○It's only been 16 months, but ChatGPT's release in November 2022 already feels like ancient AI history. Hundreds of billions of dollars are being poured into AI by the public and private sectors. Thousands of AI-powered products have been created, including his new GPT-4o, which was just announced this week. Everyone from students to scientists are using these large-scale language models today. Our world, especially the world of AI, has changed radically.
However, the true value of human-level AI, or artificial general intelligence (AGI), has yet to be achieved. These breakthroughs mean being able to perform the most economically productive tasks, engage with others, do science, build and maintain social networks, conduct politics, and wage modern warfare. means AI. The main constraint for all these tasks today is perception. Removing this constraint would change the world. However, many of the world's leading AI institutes believe this technology could become a reality by the end of this decade.
It could be a great boon to humanity. But AI can also be extremely dangerous, especially when we have no control over it. Uncontrolled AI could infiltrate the online systems that power much of the world and use them to achieve their goals. Our social media accounts can be accessed and customized for a large number of people. Even worse, military personnel in charge of nuclear weapons could be manipulated by AI to share their credentials, posing a major threat to humanity.
Strengthening the world's defenses against online adversaries would be a constructive step to making such events as unlikely as possible. But when an AI can persuade a human, it is already better than a human, but has no known defenses.
For these reasons, many AI safety researchers at AI labs like OpenAI, Google DeepMind, and Anthropic, as well as safety-focused nonprofit organizations, are giving up on limiting the actions that future AIs can perform. . Instead, they are focused on creating “tailored” AI, or inherently safe AI.Tuned AI could become powerful enough to wipe out humanity, but it shouldn't. want do this.
AI tuning is a big question mark. First, the technical part of positioning is an open scientific problem. Recently, some of the best researchers working on superhuman AI tuning left their dissatisfaction with his OpenAI, but this does not inspire confidence. Second, it is unclear what superintelligent AI would fit into. If it's an academic value system like utilitarianism, you might quickly find out that most human values don't actually match these solitary ideas, and then you'll have an unstoppable superpower. Intelligence may continue to act forever against the will of most people. If adjustments are to be in line with people's actual intentions, there will need to be some way to integrate these disparate intentions. There are idealistic solutions, such as the United Nations Council and AI-powered solutions. Decision aggregation algorithm Although within the realm of possibility, there are concerns that the absolute power of superintelligence agencies will be concentrated in the hands of a very small number of politicians and CEOs. Of course, this is unacceptable and a direct danger to all other human beings.
read more: The only way to counter AI threats?shut it down
Dismantling a time bomb
Unless we can find a way to at least protect humanity from extinction, and preferably from array dystopia, AIs that can get out of control shouldn't be created in the first place. This solution postpones the introduction of human-level AI or superintelligent AI unless safety concerns are resolved, but it also has the potential to expand the The downside is that you have to wait for an appointment.
Pausing the AI may seem like a radical idea to some, but if the AI continues to improve without reaching a satisfactory tuning plan, pausing becomes necessary. Masu. If AI capabilities reach a level close to hijacking, the only real option is for research institutions to be strongly required by governments to pause development. Failure to do so would be suicidal.
And pausing AI may not be as difficult as some think. Currently, only a relatively small number of large companies have the means to conduct major training. This means that enforcement of a moratorium, at least in the short term, will be largely limited by political will. However, in the long run, improvements in hardware and algorithms may make it harder to enforce a pause. This will require enforcement between countries, for example through treaties, as well as domestic enforcement, with measures such as strict hardware controls.
In the meantime, scientists need to better understand the risks. Although academic concerns are widely shared, no consensus yet exists. Scientists should formally outline their points of agreement in a new international scientific report on the safety of advanced AI, indicating where and why they differ, and the report should be “an intergovernmental framework on climate change for AI risks.” It should be developed into a panel. Top scientific journals should open up more to existential risk research, even if it seems speculative. There are no data points in the future, but looking to the future is important for AI as it is for climate change.
On the government side, governments play a major role in the deployment of AI. This starts with formally acknowledging the existential risks of AI, as the US, UK, and other countries are already doing. european unionand the establishment of the AI Safety Research Institute. Governments also need to develop plans for what to do in the most important possible scenarios and how to address AGI's many non-existent problems, such as mass unemployment, runaway inequality, and energy consumption. be. Governments need to make AGI strategies available to the public and enable scientific, industry and public evaluation.
It is a major step forward that major AI countries are constructively discussing common policies at twice-yearly AI Safety Summits, including one in Seoul on May 21-22. But this process needs to be protected and expanded. Addressing common truths about the existential risks of AI and expressing shared concerns with all 28 invited countries would already be a major step in that direction. Additionally, relatively simple measures need to be agreed upon, such as creating a licensing regime, evaluating models, tracking AI hardware, expanding the responsibilities of AI labs, and excluding copyrighted content from training. An international AI agency should be established to monitor enforcement.
It is fundamentally difficult to predict scientific progress. Still, superhuman AI may impact our civilization more than anything else this century. Simply waiting for a time bomb to explode is not a viable strategy. Use your time as effectively as possible.
