In a spin-off from ‘The Lord of the Rings’, two Hobbits are concerned about the safety of their forest. They request it from a Tree-Beard. Tree-Beard is an intelligent but extremely slow-moving tree. The problem was that while an entire army was cutting down and destroying the forest, Tree-Beard’s sense of time was very slow. It would take him a whole day just to say ‘Hello’ to another tree, so it was almost impossible to prepare him and his companions for such quick action. The relationship between AI and our political organizations also seems a bit like Hobbits and Tree-Beard. AI is advancing at breakneck speed. Just four years ago, AI models could barely write a single correct line of code, but today they are writing the majority of code at major AI companies. Similar progress has been seen in biology, physics, mathematics, finance, law, translation and many other fields. The scaling laws of AI are now supported by over a decade of empirical evidence. If these scaling rules continue for just a year or two more, we will likely reach what I have called ‘Powerful AI’. On the contrary, the pace of making policies and laws is extremely slow. Sometimes there are valid reasons for this. Governments have serious and wide-ranging powers, and it is generally considered better not to use them hastily. But this imbalance of timescales is troublesome. It takes several years for Parliament to take a decision, and in that time, AI reaches the capabilities of an entire country full of talent! Over the past few years—as AI has emerged as a major commercial technology—many of us who want to operate it responsibly have been faced with a dilemma. We were able to clearly see where this technology was headed. We had a feeling that within a few years, AI would become one of those rare technologies that fundamentally changes the entire policy landscape—just as nuclear weapons changed geopolitics and the Industrial Revolution fundamentally transformed the nature of economic questions. But to those who were looking at the capabilities of AI at the time, it seemed like a much more ordinary technology – perhaps like a new consumer app or cryptocurrency. It was difficult to convince most policymakers and companies that any policy other than a free-market approach could be appropriate. To be honest, the broader impacts of AI were not yet known, and we did not even know what form they would take. Therefore it was difficult to make right policies. But given these limitations, safety-focused organizations (including Anthropic) have focused on policy initiatives that safeguard future options, develop the ability to respond quickly, or give the world a better understanding of what’s to come—such as transparency laws, chip export controls, and the collection of data on AI’s labor-market impacts. Admittedly, these measures are inadequate, but so far these seem to be the only steps that could actually be made possible. However, over the past few months, the evidence of AI’s extraordinary power and its associated risks has become so clear that it is no longer possible to ignore them. Perhaps the most emblematic example of this is the Claude Mythos preview. We now know that leading AI models pose real and serious risks to cybersecurity, with the potential for disruption to the financial sector, critical infrastructure, and national security. In fact, Mythos has shaken up the global cybersecurity landscape. But it also proved that AI models have now become tools of strategic importance. Now we have to activate such a policy mechanism, which can face the risks of the times to come. It is encouraging that many of our contemporaries are now coming to the viewpoints that we have been advocating for the last few years. We need to focus on some important points, such as public safety and AI regulations, related macro-economics and tax policies, and enhancing the positive impacts of AI. Time is short, we have to be quick. (Courtesy of darioamodei.com)
Source link
Dario Amodei’s column: The pace of AI is so fast that we are lagging behind it
