Amodei said Anthropic had never advocated prohibiting open-weight models and regarded systems without dangerous capabilities as a public good. His intervention followed criticism of the company for declining to join a technology industry letter opposing broad restrictions on openly distributed AI.
The dispute has divided Silicon Valley as policymakers examine whether advanced models developed in China could expose businesses, government agencies and critical infrastructure to security threats. Nvidia chief executive Jensen Huang, Microsoft chief executive Satya Nadella and leaders from Meta, Google, OpenAI and other technology companies have defended continued access to open systems.
Open-weight models allow developers to download the numerical parameters produced during training. They can then run, modify and fine-tune the software independently of its original creator. The term differs from open-source software because developers may receive the weights without gaining access to training data, source code or full details of how the model was built.
Amodei argued that this distinction matters when assessing safety. Publishing a model’s weights can make it impossible for the developer to withdraw the system, patch safeguards or stop malicious users from modifying its behaviour. Closed models, by contrast, usually operate through controlled online services whose providers can monitor usage and update protections.
The Anthropic chief nevertheless warned that outright bans would be difficult to enforce and could damage legitimate research, competition and commercial development. Instead, he called for models to be evaluated against clearly defined capability thresholds before their weights are released.
Testing could examine whether a system can assist with biological weapons, sophisticated cyberattacks, autonomous operations or other activities carrying national security risks. Restrictions would then apply to models that cross specified danger levels rather than to every openly distributed system.
Amodei also urged tighter controls on the supply of advanced semiconductors to authoritarian governments and stronger action against efforts to copy proprietary models through distillation. Distillation involves training a smaller or competing system using outputs produced by a more powerful model. It is widely used for efficiency, although disputes arise when companies suspect it has been employed to reproduce protected technology without permission.
The debate intensified after powerful models developed by Chinese companies gained international users because of their performance, low operating costs and downloadable weights. Their expansion has challenged the business strategies of companies that charge customers for access to proprietary systems.
Supporters of open weights argue that access gives start-ups, universities and smaller countries an opportunity to build AI services without depending on a handful of technology groups. Organisations can operate models on their own computers, keep sensitive information within local systems and adapt software for specialised languages or industries.
They also maintain that open access enables independent researchers to identify weaknesses, measure bias and develop defences. Restricting distribution, they warn, could concentrate technological influence among a small number of corporations and encourage developers outside the United States to adopt competing platforms.
Huang has framed open AI as central to maintaining technological leadership. Nvidia benefits from a broad market in which companies, governments and researchers purchase chips to train or operate models, regardless of whether the software is controlled by a single provider.
Critics counter that downloadable models can have safety protections removed and may be used without oversight. Once weights have spread across international networks, sanctions or domestic regulations may have limited effect. Governments could also struggle to determine which altered version was responsible for harmful activity.
Anthropic’s refusal to sign the industry letter prompted allegations that it was seeking regulations favourable to providers of closed models. Amodei denied that commercial interests explained the company’s position and challenged the argument that openness automatically improves security.
The disagreement reflects competing approaches to AI governance. One favours broad availability, technical transparency and market competition. The other seeks controls based on the possibility that increasingly capable systems could provide dangerous expertise to individuals, criminal groups or hostile governments.
Any testing regime would face difficult questions over who sets the thresholds, which laboratories conduct evaluations and whether assessments remain valid after a model has been modified. Smaller developers may also struggle with the cost of compliance unless governments create shared testing facilities or proportionate rules.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
