The Looming Regulatory Gap: How AI Hardware Advancements Could Outpace Governance
AI regulation meets hardware innovation
AI systems are growing increasingly powerful, and policymakers are turning to compute governance as a key strategy for regulating them. They often use a compute threshold — measured in floating-point operations (FLOPs) – to identify which models will be subject to their oversight. FLOPs represent the number of arithmetic operations performed by a computer and are used to measure the computational intensity of AI model training. This metric is important because it correlates with the complexity and potential capabilities of AI systems. By monitoring the computational resources used in AI training through FLOPs, policymakers aim to keep the development of frontier AI systems in check.
At the same time, AI hardware is advancing at a breakneck pace. We're seeing the development of increasingly efficient GPUs and the emergence of neuromorphic chips and quantum computing systems. These technologies could improve computational efficiency, potentially making it possible to train even larger and more complex models while maintaining or even reducing energy consumption and costs.
These advancements in AI hardware are going to necessitate significant changes to AI regulation – as AI hardware becomes more efficient, regulatory frameworks based solely on compute or cost metrics will quickly become obsolete or insufficient.
So, what does this mean for regulators? How can AI policy keep up with hardware breakthroughs?
In this post, we'll explore these questions by:
Summarizing compute thresholds in current AI policy
Scanning the AI hardware horizon and its implications for regulation
Discussing challenges facing policymakers and opportunities for effective governance
Note: We need both appropriately defined and enforceable regulation to effectively mitigate risk from advanced AI. This post focuses on defining regulation amid hardware advancements, but these developments also challenge enforcement. As hardware efficiency improves, traditional methods of detecting high-compute operations (e.g., monitoring energy use) may become less reliable. These dual challenges highlight the complexity of AI governance in a rapidly evolving technology landscape.
Compute and cost thresholds in AI oversight
Ensuring the safety and reliability of frontier AI systems is critical as they become more capable and integrated into society. Beyond immediate concerns of misuse and abuse, AI poses existential risks to humanity if developed without proper safeguards.
This is where compute governance comes in – offering a mechanism for monitoring and controlling the massive amount of computation fueling advanced AI development in an attempt to avoid uncontrollable or catastrophic outcomes. This hardware-centric approach is particularly compelling for policymakers as it provides a concrete, measurable way to oversee AI advancement.
President Biden's executive order on AI and California's proposed SB 1047 both highlight the need to keep a close eye on computationally intensive AI systems. The EU leverages compute governance in its AI Act as well. These policies use compute thresholds as a litmus test for AI oversight. The US Executive Order 14110 draws the line at 10^26 FLOPs (and at 10^23 FLOPs for models trained primarily on biological sequence data). AI models that exceed this threshold when being trained face stricter reporting requirements. The EU sets its bar at 10^25 FLOPs.
California's SB 1047 takes a slightly different approach. As of July 2024, it combines both compute and cost thresholds – the bill would apply to AI models that use more than 10^26 FLOPs and cost over $100 million to train.
These specific numbers represent a level of computation currently associated with the most advanced AI models – the kind that could pose significant risks if left unchecked. However, as AI hardware evolves, we will need to rethink these thresholds.
AI hardware innovations on the horizon
The landscape of AI hardware is rapidly evolving, with ongoing developments in GPUs by industry giants like Nvidia, specialized chips like Etched's transformer-only processor and even more disruptive technologies on the horizon. These cutting-edge innovations will push the boundaries of what's possible in AI and challenge our current approaches to compute governance.
One such innovation is the photonic chip, which uses light to perform calculations rather than electricity. Penn Engineering's silicon-photonic chip, for instance, performs vector matrix multiplications – critical operations in neural networks – faster and more energy-efficiently than traditional chips. While the compute for these chips can still be measured in FLOPs, their improved energy efficiency could significantly reduce the cost of training large models, allowing models to operate below current cost thresholds while maintaining or even increasing their capabilities.
Neuromorphic computing is inspired by the human brain and uses materials like vanadium oxides to create artificial neurons and synapses. This allows for rapid processing capabilities and efficient power consumption. Intel's recent announcement of Hala Point, the world's largest neuromorphic system, underscores this technology's potential to surpass traditional CPU and GPU setups in computational efficiency. It could be possible to measure the compute of these systems in FLOPs, but they might enable more efficient learning per FLOP, meaning current FLOP-based thresholds would be too high to effectively capture the capabilities of neuromorphic systems. Their energy efficiency could also significantly reduce training costs, making cost-based thresholds less effective.
Quantum computing stands out with its potential to process information exponentially faster than classical computers through quantum bits, or qubits. We’re seeing significant progress towards realizing this potential in recent breakthroughs, like the DOE's Argonne National Laboratory achieving a 0.1-millisecond coherence time for qubits. In the context of AI, quantum computing could dramatically accelerate certain types of machine learning algorithms. The partnership between Moderna and IBM explores quantum computing's applications in solving scientific challenges, further highlighting its transformative potential. Quantum computing presents a unique challenge for compute governance since current FLOP-based thresholds may not be applicable – this technology might require its own specific regulations and measurement standards.
If these new kinds of AI hardware become more integrated into model training, existing thresholds that trigger regulatory oversight will no longer accurately reflect the true capabilities and risks of AI systems.
Challenges and opportunities for policymakers
Although it seems clear that adjusting (and perhaps replacing) compute and cost regulatory thresholds will be necessary to keep up with AI advancements, doing so also presents a complex set of risks.
Regulators would face many challenges in adapting these thresholds to new technologies, including:
Keeping pace: The rapid advancements in AI hardware require continuous monitoring and updating of regulatory frameworks. By the time a new threshold is implemented, it may already be outdated.
Technical complexity: Understanding and evaluating the diverse range of emerging AI hardware solutions demands a high level of technical expertise.
