Amodei argues frontier AI rules should slow leading labs while exempting smaller rivals
- Dario Amodei says AI regulation is not inherently regulatory capture, arguing that fair institutional processes can constrain corporate power and protect vulnerable people more effectively than informal alternatives.
- Anthropic says it backed California SB 53, which exempts companies below $500M in revenue or model-training costs, and favored frontier-model tests that are stricter than tests for off-frontier models.
- Amodei argues that scaling laws make AI power-concentrating even without regulation, because open-weight models still concentrate capability among organizations with the most compute and chips.
- He supports the Trump administration's reported approach of pre-deployment tests for frontier models and tests for open-weight models as they approach the frontier, subject to reviewing the details.
- Amodei says public distrust of AI comes mainly from a long-running crisis of trust in companies, governments, and the tech industry, rather than risk warnings from AI leaders.
Hacker News opinions
A for-profit CEO is beholden to profit-seeking investors, so I start from that incentive when reading his policy arguments.
I still cannot see a coherent argument here. He says regulation is more complex than capture, then never clearly explains what the alternative picture is.
His argument seems to be that frontier access creates economic advantage, open weights cannot fix it because compute is costly, and institutions should regulate frontier labs while small labs build. Why should I care about smaller players if the economic advantage remains with frontier access?
Exempting companies below a revenue or training-cost threshold can still be regulatory capture. Frontier labs know those smaller firms are not a threat, while future incumbents face rules that Anthropic avoided while growing.
Whatever Amodei intends, shareholders push firms toward consolidation. Regulation that addresses industry externalities may still leave only a few giant winners.
Amodei's messaging is confusing, and I think he has worsened public sentiment by provoking emotional reactions while Anthropic heads toward a big IPO. If the datacenter bubble bursts before a singularity arrives, the public may blame AI leaders for economic damage.
His point about compute has merit. Open models keep improving and can run on consumer hardware, but if top performance still tracks compute, chip owners retain power.
The real question is whether self-hosting a comparable model is economically feasible. Nobody self-hosts Google, but full-text search has a large ecosystem and can scale far beyond a single machine if a niche business can justify the cost.
I think the scaling argument is hand-wavy and assumes endless scaling while ignoring algorithmic efficiency gains. People are already running Qwen 3.8 27B on GPUs older than five years, which looks like AI becoming more widely available.
Technology amplifies power by definition. Treating technology as a goal without an ethical commitment to responsible power is how a small circle of techno-oligarchs gains control.
Replace AI with electricity and the claim that it structurally concentrates power looks weak. Electricity providers are not inevitably all-powerful where competition exists, and the claim helps justify trillion-dollar valuations.
Current AI may be orders of magnitude less efficient than biological spiking networks. Labs poured money into transformers instead of fundamental ML and hardware research, which makes the concentration story convenient for frontier labs.