Meta’s AI Comeback Moment, Claude Mythos | Diet TBPN
TL;DR
Meta may have finally hit its AI reset button with Muse Spark — after the Llama 4 stumble and the unreleased Behemoth, Meta launched its first major model in over a year, switched from open weights to closed, and the stock jumped about 7.5%-8% on relief that the company is back near the frontier.
The hosts frame Meta’s open-source era as always conditional, not ideological — citing John Luttig’s old thesis and Zuckerberg’s own comments, they argue Meta open-sourced Llama because it helped Facebook, Instagram, developer marketing, and platform strategy, but $10B-scale training runs and shareholder ROI were always the likely tipping point.
Muse Spark’s benchmark story is mixed, and that may actually be the point — the hosts call Meta’s chart a bit of a 'chart crime' because the blue highlight implies dominance, yet Muse Spark underperforms on some tests like ARC AGI 2 while beating rivals on others, which they read as at least a move away from pure benchmark gaming.
The funniest product moment was also the most revealing one: 'Malibu appropriate surf puns' — when the hosts tested Meta AI for a joke, it volunteered hyper-specific surf humor, then awkwardly denied any personalization, surfacing the exact tension Meta has to manage between 'personal superintelligence' and 'please don’t think we trained on your Instagram life.'
Anthropic’s Mythos reignited the old 'too dangerous to release' debate, but with a sharper cyber angle — the model is being shown first to about 50 critical-infrastructure partners including Apple, Google, Microsoft, Amazon, Nvidia, JP Morgan Chase, Cisco, CrowdStrike, and Palo Alto because it appears especially strong at finding zero-days and exploits.
The bigger theme is that frontier AI is getting more closed, more compute-constrained, and less visible to the public — between Meta shutting down its internal token leaderboard after reports of 60 trillion tokens in 30 days, Anthropic gating Mythos, and rumors of 10 trillion-parameter systems across labs, the hosts argue the best models may increasingly go first to whoever can pay or justify them.
The Breakdown
Meta’s Closed-Model Turn Feels Like the Real Headline
The show opens on Meta launching Muse Spark, its first major AI model in more than a year, with Alex Wang presenting a release the hosts treat as a genuine comeback moment. The big shift is strategic, not just technical: unlike Llama, Muse Spark is closed and will power Meta’s chatbot and app features, which immediately revives John Luttig’s old prediction that Meta would eventually stop carrying the open-source banner.
Why Meta Open-Sourced Llama Until It Didn’t
They walk through Zuckerberg’s logic in plain English: he hated being boxed in by Apple, wanted cheap performant AI to juice Facebook and Instagram, liked the optionality of AI assistants as a platform, and got good developer marketing from Llama. But the hosts stress that none of this meant permanent altruism — once capex pushes toward $10 billion training runs, proprietary data matters more, and shareholders start asking metaverse-style ROI questions, the open-source romance gets shaky fast.
Avocado, Mango, and the Benchmark 'Chart Crime'
The hosts connect Muse Spark to prior reporting that Meta had two internal efforts, Avocado for text and Mango for image/video, guessing this new release is probably the Avocado line. Then they pick apart Meta’s benchmark chart: Muse Spark is highlighted in blue and looks like the winner at first glance, but on closer inspection it beats some models and clearly trails others like on ARC AGI 2, which makes the presentation feel a little slippery even if not outright benchmark-hacky.
The 'Malibu Surf Puns' Demo Was More Interesting Than the Scores
Their first instinct was not to run evals but to ask for a joke, and Meta AI answered with a skeleton joke plus an offer for 'dad jokes, nerdy ones, or Malibu appropriate surf puns.' That weirdly specific phrase sends them down a rabbit hole about whether Meta is pulling from Instagram context, and the model’s repeated half-apology, half-denial becomes a perfect little case study in how awkward 'personal superintelligence' looks when it can’t admit why it sounds creepily personal.
The Internal Claude Token Dashboard Made Meta’s Cost Problem Hard to Ignore
They tie the launch to reporting that Meta employees had been burning through Claude at huge scale, with one internal dashboard reportedly showing more than 60 trillion tokens over 30 days before it was taken offline. Their read is simple: if thousands of engineers are maxing Claude, Meta has every incentive to convert that opex into capex by training in-house models and serving them cheaply on its own hardware.
Wall Street Likes 'Close to Frontier' More Than It Likes AI Vibes
The hosts say the market reaction — Meta up nearly 8% — is basically relief that the company’s billions in AI spend might amount to something concrete. There are still open questions about whether Meta will chase codegen aggressively or focus on consumer assistants across its family of apps, but a near-frontier model that can be deployed to billions of users is a much easier story to underwrite than vague AI ambition.
Anthropic’s Mythos Brings Back the 'Too Dangerous to Release' Playbook
The second half shifts to Anthropic’s Mythos, which is only being previewed with around 50 critical-infrastructure organizations because of its apparent strength at finding zero-days and vulnerabilities. The hosts note both the serious upside — preemptively hardening software at places like JP Morgan, Cisco, CrowdStrike, and Palo Alto — and the eye-rolling backlash, since the industry has heard 'too dangerous to release' before, all the way back to GPT-2.
The Real Pattern: Frontier Models Are Going Dark
They end on the broader market structure: compute is tight, the best models depreciate fast, distillation is a threat, and safety is now only one of several reasons labs may keep systems private. Whether it’s Meta going closed, Anthropic gating Mythos, or Elon Musk reportedly training multiple giant xAI systems up to 10 trillion parameters, the recurring message is that the most important models may no longer be the ones the public can casually touch.