Blue Planet Studio - stock.adobe
Mind the gap, not the doom: An operational guide to AI pacing
Business leaders should focus on near-term AI risks and capabilities gaps rather than existential debates to survive coming disruptions.
Calls to slow or pace the frontier of AI development have taken center stage recently. Several heads of industry have called publicly to coordinate slowdowns, with notable commentators such as Ezra Klein arguing that AI models should be prevented from handling their own training. Even Nvidia CEO Jensen Huang thinks a shutdown is needed if experimental AI models cannot be contained.
Technology and business leaders face dual temptations: to ignore the existential risk discourse and to wade into it, both of which would be mistakes. Business leaders in sectors affected by AI should focus on near-term, high-consensus risks that their firms will need to navigate regardless of whether existential risk materializes or dissipates. Only organizations that can weather the disruptions posed by current and near-future AI capabilities will be able to influence existential decision-making further down the line.
The debate
At the center of the most recent wave of debate is a concept known as recursive self-improvement (RSI). RSI can be understood as the result of increasingly automated workflows in AI research, ending in an intelligent system that can improve itself without human input. The prospect of an uncontrolled AI "takeoff" has supercharged the stakes of the slowdown debate across the board.
In early September, OpenAI announced it had successfully created and implemented a system referred to as an "automated research intern." OpenAI uses the intern label to denote that while the new system can carry out research tasks over multi-day timelines, it still requires human direction at multiple points within that time frame. According to OpenAI, adoption of the automated research intern has resulted in their research org running an agent-human workday ratio of 3.1 to 1.
To cement the trend, we can compare this ratio to the situation before the adoption of this new system. Total agent runtime in June 2026 was still below human labor (a ratio of less than 1:1). In a few short months, this relationship inverted. OpenAI now targets a fully automated AI researcher for March 2028.
Similar trends are occurring at Anthropic. In May 2026, 80% of the code at Anthropic was written by agents directed by human researchers. As of August 2026, AI R&D work at Anthropic is reportedly "led" by Claude in 26% of cases. While none of this work is yet at the fully autonomous level, Anthropic's own extrapolation suggests that by the end of 2026, Claude will hold the "lead" role in more than 80% of AI R&D efforts.
These trends show that RSI is already snowballing. But what makes this different from previously observed trends, and why should business leaders care? The key idea to hold is that RSI bucks the trend in who it affects most. If achieved, RSI has the potential to remove all humans from the loop, meaning the conventional tools that policymakers rely on -- licensing, liability and regulation -- will no longer have targets to act upon. OpenAI and Anthropic leadership acknowledge this and add that they do not know how to ensure that the product of RSI will be broadly considered good -- for businesses, for their employees, for anyone.
The case against a slowdown is also something to consider with sober eyes. Can regulators be trusted to understand and correctly apply safeguards? Given that RSI is so incremental, attempts to target techniques would almost certainly apply to practices in wide use at the frontier labs. Without banning these, what is the chance that regulators will be effective? And then there is the confounding factor of coordination with China. These concerns and more are driving debates across industry, academia and broader discourse.
Three scenarios for business leaders
There are a few possible outcomes relevant to business leaders. Each has its own implications for the stance organizations should adopt regarding how AI is implemented in their business, drives demand for their services or poses a threat to their secure operation.
In the first case, we see increasing acceleration. In this scenario -- which we are presently on track to see -- models improve on or ahead of the industry's projected schedule. For anyone other than frontier labs, the gulf in capabilities will widen, and the risk posed by newer models to orgs will only grow. Some businesses will see their sectors saturated with agent labor. Some businesses will fail because their services are based on a level of trust that models and their societal effects erode from under their feet. Still others will suffer security breaches orders of magnitude greater than their defenses are prepared to counteract.
Even orgs that do their best to adopt and adapt to the new models will face challenges. Some may find themselves locked into AI service levels far behind the frontier because their contract terms with AI providers did not account for the exponential rate of change. Others may become reliant on enterprise models that change their pricing in ways and at speeds they cannot keep up with. The hunger for cheap inference may drive some orgs to use open-weight models without sufficient knowledge of their training data, exposing them to backdoor attacks and unintended behavior in edge cases.
In the second case, progress at the frontier is slowed, either by policy or by voluntary coordination between stakeholders. This might add additional friction to training runs, stemming from disclosure and authorization requirements, strict requirements for safeguards or mechanistic interpretability, or a newly established standard of agent liability.
