Frontier AI’s Road…
Painless AI Lab : A Hong Kong Practitioner’s Real-World AI Journey (Part 32)

The most useful AI argument this week is not a bigger story. It is whether frontier AI should ease its pace.

Anthropic CEO Dario Amodei published “We Must Pace the Frontier”. He is not telling anyone to stop. He argues that major labs should slow down so safety work can keep up. His plan includes third-party evaluators with employee-level access, common standards among leading labs, and coordination beyond any single company. Sam Altman said OpenAI also agrees the industry needs to “pace the frontier,” and that it will give independent evaluators similar access. Elon Musk wrote that Amodei was “right.” Demis Hassabis backed the same direction.

During the All-In Summit in Los Angeles, President Trump called Nvidia CEO Jensen Huang while Huang was onstage. Trump said robots and AI would not “destroy the world,” called that claim a “hoax,” and argued that opposing a rapid AI infrastructure build-out would only help competitors. He added that the United States still had to be “a little bit careful,” but should not stop. Huang said AI safety still matters. He also told Trump that America must not let competitors set the pace.

Timelines
Most lab leaders expect AI to handle most human knowledge work within one to three years. The harder question is what happens if the models become strong enough. “Recursive self-improvement” does not require a sudden jump from AGI to superintelligence. It only requires systems that can steadily and reliably speed up the building of the next generation. Whether that loop has already begun — and how steep it is.

Control
If an AI system did become comprehensively superhuman at strategy and research, would human oversight become too weak?

Before that point, the questions are:
What data are the models trained on?
What tools and privileges does the system get?
What evaluation, interpretability, and shutdown mechanisms are in place while developers still set the terms?
Are labs and governments coordinating, competing, or doing both at once?
In practice, a self-imposed pause is extremely hard. So is cross-lab coordination.

What actually decides this
Frontier AI has a large upside: medicine, materials science, and productivity. It also has an governance problem. Once systems start doing more of their own research, the test is no longer who makes the better speech. It is whether the evaluation and limits can be built in time and whether enough developers will accept the same floor.