All-In PodcastWhy Anthropic's best model is locked up despite a $30B ramp
Mythos scored so well on cyber offense that Anthropic delayed its release; a 100-day hardening window did not stop Claude Code from driving a $30B run rate.
CHAPTERS
- 0:00 – 4:19
Brad Gerstner sits in for Friedberg: banter, “retard maxing,” and the show’s tone
The episode opens with Brad Gerstner joining as a guest bestie, plus the usual ribbing about moderation, intros, and personal habits. The group riffs on “retard maxing” (detachment/non-rumination), internet culture references, and setting an action-oriented mindset for the rest of the discussion.
- •Brad Gerstner introduced as the “fifth bestie” filling in for Friedberg
- •Back-and-forth about show format, moderation, and cutting “hot air”
- •Discussion of ‘retard maxing’/detachment and Elijah Long’s philosophy content
- •Cultural jokes and references (Dune, modern internet philosopher framing)
- 4:19 – 6:16
Anthropic withholds ‘Mythos’: claimed autonomous vuln-finding and exploit chaining
Jason frames Anthropic’s decision to delay releasing its new model ‘Mythos,’ citing security concerns: finding long-dormant vulnerabilities across major OSes and browsers. A clip from Dario and Anthropic staff emphasizes the model’s ability to discover and chain multiple vulnerabilities into sophisticated exploits.
- •Anthropic claims Mythos found thousands of vulnerabilities, including decades-old issues
- •Examples: OpenBSD, FFmpeg, Linux kernel and browser/OS bugs
- •Anthropic’s messaging: model is ‘as good as a professional’ at bug finding
- •Exploit chaining: combining 3–5 smaller vulnerabilities into a major compromise
- 6:16 – 9:51
Self-regulation vs fear-marketing: Brad praises ‘Project Glasswing’ and sandboxing
Brad argues Anthropic deserves credit for not ‘moving fast and breaking things’ and for organizing a coordinated 100-day effort to harden software before broad release. He positions this as a threshold moment toward AGI-grade models and a pragmatic alternative to government moratoriums.
- •Brad calls the move responsible: delay release to prevent widespread exploitation
- •Project Glasswing: coalition of major firms to find/fix vulnerabilities for ~100 days
- •Claimed shift to ‘sandboxing’ frontier models before GA release
- •Market forces + coordination with government, without heavy-handed regulation
- 9:51 – 15:00
Sacks’ two-part take: Anthropic’s ‘scare pattern,’ but cyber risk is plausibly real
Sacks critiques Anthropic’s history of dramatic safety studies used alongside releases (e.g., the prior ‘blackmail’ narrative). However, he concedes that cyber capability scaling with coding models is logically credible, advocating a pre-release patching window and preparation for an offense/defense arms race.
- •Sacks: Anthropic is good at ‘product releases’ and ‘scaring people’
- •References prior safety studies as reverse-engineered headline generators
- •Cyber angle deemed more legitimate: better coding → better bug discovery → exploit creation
- •Suggests a limited window before others (incl. Chinese models) catch up; use it to patch
- 15:00 – 19:45
Chamath calls it ‘mostly theater’: parallels to GPT-2 rollout and limits of patching
Chamath argues Anthropic/OpenAI have used staged-release safety narratives before (GPT-2 ‘end of days’), which proved overblown. He doubts the 100-day effort can materially harden the internet and suggests capable attackers could already do similar work with existing models like Opus.
- •Compares to GPT-2 (2019) staged rollout that ended as a ‘nothing burger’
- •Claims if exploits are easy to find, fixing them could take ‘years’ not months
- •Belief that sophisticated hackers could already use existing Claude models effectively
- •Credits Anthropic’s go-to-market execution despite skepticism of the framing
- 19:45 – 24:04
The ‘100 days’ debate: what can realistically be hardened—and why it still matters
The group argues about whether 100 days can produce meaningful security outcomes, with jokes underscoring the seriousness of mass vulnerability exposure (browser history, sensitive data). Brad and Sacks emphasize that even partial patching and coordinated attention can reduce harm during the transition to stronger models.
- •Question of practical impact: number of PRs/patches that can land in 100 days
- •Brad: even preventing a subset of catastrophic leaks is worth the delay
- •Sacks: regardless of theater, risk must be treated as real; coordinated patching helps
- •Broader point: AI is accelerating code production, so security must become embedded
- 24:04 – 32:06
OpenClaw ‘ankled’: subscription clampdown, API pricing, and agent competition accusations
Sacks and Jason unpack claims that Anthropic cut off OpenClaw’s access by disallowing heavy agent usage via flat-rate subscriptions, forcing a move to usage-priced APIs. The conversation escalates into whether this is normal pricing rationalization or a strategic move to suppress an open-source agent harness ahead of Anthropic’s own agent product.
