Column|AI Doesn’t Move at One Speed, So Why Slow it Down as if it Does?

Starting this week, we are launching a new Column series where we discuss interesting topics in a more casual way than our typical research. For this debut edition, we would like to share our view on the most widely debated topic over the weekend: whether AI should slow down its frontier expansion, and whether it would.

Once again, X is awash in debate over slowing down AI. Anthropic is calling for slower development of frontier models and stronger safety oversight.

We had a similar discussion a few weeks ago. This time, though, Dario has laid out a serious, detailed argument in a lengthy essay, and both Sam and Elon have expressed support.

There’s plenty of skepticism. The most common argument is that training models has become too expensive, and frontier labs want an excuse to slow down. Others argue they want demand to catch up with model capabilities, then ramp training again once the market grows.

Release Pace Is Not Training Pace

An important distinction: slowing releases does not necessarily mean slowing training. Both Anthropic and OpenAI have developed advanced internal models that are not publicly available. And the safety risks AI poses vary substantially across industries.

In some fields, AI is rapidly dismantling organizational structures and ways of working that have been in place for decades.

For the first time, capital can rapidly scale compute and concentrate it on a single problem, without the years once required to train people and build organizations. Agents are designed to follow instructions.

Consider what just happened in mathematics. On September 8, OpenAI announced a solution to the Navier–Stokes existence and smoothness problem. The Hodge conjecture and the Riemann hypothesis could be next.

A similar way of organizing work is emerging in cybersecurity. In cases recently disclosed by Anthropic, attackers were already using multiple subagents for reconnaissance, code review, and verification of findings. Offensive and defensive operations will increasingly move too fast, and at too large a scale, for human engineers to keep track of.

This also gives frontier labs, with their vast compute and capital, something resembling a god’s-eye view of human society. In these fields, it matters to slow the pace or find ways to work with existing education and talent systems, even if every industry will eventually go through the transformation coding has experienced over the past two years.

Yet many industries still lack the data needed for reinforcement learning, particularly workflow traces and action logs. Coding, mathematics, and cybersecurity are fields where data is relatively accessible and feedback is often readily available. In other industries, even a standardized context layer is missing.

Take finance. You have the questions you ask GPT every day and the notes you keep in Notion. But much of an analyst’s work happens in the process of conducting research, deciding how and why to revise an EPS model, and reaching a conclusion. Analysts have long lacked, and still lack, the infrastructure to systematically capture these decision-making traces the way software engineers can record and revisit their work. We’ve spent a long time tackling this problem while building FUNDA’s own context layer.

One reason coding can support progress toward recursive self-improvement (RSI) is the abundance of accessible, high-quality data that can be reviewed and traced to its source. In most other fields, we first need to address the data bottleneck. This challenge extends beyond frontier models: robotics has been grappling with similar issues for years.

You can’t assume that every industry will compress five to ten years of progress into two years, as coding has.