Why did Anthropic hold back Mythos?
Mark Andreessen theorizes that Anthropic withheld its Mythos model primarily due to prohibitive compute costs rather than purely altruistic safety concerns. The delay may have served a dual purpose: conserving compute for the upcoming Opus 4.7 release while generating significant marketing buzz around perceived scarcity and responsibility.
Summary
The transcript discusses Mark Andreessen's theory about why Anthropic chose to hold back its Mythos AI model from commercial release. According to Andreessen, the primary driver was likely financial and logistical: Mythos was an extremely large and expensive model to serve, estimated at 10 to 20 times the token cost of Opus. With Opus 4.7 already in the pipeline, Anthropic may have made a strategic decision to withhold Mythos to conserve compute resources for the next generation model rather than strain infrastructure on an already-oversized release.
The speaker also notes that the withholding of Mythos generated an unintended — or perhaps intentional — marketing benefit. By framing the decision around safety and responsibility, particularly given that Mythos reportedly revealed previously unknown coding vulnerabilities, Anthropic created an impression of altruism and scarcity. Government officials and the broader public interpreted the decision as a bold, principled safety stance, turning the non-release into a major public relations event.
The speaker acknowledges that the altruistic framing may not be entirely cynical — Mythos genuinely did expose dormant security vulnerabilities in codebases, and giving organizations time to patch those issues before the model's capabilities became widely accessible is a legitimate safety rationale. Nonetheless, the speaker concludes that the compute economics likely made commercial release impractical regardless, suggesting the safety narrative and the business reality conveniently aligned in Anthropic's favor.
Key Insights
- Mark Andreessen argues that Mythos was 10 to 20 times more expensive per token than Opus, making it commercially unviable to serve at scale regardless of safety considerations.
- The speaker suggests Anthropic strategically withheld Mythos to preserve compute resources for the upcoming Opus 4.7 release rather than waste infrastructure on an oversized model.
- The decision to withhold Mythos inadvertently — or deliberately — created a massive marketing event, with government officials interpreting the non-release as a bold, altruistic safety stance.
- The speaker concedes that the safety rationale has genuine merit, as Mythos did surface previously unknown coding vulnerabilities, giving organizations time to patch dormant bugs before the model's capabilities spread.
- The speaker concludes that Anthropic's safety narrative and its business constraints likely aligned conveniently, making it impossible to determine how much of the decision was principled versus economically forced.
Topics
Transcript
[0:00] Mark Andreessen pointed out that one of the reasons why Anthropic might have wanted to hold back Mythos is they simply didn't have the compute to serve it. The model was huge and very expensive to serve, something like 10 or 20 times the token cost of say Opus. They knew Opus 4.7 was coming out, right? So, they hold it back knowing that they don't have the compute to serve it anyway, and they save their compute for the next iteration of Opus. And then, by holding it back, they create this impression of scarcity and altruism, and it turns into this gigantic marketing event for their [0:30] product because everyone in the government's like, "Oh, wow, they're…
Full transcript available for MurmurCast members
Sign Up to AccessMore from All-In Podcast
How Robots Learn: Much Slower Than Humans At First, Then Infinitely Scalable
The speaker discusses the evolution of learning in robots compared to humans, emphasizing that while humans learn efficiently with minimal data, robots require significantly more data to learn effectively. However, once a robot learns a task, that knowledge can be shared across all robots of that type, leading to infinite scalability.
Former Navy Seal Details First Autonomous Rescue Mission Near the Strait of Hormuz
A former Navy SEAL discusses a significant autonomous rescue mission involving a fully autonomous speedboat named Corsair, tasked with rescuing downed American pilots in the Strait of Hormuz. This mission represents a critical milestone in military operations, showcasing the ability to conduct rescues without further endangering soldiers.
GameStop CEO: “Why Does Everyone Want Us to Fail?”
GameStop CEO argues that the media and management are biased against GameStop's success, despite his financial commitment to the company. He highlights the disparity between overpaid management teams and those willing to risk their own capital.
Friedberg: Elon’s $17B Terafab Could Be the Greatest Chip Fab on Earth
The discussion highlights Elon Musk's ambitious vision for a semiconductor fabrication site, which could reduce U.S. dependence on Taiwan and China for chips. The potential success of this venture is tied to profits from Starlink and Musk's investment strategies.
Google’s AI Brain Drain, SpaceX's Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AI
The All-In podcast discusses Google's AI leadership exodus and restructuring, SpaceX's spectacular earnings with $7.8B revenue (up 92% YoY) and $2.6B in AI compute rental revenue, Airtable's acquisition by Bending Spoons for $1.28B (90% below peak valuation), and concerns about U.S. training data being sold to Chinese AI companies.