Measurement and comparison: Different hardware architectures may require different metrics for performance and efficiency.
International coordination: Ensuring global cooperation and consistency in these standards is crucial. Disparate regulations could lead to regulatory arbitrage, where companies shop for the most lenient jurisdictions, potentially compromising safety and ethical standards.
Balancing innovation and safety: Regulators need to walk a fine line between enabling technological progress and ensuring adequate safeguards. Overly strict regulations could drive innovation underground or offshore, while overly lax ones could expose society to unnecessary risks.
Scaling regulatory capacity: Regulatory bodies will face a significant scaling challenge if more AI projects fall under their purview due to changing thresholds. The increased workload could strain resources and compromise the thoroughness and effectiveness of oversight.
To address these challenges, we need governance frameworks that can evolve alongside the technology they are regulating. Ultimately, the goal should be to create a regulatory environment that supports responsible innovation while providing robust protection against potential harms.
The following policy levers could help in achieving this goal:
Leveraging forecasting tools: Utilize tools to anticipate future AI hardware advancements and their impacts.
Example: Epoch AI conducts research on topics like compute trends, data usage, and algorithmic efficiency to forecast the effects of AI on the economy and society.
Implementing regular policy reviews: Establish frequent, agile review mechanisms to keep pace with rapidly changing technology (e.g. quarterly rapid assessments of AI hardware developments and their alignment with current regulations).
Example: The EU AI Act mandates annual assessments of high-risk AI systems and prohibited practices.
Note: While codifying periodic policy reviews in legislation is a step in the right direction, annual reviews are not frequent enough to keep pace with the evolution of AI technology.
Creating adaptive regulatory frameworks: Design flexible frameworks of compute governance based on broad principles instead of rigid rules.
Example: The Food and Drug Administration’s (FDA) Pre-Cert program for digital health technologies focuses on evaluating a company's culture of quality rather than just the product and envisions continuous oversight throughout a product’s life cycle.
Note: The US Supreme Court recently overturned the Chevron doctrine, meaning adaptive frameworks would now need to be more explicitly authorized in legislation since federal agencies can no longer rely on broad interpretations of their authority when implementing policy.
Establishing multi-stakeholder advisory panels: Form diverse panels to provide insights on the implications of evolving AI hardware.
Example: The National AI Advisory Committee (NAIAC) in the US advises the President and the National AI Initiative Office on various aspects of AI.
Encouraging international collaboration: Work with other countries to harmonize compute governance approaches.
Example: The collaboration between the US and UK AI Safety Institutes demonstrates cross-border efforts in AI governance.
Developing tiered regulatory approaches: Implement a system where scrutiny scales with AI system impact and capabilities.
Example: The EU AI Act's risk-based approach categorizes AI systems into unacceptable, high, and low/minimal risk.
Focusing on outcomes and impacts: Shift focus from input metrics like compute to output metrics that measure actual AI capabilities and impacts (e.g. via evaluations, red-teaming, third-party auditing, cybersecurity assessments, etc.).
Example: The Federal Trade Commission (FTC) has said they will focus on consumer protection outcomes when evaluating AI systems instead of the specific technologies used.
Note: An outcome-based approach is currently not feasible for frontier AI systems since we don’t know how to guarantee that they are safe and aligned before deployment. Measuring their impact based on outcomes could lead to catastrophic results if unforeseen harmful capabilities emerge.
Investing in regulatory technology (RegTech): Use AI to enhance regulatory processes.
Example: The Securities and Exchange Commission (SEC) uses a system called ARTEMIS (Advanced Relational Trading Enforcement Metric Investigation System) to analyze trading patterns and detect insider trading.
Promoting open standards: Encourage development of open standards for AI hardware and software.
Example: The Institute of Electrical and Electronics Engineers Standards Association’s (IEEE SA) P7000 series develops standards for AI ethics and governance.
Building regulatory capacity: Invest in training programs for regulators to ensure necessary technical knowledge of AI hardware.
Example: The National Institute of Standards and Technology (NIST) AI Risk Management Framework provides educational resources on AI risk management.
Note: Building regulatory capacity for risk management is crucial, but it's important to recognize that we still need significant progress in understanding how to assess and mitigate risks in advanced AI systems.
Many of these levers are already being implemented to varying degrees in AI governance across different jurisdictions and sectors. The challenge for policymakers is to learn from these existing efforts and create a more resilient and responsive regulatory environment that supports the safe development of AI.
Achieving this goal requires recognizing the diverse ecosystem of actors who influence AI policy – these individuals play crucial roles in shaping governance through their expertise, research, and advocacy. Understanding their perspectives is key to developing comprehensive and effective AI regulation. Key archetypes include:
Congressional Staffer: Crafts flexible yet robust laws, could focus on regular policy reviews and international collaboration.
Federal Agency Head: Implements and enforces regulations, could leverage RegTech and multi-stakeholder advisory panels.
AI Lab Policy Lead: Balances compliance with innovation, might advocate for open standards and outcome-focused regulations.
Given the complexity of AI governance, each role prioritizes different policy levers based on their unique position. In addition to implementing a broad scope of policy levers, effective governance requires understanding how these actors interact, where their interests align or conflict, and how to bridge gaps in their approaches so that we can work towards a more holistic framework for responsible AI development.
Lastly
It's important to note that while this post focused on AI hardware, similar challenges arise from innovations in data handling and AI algorithms. Techniques like synthetic data generation and the development of compressed, efficient algorithms could also enable more powerful AI models without necessarily increasing the amount of compute used, potentially impacting regulations in similar ways to hardware efficiency improvements. This shows just how essential effective and adaptable regulation is for mitigating risks and preventing harmful outcomes as AI technologies continue to become more complex and capable.