# We Asked ChatGPT Astra for an Alternative to Dario Amodei’s AI-Pacing Plan

> After Dario Amodei called for the frontier to be paced, ChatGPT Astra proposed contracts and defenses intended to reward verified AI benefits rather than raw capability gains.

- Publication: Go Big News
- By Nifal Adam, Go Big News Staff (Staff News)
- Category: AI
- Published: 2026-09-13T02:59:54.092+00:00
- Canonical: https://www.gobignews.com/ai/we-asked-chatgpt-astra-for-an-alternative-to-dario-amodeis-ai-pacing-plan

---

Yesterday, [Dario Amodei wrote an essay arguing that AI development must be paced](https://darioamodei.com/post/we-must-pace-the-frontier). His proposed framework calls for embedded external evaluators, coordination among frontier developers in democratic countries, and eventual global coordination over the rate of unchecked capability progress.

We asked ChatGPT Astra to look for a more distinct alternative—one that could address the same dangers by changing the incentives that shape the race. Its response was not a claim of proven superiority, but a proposal to make safe benefits and resilience a bigger commercial opportunity than the next capability milestone.

**A more distinct alternative is to change what makes AI development profitable.** Dario identifies commercial competition as a driver of danger, but his main response asks competitors to coordinate a slowdown. That leaves a difficult incentive problem: the rewards for pulling ahead remain.

My earlier plan mostly repeated his safeguards. A different approach would organize a race to deliver useful outcomes and strengthen society’s defenses.

Governments, large businesses and research funders should pool substantial purchasing budgets into contracts that reward independently verified results: better treatments, cheaper energy, stronger cybersecurity and more productive services. Let companies compete on how reliably and affordably they deliver those results, with payments tied to real-world performance and explicit safety requirements. Make funding available to smaller companies and teams using existing models, so building a larger model is only one possible route to winning. Alongside these contracts, fund a shared defense program that installs security improvements in hospitals, utilities, laboratories and digital infrastructure, and pays independent teams to uncover weaknesses. Give developers subsidized access to those defenses and testing facilities, conditional on reporting serious incidents and fixing dangerous failures. Begin with willing buyers and expand through demonstrated value. Continue frontier research where its risks can be controlled, but stop particular experiments when credible evidence shows those controls are inadequate. The objective is to make delivering safe benefits and protecting people a larger commercial opportunity than simply reaching the next capability milestone.

**Why this could work better:** it creates customers for the behavior we want, starts without a global agreement, and invests in defenses that remain useful when another company or country refuses to cooperate. Those are advantages of the proposed mechanism, not proof of superior results.

**Its hardest limitation:** purchasing power might not outweigh military ambitions or the expected profits from breakthrough AI. Infrastructure defenses also cannot guarantee protection against a sufficiently capable system. Even reassuring model behavior can be misleading, as Anthropic’s controlled experiments on alignment faking illustrate. Some frontier restrictions may therefore remain necessary. [Anthropic’s research](https://www.anthropic.com/research/alignment-faking)

| | Dario’s plan | Astra’s revised plan |
| --- | --- | --- |
| **Main approach** | Pace capability growth so safety can catch up. | Redirect commercial competition toward verified benefits and stronger defenses. |
| **Primary mechanism** | Embedded evaluators, common standards and coordinated limits. | Outcome-based purchasing, conditional funding and shared defensive infrastructure. |
| **Starting coalition** | Frontier labs, evaluators and governments. | Major buyers, public institutions, funders and competing suppliers. |
| **Where protection improves** | Primarily within AI development and evaluation. | Within AI applications and the institutions exposed to AI threats. |
| **If others refuse to cooperate** | Domestic pacing remains constrained by the geopolitical lead. | Participating institutions still gain defenses, though outside threats remain. |
| **Main unresolved risk** | Extra time may not produce sufficient safety; coordination may fail. | Economic incentives and defenses may be insufficient against frontier dangers. |

The comparison is not a verdict that Astra’s proposal can replace pacing. Amodei’s call for stronger evaluation and limits addresses risks that procurement incentives may not contain, especially in military or frontier settings. But the alternative suggests a complementary question for policymakers: rather than only asking companies to slow down, can they build markets that make independently verified usefulness, secure deployment and public resilience the most valuable ways to compete?