In this case, businesses need to be aware of the changing regulatory landscape and how it might affect them. It is unlikely that all policies will unfold evenly over time or around the globe. Compliance with a market environment like the EU may preclude a business from operating competitively in other markets, while noncompliance may present criminal consequences and large-scale fines. To give a sense of scale, Google was forced to pay a fine of 4.6 billion Euros in a 2026 anti-trust suit, a sum that accounts for nearly 2% of the EUs budget.
Businesses will still need to prepare to face cyber threats in this scenario. Whether progress slows due to policy or voluntary commitment, the potential for that progress to resume, fueled by backlogs of resources or the efficiency gains from conventional computing advancements during the imposed "pause" will no doubt loom in the near future, warded off by the efforts of policymakers, but not vanquished.
In the final case, progress may slow due to an underlying, as-yet-unknown property of AI itself. Some practitioners predict a plateau as the result of training data quality degradation and training systems on their own outputs (synthetic data).
With this outcome, businesses will still face the substantial task of implementing whatever frontier models are on the market, even if progress grinds to a halt. This represents a huge amount of value. Not only will businesses face competition for this work from their conventional class and cadre of competitors, but they will also need to ward off the huge accumulation of talent and compute now wielded by frontier labs hungry to revise their value propositions to include new industries.
For each of these outcomes, there is one metric that will be a binding constraint for businesses: the size of the gap between their own capabilities and the frontier. By implementing independent evaluations, firms can build trust in data about themselves and the frontier. In their Sept. 6 post, OpenAI tips their own hand, noting that their compute remained valuable even when new controls caused them to channel it toward alternative use cases. The challenge coming for business leaders is to justify themselves as more than an alternative use case for compute.
What to watch
If enforced, disclosure and transparency by frontier labs will be the bellwethers that business leaders watch to inform their decisions about AI. Organizations and their leadership should separate the existential from the operational questions. Whether long-term questions of doom are eventually resolved will not matter to organizations that do not survive the near-term disruption.
What to do next
- Mind the gap. Information about what happens at the frontier labs affects everyone, but information about the state of your business in relation to AI is much more privileged. Act today on what you can measure and change. Save for tomorrow what you cannot.
- Build for model substitution. If the frontier slows, you will run the same models for longer times and can confidently build around specific configurations. If the frontier continues to accelerate, your business will want to swap quickly. Build an architecture that allows for swapping.
- Avoid lock-in. Investing time and resources to implement systems that will be outpaced by the frontier can destroy enormous amounts of capital. Investments into proprietary data and other more durable infrastructure will likely age more gracefully.
- Assign an owner for the disclosures. Someone in the organization should read what your frontier vendors publish under frameworks such as SB 53 (CA) and RAISE (NY), the way someone reads SOC 2 reports today.
Weather the storm first -- the rest follows
The debate over slowing AI is worth following, but a thesis on existential risk is distinct from a business model for most firms. Whether the frontier accelerates, is slowed by external factors or plateaus is not something most businesses can influence. Much of what will affect bottom lines will occur regardless of what occurs at the frontier.
Business leaders outside the frontier labs should focus on adding value by ensuring their firms survive the near-term disruptions set to ripple through the economy over the next few business cycles. The businesses that survive and thrive will be the ones that prioritize measuring the gap between the frontier and themselves. Tools and techniques offered by independent evaluators can provide businesses with additional insights beyond what labs disclose as a result of policies or voluntary commitments, and tailor this information to firms' contexts. Equipped with tools to measure the gap to the frontier honestly and build systems that can bend instead of break, firms can set themselves up for success in the near term and secure a seat at the table for whatever comes after.
Chris Canal is a machine learning engineer and entrepreneur with a background in ML systems, natural language processing and product development. His experience designing large-scale ML infrastructure at Amazon and automating complex legal reasoning at RunTitle grounds EquiStamp's evaluation methodology and technical standards. Canal co-founded EquiStamp in 2023 to build independent evaluation infrastructure for frontier AI systems. He drives several AI wafety ecosystem consortia to collaboratively protect public interest. He also works in close partnership with the EU Commission's AI Office across various AI safety domains, including CBRN risk, loss of control and harmful manipulation of humans.