- •OpenClaw users allegedly consumed far more tokens than typical $200/mo subscribers
- •Anthropic moved heavy usage to API billing (metered pricing)
- •Timing controversy: access change followed by Anthropic’s agent product announcement
- •Debate over antitrust concepts: bundling, price discrimination, dominant share in coding
- 32:06 – 36:25
Is Anthropic dominant in AI coding? Market-share semantics vs TAM reality
Sacks pushes for a direct answer on whether Anthropic holds dominant share in ‘coding tokens,’ with Jason and Chamath suggesting ~50–60% while Brad disputes dominance by pointing to the broader TAM and rapid market shifts. Chamath argues AI coding is still a small slice of total software creation and that enterprise-grade autonomy remains limited due to tech debt.
- •Sacks: >50% share of coding tokens implies dominance (at least currently)
- •Brad: dominance claims are hard in a fast-changing market; TAM framing matters
- •Chamath: AI-enabled coding still small portion of overall market; enterprise automation not ‘out-of-the-box’
- •Discussion of legacy tech debt (COBOL/Fortran) and slow enterprise migration
- 36:25 – 42:18
Open source vs frontier models: Bittensor/Bridges and distributed training as disruption
Jason argues open-source coding tools and incentive-driven networks (e.g., Bittensor subnets) are rapidly closing the gap, citing fast progress toward Claude-like performance. Chamath agrees open-source pretraining/orchestration could become a major disruptive path if capital markets tire of funding mega-model training—while maintaining that enterprises won’t outsource proprietary production code to open projects.
- •Jason cites Bittensor subnet ‘Bridges AI’ reaching ~80% of Claude capability quickly
- •Thesis: open source + crypto incentives can accelerate improvements and undercut frontier pricing
- •Chamath: distributed training/orchestration is a real orthogonal threat to mega-capex models
- •Caveat: major enterprises won’t open-source re-engineering of proprietary production codebases
- 42:18 – 49:26
Anthropic’s $30B run rate: fastest ramp ever, and the ‘TAM for intelligence’
Jason presents the headline numbers—Anthropic reaching an estimated $30B run-rate after rapid enterprise adoption, including many $1M+ annual customers. Brad contextualizes the growth as evidence of a near-infinite ‘intelligence TAM,’ compute constraint-driven throttling, and an inflection where AI shifts from IT budget to labor augmentation/replacement.
- •Timeline: API monetization → $1B run rate → Claude Code catalyst → $30B run rate claim
- •Enterprise adoption: 1,000+ customers reportedly paying $1M+ annually
- •Brad: evidence that revenue can scale exponentially alongside model capability gains
- •Compute constraints limit growth; intelligence unit-cost falling (Jevons-style dynamic)
- 49:26 – 57:59
Profitability and margins: gross vs net revenue, inference cost declines, ‘accidental profitability’
Chamath warns the conversation is still stuck at gross/net revenue and run-rate talk, far from steady-state cash flow clarity. Brad counters that compute is a largely fixed near-term input, inference costs are falling sharply, and margins may be improving rapidly—suggesting burn could be lower than many assume.
- •Chamath: market still debating gross vs net; profitability discussion is premature
- •Brad: compute is key cost; ramping revenue against constrained compute can lift margins
- •Rumors cited: ~50–60% gross margin; inference cost down ~90% YoY (claimed)
- •Debate: whether incumbents (Meta/Google) will turn AI into a capital/compute moat
- 57:59 – 1:10:01
Vibe shift: Anthropic ‘ripping,’ OpenAI ‘reeling’—but Brad says don’t count OpenAI out
Jason highlights secondary-market signals and talent migration narratives, suggesting OpenAI is facing strategic and cultural pressure. Brad argues OpenAI remains firmly on the wave, citing upcoming ‘Spud’ model previews and Codex momentum, emphasizing the market isn’t zero-sum and that multiple US frontier labs can win.
- •Jason: OpenAI facing employee churn, strategy questions; Anthropic trading higher in secondaries
- •Brad: Anthropic’s focus (coding) produced a 90-day surge; but OpenAI’s next model could match
- •Claim: ‘Spud’ is being previewed and is on par with Mythos (per Brad’s conversations)
- •Broader frame: ‘Team America’ benefits from multiple competitive frontier labs
- 1:10:01 – 1:29:17
Geopolitics: Iran ceasefire, Israel’s influence, market impact, and US public opinion
The show closes with a wide-ranging discussion on the Iran conflict, ceasefire negotiations, and how markets are pricing the risk. Jason raises concerns about Israeli influence on US policy and the domestic backlash; Sacks and Brad focus on outcomes, de-escalation value, and Israeli leaders acknowledging declining US support.
- •Sacks: avoids speaking as White House; supports ceasefire and de-escalation dynamics
- •Brad: market drawdown modest vs tariff episode; optimism if ‘landing the plane’ on multiple conflicts
- •Jason: argues Americans perceive outsized Netanyahu influence; links to domestic political/antisemitism tensions
- •Bennett poll tweet: Israeli politicians note worsening US public sentiment; calls to repair relationship
