Reactive firefighting as a form of governance? No, I'm not panicking.
As recent security incidents with unreleased frontier models demonstrate, closed systems alone will not guarantee safety against sophisticated autonomous threats. Instead, policymakers must treat strategic transparency and open-weight collaboration as defensive principles. But will they? I hope so.
AI cannot be contained by Western regulation alone. If democracies shackle development, they do not stop AI—they hand the strategic advantage to China, Russia, North Korea and other less constrained actors. The technology, knowledge and incentives already exist globally. Governance may reduce specific harms, but it cannot reverse diffusion or competition. The realistic strategy is rapid learning, adaptation and defensive capability, not imagined control. If machine intelligence surpasses human intelligence within years, societies that failed to understand and use it will be governed by those that did.
Hi Luiza, thanks for this piece. Two huge things you correctly point to that I think deserve much more attention:
1. "More important than any specific AI governance measure... are the robustness, preparedness level, and integration between the various entities within a given AI ecosystem."
The obvious question is how can the public be assured of anything that happens WITHIN an AI ecosystem, especially when one business controls the full stack all the way to consumer. My answer is that the full stack is only whole because we let it be. The last hop, the client, is separable, and it is the one seat where the public could actually be represented.
2. "What I and many others are proposing is to strategically strengthen transparency and openness where we can, so that we make the research, development, security, oversight, and governance layers more refined, robust, integrated, and prepared."
Yes. And I think the open versus closed camps are irreconcilable at the model layer but jointly satisfiable at the delivery layer. Neither side gets what it wants by arguing about weights. Both get most of it if the path between a provider and a user is instrumented, verifiable, and governed by a standard rather than by whoever owns both ends of it.
The key word you used is BETWEEN the entities WITHIN a system. That is exactly where on the web we have historically reached for protocols. The rules and standards that govern how components of a system communicate with each other, not designed merely for speed of transmission, but for monitoring, gating, directing, and restricting the flow of information where necessary.
The system I propose to implement for all AI inference delivery is analogous to e-mail. Many of the components are proven mechanisms which all internet users already rely on today. The new and novel ones need broad support and adoption but they are incredibly feasible. I'm a practitioner, not an academic, but this issue deserves extraordinary effort from all of us.
I wrote up the case last night. The precedent I lean on is Carterfone, 1968. The FCC opened the interface layer of the phone network and let the carriers file technical standards to protect it, so the safety case is what precipitated decades of innovation nobody in that hearing room could have described. The argument is that you don't build a system that depends on the occupant of the middle seat being good. You put something good in that seat.
I see this governance gap play out personally with clients.
Most salaried professionals assume their job is safe if they keep upskilling. But the real risk is systemic — a model escapes, policy lags, and suddenly an entire sector's skills are obsolete.
A 12-month emergency fund and a dedicated learning budget aren't optional anymore. They're the only buffer against events no individual can control.
Reactive firefighting as a form of governance? No, I'm not panicking.
As recent security incidents with unreleased frontier models demonstrate, closed systems alone will not guarantee safety against sophisticated autonomous threats. Instead, policymakers must treat strategic transparency and open-weight collaboration as defensive principles. But will they? I hope so.
AI cannot be contained by Western regulation alone. If democracies shackle development, they do not stop AI—they hand the strategic advantage to China, Russia, North Korea and other less constrained actors. The technology, knowledge and incentives already exist globally. Governance may reduce specific harms, but it cannot reverse diffusion or competition. The realistic strategy is rapid learning, adaptation and defensive capability, not imagined control. If machine intelligence surpasses human intelligence within years, societies that failed to understand and use it will be governed by those that did.
Hi Luiza, thanks for this piece. Two huge things you correctly point to that I think deserve much more attention:
1. "More important than any specific AI governance measure... are the robustness, preparedness level, and integration between the various entities within a given AI ecosystem."
The obvious question is how can the public be assured of anything that happens WITHIN an AI ecosystem, especially when one business controls the full stack all the way to consumer. My answer is that the full stack is only whole because we let it be. The last hop, the client, is separable, and it is the one seat where the public could actually be represented.
2. "What I and many others are proposing is to strategically strengthen transparency and openness where we can, so that we make the research, development, security, oversight, and governance layers more refined, robust, integrated, and prepared."
Yes. And I think the open versus closed camps are irreconcilable at the model layer but jointly satisfiable at the delivery layer. Neither side gets what it wants by arguing about weights. Both get most of it if the path between a provider and a user is instrumented, verifiable, and governed by a standard rather than by whoever owns both ends of it.
The key word you used is BETWEEN the entities WITHIN a system. That is exactly where on the web we have historically reached for protocols. The rules and standards that govern how components of a system communicate with each other, not designed merely for speed of transmission, but for monitoring, gating, directing, and restricting the flow of information where necessary.
The system I propose to implement for all AI inference delivery is analogous to e-mail. Many of the components are proven mechanisms which all internet users already rely on today. The new and novel ones need broad support and adoption but they are incredibly feasible. I'm a practitioner, not an academic, but this issue deserves extraordinary effort from all of us.
I wrote up the case last night. The precedent I lean on is Carterfone, 1968. The FCC opened the interface layer of the phone network and let the carriers file technical standards to protect it, so the safety case is what precipitated decades of innovation nobody in that hearing room could have described. The argument is that you don't build a system that depends on the occupant of the middle seat being good. You put something good in that seat.
https://logosanalog.com/p/ai-is-two-things-which-one-is-on
The full protocol spec is in a DOI-registered paper linked at the end of it: https://doi.org/10.5281/zenodo.21610185
I see this governance gap play out personally with clients.
Most salaried professionals assume their job is safe if they keep upskilling. But the real risk is systemic — a model escapes, policy lags, and suddenly an entire sector's skills are obsolete.
A 12-month emergency fund and a dedicated learning budget aren't optional anymore. They're the only buffer against events no individual can control.