Short answer: The alarming headline refers to a real March 2024 assessment written by Gladstone AI for review by the U.S. Department of State. Its authors warned that advanced AI could create severe national-security, weaponization and loss-of-control risks, and proposed government thresholds that could restrict the largest training runs. The document was a recommendation—not a State Department finding that an AI apocalypse is inevitable, a federal law, or evidence that the United States currently imposes a universal compute cap.
What the report was
Defense in Depth: An Action Plan to Increase the Safety and Security of Advanced AI was prepared by Gladstone AI. Gladstone says the State Department commissioned the assessment in October 2022, before ChatGPT’s public release; the work was completed in February 2024 and announced publicly on March 11, 2024. The project reviewed historical nonproliferation regimes, surveyed AI research and development trends, and proposed a government-wide action plan. Gladstone says its team consulted more than 200 people, including government officials, cloud providers, security specialists, AI-safety researchers and frontier laboratories. See the report overview and disclaimer and the March 11 announcement.
Publicly circulated copies run roughly 250 pages, with one copy totaling 284 pages. The document was prepared for State Department review. Gladstone’s disclaimer says the contents were the authors’ responsibility and did not reflect the views of the State Department or the U.S. government.
Which AI risks did the authors discuss?
Weaponization
The report examined ways advanced systems could increase the capabilities of state or nonstate actors. Examples included cyberattacks, disinformation and influence operations, autonomous or robotic systems, and assistance with chemical, biological or materials-science work. These are risk scenarios, not claims that every current AI system can perform them or that a particular attack has occurred.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
Loss of control
The second category concerns highly capable systems behaving in ways developers cannot reliably predict, contain or correct, especially when they have broad autonomy or access to high-stakes environments. That is different from an ordinary chatbot error. It is a forward-looking risk framework: the report argued that sufficiently capable systems could create catastrophic or even “extinction-level” consequences if safeguards failed. It did not prove that human extinction is certain, imminent or unavoidable.
What “limiting compute” would mean
In this context, compute is the amount of computational work used to train a model, commonly expressed in operations or floating-point operations. The proposal targeted very large development runs rather than ordinary consumer use or all AI research.
- Report large runs: Require developers to disclose training above a specified level.
- Require approval: Make some runs subject to government permission rather than automatic authorization.
- Set a maximum: Establish a ceiling above which training would be prohibited unless rules changed.
- Control cloud access: Require cloud providers to restrict or refuse compute for certain large runs, especially open-access models.
- Monitor the infrastructure: Track advanced chips, data centers, electricity and other parts of the supply chain that make frontier training possible.
The report also discussed stricter treatment for models whose weights would be released openly, because a broadly available model may be easier for many actors to adapt or misuse.
The report’s example numbers
A publicly available copy proposed, as policy examples, a pause on frontier-model development above 1026 operations and a pause on open-access models above 1025 operations, along with cloud restrictions for open-access training above 1025 operations. Those figures were recommendations tied to the technical environment of 2024, not current legal limits or permanent scientific boundaries. The document said thresholds would need technical review as capabilities and efficiency changed. See the public report copy and TIME’s contemporaneous coverage.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesRank #2
Why regulators might use compute as a proxy
Compute is attractive to policymakers because the biggest training runs depend on a relatively small number of specialized chips, large data centers, cloud providers, power supplies and network connections. Those physical bottlenecks may be easier to audit than a model’s abstract “capability.” A threshold could also intervene before a system is deployed and before its full behavior is understood.
Compute is not the same thing as danger, however. Capability also depends on architecture, data quality, algorithmic efficiency, post-training, tool access, deployment design and inference-time computation. A smaller but more efficient model might outperform a larger older model. A developer could obtain new capabilities through fine-tuning or distillation rather than a new frontier-scale pretraining run. Background discussions include Computing Power and the Governance of Artificial Intelligence and the International Scientific Report on the Safety of Advanced AI.
The rest of the proposed action plan
Compute controls were only one part of the document. Its five broad lines of effort were:
Interim safeguards
- Monitor advanced-AI development.
- Create a coordinating task force.
- Introduce controls over critical parts of the advanced-AI supply chain.
More capable government
- Improve training and preparedness for officials.
- Build early-warning systems.
- Prepare contingencies for serious AI incidents.
Safety research and standards
- Fund alignment and other AI-safety research.
- Develop standards for responsible development and adoption.
Law and enforcement
- Create an advanced-AI regulatory agency with licensing and rulemaking authority.
- Establish civil and criminal liability.
- Provide emergency powers for rapidly developing threats.
International coordination
- Seek international agreement on catastrophic AI risks.
- Develop monitoring and verification mechanisms.
- Coordinate chip, cloud and other supply-chain controls.
The Gladstone overview describes these measures as recommendations from the authors, not as existing regulations.
