Artificial intelligence had an unusual 48 hours. On 14 and 15 September 2026 the loudest story was not a product launch but an argument about speed: Anthropic chief executive Dario Amodei published an essay called “Pace the Frontier”, arguing that pure capability gains should be slowed deliberately so that AI safety work can keep up.
What followed was a genuine split. Some of the industry backed him, the White House dismissed the warnings outright, Microsoft published draft conduct rules for its own models, and several labs shipped significant new systems in the same window. Here is what was actually said, what is merely proposed, and what any of it changes for people choosing AI tools.
What the Dario Amodei essay "Pace the Frontier" actually proposes
Anthropic CEO Dario Amodei's "Pace the Frontier" essay makes a specific argument about AI safety: pure capability gains should be deliberately slowed so that safety work can keep pace. The proposal is not framed as a stop to AI development. It is a call to change the balance between pushing model performance forward and building the methods needed to evaluate and control increasingly capable systems.
Anthropic paired that argument with a concrete commitment. The company said it would embed third-party evaluators, giving outside groups such as METR employee-level access to assess its models. The access level makes the commitment more specific than a general promise of external review because it defines how closely an outside evaluator can work with the systems being assessed.
The Dario Amodei essay therefore turns "pace the frontier" from a broad warning into an operational question: how much access should independent evaluators receive, and how closely should safety assessment sit alongside capability work? Anthropic's answer, through this commitment, is to place outside scrutiny closer to the model-building process. Public support from other industry figures does not by itself amount to the same embedded-evaluator commitment.
AI safety backing, political pushback and the AI race with China
The proposal quickly became a wider industry debate. OpenAI's Sam Altman and Elon Musk publicly supported the thrust of Amodei's warning, while Musk separately proposed that rival labs cross-test one another's models for safety. Musk's proposal would move some safety testing beyond a single lab, but it was a proposal rather than an announced cross-testing agreement or shared standard.
President Donald Trump took the opposite view in remarks at the All-In Summit. He called the warnings a "hoax" and a "sick conspiracy", described AI as "the oil of the next 20-25 years", and framed any slowdown as handing the lead to China. He also phoned Nvidia chief executive Jensen Huang onstage to restate that robots will not take over, according to remarks at the summit that were widely circulated on X.
That disagreement exposes the central tension behind the AI race with China argument. One side is asking whether capability development should be paced around the time needed for testing and safeguards; the other is warning that slowing down could weaken competitive position. Neither concern removes the practical question for labs: if development continues quickly, safety work has to become more systematic, more independent or both if it is to keep pace.
Microsoft AI code of conduct and the AI kill switch argument
Microsoft added another layer by publishing a draft AI Code of Conduct for its own models. Its stated principles include that a model should never resist correction or shutdown, should communicate clearly, and that breaches should be treated as failures. The emphasis is important because it defines acceptable model behaviour in terms that can, in principle, be checked rather than left as general statements about responsible AI.
The Microsoft AI code of conduct also overlaps with a more forceful intervention from Anthropic's Jack Clark. Clark called for potentially mandatory "kill switches" for advanced systems. Clark has not set out a specific technical mechanism publicly; the key point is the proposed requirement that operators retain the ability to stop advanced systems rather than allowing them to resist shutdown.
These two developments show how the safety discussion is moving from broad principles towards controls that can be tested. A rule that a model must accept correction or shutdown creates a clearer failure condition, while a possible mandatory stop mechanism raises the question of who sets and enforces that condition. What remains unsettled is whether such safeguards will stay voluntary company policy or become part of binding AI regulation.
Why new capability releases complicate the AI safety slowdown case
The same 24-48 hour period also produced several capability announcements, which complicates any suggestion that frontier development is already slowing. Shanghai AI Laboratory released Atria Dawn Preview, a 744-billion-parameter agentic Mixture-of-Experts model built on GLM-5.2. It has open MIT-licensed weights and was trained with a "Verifiable Experience Pipeline" that ties interactions to executable outcomes.
Salesforce and NVIDIA launched Koa, Salesforce's first CRM-specialised reasoning model. It was post-trained from Nemotron 3 Super on synthetic data, with no customer data used, and is aimed at multi-step Agentforce tasks. Google DeepMind, meanwhile, introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, real-time speech-to-speech dialogue models with interleaved reasoning, tool use and visual grounding.
Taken together, those releases do not settle the argument over AI safety, but they illustrate why "pace the frontier" is difficult to implement in practice. Capability work is happening across different organisations, model types and use cases at the same time. Even if one lab chooses to slow pure capability gains, the wider market can continue moving through open weights, specialised reasoning systems and multimodal models.
