In recent days, the safety arguments advanced by Dario Amodei, CEO of Anthropic, have been challenged from two distinct angles. On one side, Jensen Huang, CEO of NVIDIA, publicly pushed back against Amodei’s claim that advanced AI could eliminate half of entry-level white-collar jobs. On the other, results published by AISI indicate that the publicly available GPT-5.5 model performed nearly on par with a restricted model called Mythos on a set of expert cyber tasks.
Jensen Huang’s remarks and their significance
Jensen Huang urged CEOs to stop speaking like prophets about imminent labor collapse, implicitly questioning the more apocalyptic scenarios voiced by Amodei and others. That pushback matters because prominent industry leaders shape both public debate and the reactions of investors and regulators.
AISI’s comparison: GPT-5.5 versus Mythos on cyber tasks
According to AISI’s reported measurements, the public GPT-5.5 achieved a 71.4% score on a series of expert cyber security tasks, while Mythos scored 68.6%. These numbers suggest the public model’s performance in the tested tasks approaches that of the more restricted system. It is important to note the brief summary did not include detailed methodology or a full task list, so interpreting these results requires caution.
Why this matters for the “too dangerous to release” argument
Part of Amodei’s and others’ case for withholding certain models rests on the premise that the most capable systems must be tightly controlled to prevent misuse. If a public model approximates a closed model’s capabilities on professional tasks that motivated access restrictions, the argument for blanket non-release weakens. That dynamic is amplified when industry figures such as Jensen Huang openly dispute catastrophic labor predictions.
Consequences and open questions
Taken together, the two developments raise fresh questions about model releases, access controls and how safety is communicated:
- How can the community consistently measure and compare risks and capabilities when methodologies are not always transparent?
- To what extent does the rhetoric of industry leaders influence regulatory and policy responses in AI?
- What are the implications if public models quickly catch up to previously restricted systems, especially for preventing harmful uses?
These developments do not rule out that some models could pose genuine dangers, but they highlight that debates over access and safety narratives are evolving. Robust, transparent comparisons and clear methodologies will be essential for the next stage of the discussion.
