there really isn't a great or easy path or ability to do that. And there certainly won't be in 5 or 10 years. Once these models all move on to the edge of the network, and they're all being turned around all the time, every day, and there's a great evolution underway.View on YouTube
As of November 30, 2025, it is too early to judge a prediction whose explicit horizon is 5–10 years after May 2023 (i.e., roughly 2028–2033). We are only about 2.5 years into that window, so the core claim about what will or will not be possible for regulators by that time cannot yet be tested.
What we can say so far:
-
On-device / edge models are clearly advancing, but not dominant for frontier capabilities. Apple’s “Apple Intelligence” strategy uses a relatively small ~3B-parameter model on-device plus larger models in a privacy-preserving cloud, illustrating a hybrid approach rather than a full shift of powerful models entirely to the edge. Apple emphasizes that many features run on-device, but more complex tasks are still offloaded to server-side models via Private Cloud Compute. (apple.com)
-
Regulatory systems are still in the process of being built, not demonstrably obsolete yet. The EU AI Act, approved in 2024 and entering phased effect starting August 2025, creates obligations specifically for general‑purpose / foundation models (called GPAI models), focusing on transparency, safety, and risk controls rather than trying to track every individual device instance. (ibanet.org) Enforcement and guidance are still being rolled out and even debated (e.g., calls to delay enforcement and soften rules), but there is not yet clear evidence that regulators in principle cannot monitor major model families or providers. (reuters.com)
-
The prediction’s key test (“it will be effectively impossible for regulators to comprehensively track, audit, and approve which models are being run on which machines”) is inherently about the end state of a technological and regulatory race. Current trends (rapid open-source proliferation and stronger on-device hardware, alongside regulatory experimentation) could plausibly support Friedberg’s concern, but by 2025 we do not yet have decisive evidence either way.
Because the specified 2028–2033 timeframe has not arrived and present evidence does not conclusively show that comprehensive regulatory oversight over models-on-edge is either definitively possible or definitively impossible, the prediction’s truth value cannot yet be determined.