Why did people miss the point on Mythos?

Anthropic's Claude Mythos has ignited a firestorm of criticism, with outlets like The Guardian and computer science professor Cal Newport labeling it a hype-driven stunt to secure funding. Newport argued that Mythos doesn't represent a significant leap in detecting vulnerabilities compared to previous models, while the YouTube channel Internet of Bugs called it a "$x00 million marketing stunt." However, a detailed analysis on LessWrong by draganover contends that the real point is being missed: Mythos demonstrates that scaling laws are holding, leading to the inevitable emergence of dangerous capabilities like advanced cybersecurity exploits. The author warns that this acceleration toward unpredictable risks is a larger story buried under the noise.

The post also addresses the tendency of critics to cherry-pick evidence, such as a tweet from a cybersecurity org claiming smaller models found the same bugs as Mythos. The author likens this to isolating a needle in a haystack and giving it to a child, arguing that the hard part is locating the vulnerability, not analyzing it post-hoc. The author calls for better protocols for announcing emergent capabilities, emphasizing the need to build public sympathy for AI safety rather than squandering it. They acknowledge Anthropic's conflict of interest but urge the community to focus on the broader implications of scaling laws, not just the immediate hype.