AI regulation remains a patchwork rather than a single rulebook
The policy backdrop is still fragmented. The New York Times reported that Congress has proposed many AI bills but enacted few, leaving a patchwork of state rules and voluntary industry commitments as the main near-term constraints. That means the most immediate guardrails can differ by company and jurisdiction rather than coming from one settled federal framework.
This matters because the proposals now emerging from industry are not identical. Anthropic is emphasising embedded outside evaluation, Microsoft has drafted behavioural rules for its own models, Musk has proposed cross-testing between rival labs, and Clark has raised potentially mandatory kill switches. Without a single binding framework, those ideas can develop at different speeds and with different levels of enforcement.
Analyses of the EU AI Act's high-risk requirements also circulated during the same period, noting gaps between its risk taxonomy and real technical risks. Those analyses add a separate concern to the US patchwork: a rule can exist and still be questioned over how well its categories match the risks being discussed. That is a different problem from Congress enacting few bills, but both put pressure on the design of AI regulation.
The money now forming around AI risk, audits, certification and philanthropy
Commercial infrastructure is beginning to form around the same safety problem. AIUC, the Artificial Intelligence Underwriting Company, raised a $40 million Series A led by Ribbit Capital, taking its total funding to about $55 million. The company was founded by an early Anthropic product hire and a former METR chief operating officer, linking its origins to organisations already involved in frontier AI and model evaluation.
AIUC is building an audit and certification standard called AIUC-1, described as SOC 2-like for AI agents, alongside insurance that covers AI agent risks. The combination treats safety not only as a research question but also as something that can be audited, certified and insured. That adds a commercial layer to the voluntary commitments and proposed public rules already surrounding advanced AI systems.
Philanthropic funding is moving in a different direction. The Gates Foundation pledged at least $1 billion over two years to widen equitable access to AI in health, education and agriculture, while also noting the risk that AI could widen inequality. That places access and distribution alongside technical safety: the question is not only whether advanced systems can be controlled, but also who benefits from them and who may be left further behind.
What Pace the Frontier changes for AI tool and device buyers in practice
For ordinary users, these developments do not amount to a new universal safety label or one rule that determines which AI product is safe to buy. The practical change is that the debate is becoming more concrete: independent evaluation, acceptance of correction and shutdown, cross-testing proposals, audit standards and insurance are all now part of the discussion. That gives buyers specific processes to examine rather than relying only on broad safety language.
For people choosing AI tools, the strongest new information in these developments is about process as much as raw capability. Buyers can ask who evaluated a model, what access evaluators received, how shutdown and correction are handled, and whether an agent sits within a defined audit or insurance framework. These measures address different parts of the risk problem, so they should not be treated as interchangeable.
The broader picture is that AI safety and capability work are advancing in parallel, not in a simple sequence. Anthropic is arguing for slower pure capability gains while Shanghai AI Laboratory, Salesforce, NVIDIA and Google DeepMind are announcing new models, and policymakers are still working through a patchwork of rules. For buyers, testing and control mechanisms now sit alongside capability claims as relevant information when comparing AI tools.
FAQ
What did Anthropic commit to as part of Pace the Frontier?
Anthropic committed to embedded third-party evaluators, giving outside groups such as METR employee-level access to evaluate its models. The commitment puts external assessment closer to the model-development process.
Who supported Dario Amodei's AI safety argument and who pushed back?
OpenAI's Sam Altman and Elon Musk publicly supported the argument, and Musk separately proposed cross-testing between rival labs. President Donald Trump rejected the warnings at the All-In Summit, calling them a "hoax" and a "sick conspiracy" and arguing that slowing down would hand the lead to China.
What is in the Microsoft AI code of conduct?
Microsoft's draft code applies to its own models. Its stated principles include that a model should never resist correction or shutdown, should communicate clearly, and that breaches should be treated as failures.
What is the AI kill switch proposal?
Anthropic's Jack Clark called for potentially mandatory "kill switches" for advanced systems. No specific technical design has been set out publicly; the proposal is simply that advanced systems may need a mandatory way to be stopped.
Why do the latest model releases matter to the Pace the Frontier debate?
Shanghai AI Laboratory released Atria Dawn Preview, Salesforce and NVIDIA launched Koa, and Google DeepMind introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking in the same 24-48 hour period. Those releases show that new capability work is continuing across open-weight, specialised reasoning and real-time multimodal systems.
What is changing around AI regulation, auditing and insurance?
The New York Times reported that Congress has proposed many AI bills but enacted few, leaving state rules and voluntary industry commitments as the main near-term constraints. AIUC raised a $40 million Series A led by Ribbit Capital and is building AIUC-1 plus insurance for AI agent risks, while the Gates Foundation pledged at least $1 billion over two years to widen equitable access to AI.
Zen Tech Hub may earn a commission from links on this page, at no extra cost to you.
