Misalignment Museum

AI Regulation — just slap a 🙂 on it Existential pessimism has received $883M in philanthropy, sparking a rush of premature regulation and regulatory capture. There are now 407 AI regulation bills in the works in America. Let’s look at the bellwether bill from California: SB-1047 SB-1047 requires developers to ensure safety before training large AI models, with civil penalties for violations of the Act. This reflects a fundamental misunderstanding of how AI is built. It is not like engineering an artifact with known functional blocks. Metaphorically, it’s less like purposeful design and more like directed evolution, or parenting. The absurdity of this regulatory approach becomes stark if you substitute “teenager” for “AI” as a thought experiment. Testifying that a teenager will be “safe” is an empty promise providing a false sense of security at best. Doing so before you start parenting is rank absurdity. You can't prove safety in any cutting-edge complex system built by iterative algorithms — whether AI or teenager. Nor can you control them, in the engineering sense. (Summary of why in my 2006 Tech Review article: technologyreview.com/2006/07…) In SB-1047, developers must make “a positive safety determination that the model does not have the capability to enable certain harms, such as the creation of chemical or biological weapons that would result in mass casualties or a cyberattack on critical infrastructure resulting in $500 million or more in damage.” It sounds like a reasonable goal, but how would you parent a teenager to prevent this? How far would you have to cripple an AI to remove its chemical and biological reasoning? Given the vast array of industrial chemicals that could become a WMD if released, it becomes a fool’s errand. In my neighborhood, the arsine and phospine at our semi fabs could kill their local community if released. Did I just enable a harm by mentioning it? Even if you could purge most of chemistry and biology and the references to their powers, the mind control becomes crippling. The core reasoning powers of the AI models degrade with the more mind control you impose. Just like a teenager. Like typical regulatory burdens, testing needs to be according to “a written safety and security protocol” with annual compliance certification. Lots of work for lawyers and bureaucrats, and a hindrance to the builders. The burgeoning bureaucracy is the shoggoth on the face of America. As with cellular ag, the regulatory chokehold on innovation and progress is coming in the new industry's infancy. I would hope the government would look back on the early days of the Internet and all that blossomed in the unregulated domains. Innovation is what makes America great, again and again, but for the ignorant impulse to preemptively regulate that which we don’t understand. The regulators do not understand that which they regulate… and they never will. This may make them uneasy, but we are routinely building complex systems that exceed human understanding. We cannot regulate the product if we cannot understand the product. The locus of learning and progress is the process, not the product of creation.