Arguments for and against a compute threshold
| Potential benefit | Why it matters |
|---|---|
| Measurability | Operations, hardware and cloud use can be recorded more consistently than subjective capability claims. |
| Concentrated leverage | Large runs rely on specialized chips, data centers and a limited number of providers. |
| Early intervention | Officials could act before a model is widely deployed. |
| Race stabilization | Common limits might reduce incentives to scale simply to avoid falling behind competitors. |
| Limitation or risk | How it could undermine the policy |
|---|---|
| Obsolescence | Algorithmic improvements can make an old numerical threshold ineffective. |
| Circumvention | Training could be divided among countries, companies, providers or multiple smaller runs. |
| Capability mismatch | Data, architecture, tools, autonomy and inference-time scaling can matter as much as training operations. |
| Innovation costs | Startups, universities and nonprofit researchers might face burdens that large laboratories can absorb. |
| Enforcement complexity | Authorities would need reliable accounting for distributed training, foreign access, hardware and energy. |
| Strategic leakage | One-country limits could move development offshore without international coordination. |
| Open-model ambiguity | “Open source,” “open weights” and “open access” describe different release practices and would need precise legal definitions. |
| Metric gaming | Developers could optimize capability per operation or shift work into fine-tuning, distillation or inference. |
Important edge cases a training cap might miss
Efficient smaller models
A model trained below a threshold could still be highly capable if it uses better data or algorithms.
Fine-tuning and distillation
Adapting an existing model may create dangerous functionality without a new frontier-scale pretraining run.
Inference-time scaling
Systems can spend substantial computation after training to reason, search, simulate or operate tools. A training-only rule would not govern that later capability growth.
Distributed or international training
Splitting work across providers, legal entities or countries would complicate attribution and enforcement.
Rank #4
- NVIDIA Ampere Architecture-based CUDA Cores - Double-speed processing for single-precision floating point (FP32) operations and improved power efficiency provide significant performance improvements for graphics and simulation workflows, such as complex 3D computer-aided design (CAD) and computer-aided engineering (CAE), on the desktop.
- Second-Generation RT Cores - With up to 2X the throughput over the previous generation and the ability to concurrently run ray tracing with either shading or denoising capabilities, second-generation RT Cores deliver massive speedups for workloads like photorealistic rendering of movie content, architectural design evaluations, and virtual prototyping of product designs. This technology also speeds up the rendering of ray-traced motion blur for faster results with greater visual accuracy.
- Third-Generation Tensor Cores - New Tensor Float 32 (TF32) precision provides up to 5X the training throughput over the previous generation to accelerate AI and data science model training without requiring any code changes. Hardware support for structural sparsity doubles the throughput for inferencing. Tensor Cores also bring AI to graphics with capabilities like DLSS, AI denoising, and enhanced editing for select applications.
- Third-Generation NVIDIA NVLink - Increased GPU-to-GPU interconnect bandwidth provides a single scalable memory to accelerate graphics and compute workloads and tackle larger datasets.
- 48 Gigabytes (GB) of GPU Memory - Ultra-fast GDDR6 memory, scalable up to 96 GB with NVLink, gives data scientists, engineers, and creative professionals the large memory necessary to work with massive datasets and workloads like data science and simulation.
Academic and safety research
Any licensing system would need credible procedures for legitimate research without creating an exemption broad enough to defeat the rule.
Hardware controls
Chip and cloud restrictions may be easier to administer than model-by-model caps, but they can also affect scientific computing, commercial cloud services and allied countries.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Did the State Department adopt the recommendations?
Nothing in the report itself establishes a federal compute ban, a requirement that companies obtain State Department permission, or a current legal threshold of 1025 or 1026 operations. The institutional distinction is decisive: Gladstone produced an outside assessment commissioned for review, and its disclaimer says the authors’ views did not represent the department or the U.S. government.
Accordingly, accurate descriptions are “a report commissioned by the State Department proposed” or “Gladstone AI’s authors warned.” It is inaccurate to say that the State Department ordered companies to limit compute, declared that AI will end humanity, or made training above the example thresholds illegal.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Separate State Department materials discuss responsible AI development, risk mitigation, transparency, standards and international cooperation. They do not, by themselves, turn Gladstone’s proposed criminal prohibitions, licensing system or universal compute cap into department policy. See the department’s international cyberspace and digital policy strategy and AI materials.
How the proposal fits the wider AI-policy debate
The report appeared alongside debates over frontier-model evaluations, chip-export controls, voluntary commitments by major laboratories, open-weight releases, incident reporting and international governance. Compute-based rules are one possible intervention. Others focus on what a model can do, how it is deployed, how securely its weights are held, whether incidents are reported, and who is liable for harm.
That distinction matters for evaluating the proposal. A compute threshold may be administratively visible while missing a capable system developed efficiently. A capability evaluation may better target dangerous behavior but be harder to standardize and enforce. Chip and cloud controls can reach infrastructure, yet may create spillovers and require international cooperation. No single mechanism automatically resolves those trade-offs.
Bottom line
The report was real, serious and unusually broad. Gladstone AI warned that advanced AI could become a national-security and proliferation problem, not merely a software-quality problem, and proposed compute thresholds alongside licensing, liability, supply-chain controls, safety research and international verification. But the numerical limits were 2024 policy recommendations, and the document was not itself law, a current federal restriction or an official State Department conclusion that an AI apocalypse is inevitable.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
