Safe While Competitive

As AI has progressed, the economy has become more confusing than ever. Along with the safety of AI itself, the economic evaluation of AI seems to just be speculation; speculation of the potential efficiency AI can cause, along with speculation of the efficiency of the AI models themselves.

Recently, many AI leaders and researchers have expressed concern about the safety of AI, and have questioned why the government hasn't put in any regulations. Because of this, people are doubtful of AI's safety and its potential to benefit our flourishing. As people lose trust in their government, they also lose trust in the motives of these AI labs. Why hasn't the government attempted to regulate AI? In the past, there have been races for innovation, especially for something like AI that has the potential to dominate efficiency, innovation, and war. As we live in a capitalist society, we are basically structured and built to compete in this AI race. Can we win a race to AI while staying safe? I think I've thought of a way to keep AI innovation moving while keeping that innovation structured and safe.

Right now, our economy is propped up on speculation of the efficiency AI can cause, along with speculation about how cheap we can get AI models to run. Right now, AI companies and the government are more focused on the possible power AI has, not the efficiency and expense it takes to run that AI. To make an AI more powerful, you can optimize it, which takes longer and doesn't hold a guaranteed return on the AI's power. Currently, we view this race as a race for AI power. We better AI power by building more and more data centers, giving AI more computational power, which basically guarantees the AI becomes more powerful and capable.

But people don't want data centers. People are starting to see the effects of global warming, along with rising taxes for people living near the data centers, and rising land prices, but onto the main point. If we continue down this route of building data centers, the economy becomes more and more dependent on AI eventually becoming more efficient or rather, more dependent on the speculation that AI will become more efficient and cheaper to run. If the cost of AI doesn't come down, that speculation is wrong, and the economy collapses. If AI goes rogue and turns out to be unsafe, the economy also collapses.

A safer way to approach this innovation is to put natural regulations on the efficiency of these AI models. Instead of bettering AI by just giving it more data centers which is quick and easy to do, we better it by making it more efficient with the limited computational power it already has. This guarantees a better investment, since we're building toward AI that costs less to run because it's actually more efficient. It also takes risk out of the picture in two ways: it causes natural limits on how fast AI models can progress, which gives us more time to build safety nets against rogue AI, and it keeps the economy from being propped up on speculation about future efficiency instead grounding it in efficiency we've actually built and guaranteed.

For the economy to avoid a crash, AI models eventually have to become more efficient, that much we've already said. But "eventually" is still just speculation. There's no guarantee it happens in time, or happens at all. So instead of relying on that speculation and hoping efficiency catches up later, why not require and build that efficiency now, before the economy gets any more dependent on AI? If we make efficiency a condition of progress from the start, instead of a bet we're hoping pays off down the road, we remove the risk of a crash instead of just hoping we dodge it.

The common argument against AI safety regulation is that being safe means going slow, and going slow means losing the race. But safe and slow aren't the same thing they're just related. What actually makes AI safe is testing, and testing takes time. Slowness isn't the goal; it's a side effect of doing the testing properly. That distinction matters, because if you ask a company to slow down on purpose for the sake of safety, you're asking it to sacrifice its competitive position for no direct payoff, and in a race-driven, capitalist system, that's a demand companies and governments will keep ignoring.

Efficiency regulation doesn't have that problem. Nobody has to argue for caution as the goal, because the slowdown isn't the point, it's a byproduct. Building real efficiency into a model takes harder research and more time than just adding more data centers. So the delay safety needs happens naturally, justified by something the market already wants: cheaper, more efficient models. This is what makes efficiency regulation stronger than a direct pace cap. It produces the same slowdown that safety testing needs, without asking anyone to fight against their own incentive to compete.

But efficiency regulation and safety regulation shouldn't just run side by side, with one happening to buy time for the other. They should be locked together directly. A lab shouldn't be allowed to scale past a certain point unless it has hit the required efficiency bar and completed the corresponding safety testing for that level of capability. The efficiency requirement becomes the literal gate for the safety testing window, not just extra time that exists and might get used well. That closes the gap in the argument. It's no longer an assumption that labs will spend their slower pace on safety work, it's a requirement that they do.