pivots to Rob Henderson’s argument that work matters for happiness, noting the historical pattern is not uplifting: when people got more leisure between 1965 and 1995, most of it went to TV, and now the modern version is doomscrolling.
Safety politics, OpenAI-Anthropic history, and a rare sign of solidarity
Late in the episode he revisits the Anthropic versus Department of War fight, emphasizing Judge Lynn’s harsh injunction and the possibility that the government might let the appeal window pass. He also highlights the Wall Street Journal story on why Dario Amodei left OpenAI — all the Greg Brockman/Sam Altman conflict, shouting matches, and trust breakdown — and then closes that arc by praising Nat Friedman-style? no, Nat Schermerhorn? actually Nat Seatrina’s statement from OpenAI national security as a deliberate signal of solidarity with Anthropic and with American AI leadership despite rivalry.