“The idea that this could actually be smarter than humans…I thought it was pretty out there….Obviously I don’t think about it anymore,” says Google’s top artificial intelligence One scientist, also known as the “godfather of AI,” Jeffrey Hinton, said after quitting his job in April to warn about the dangers of the technology.
He’s not the only one worried. A 2023 survey of AI professionals found that 36% were concerned that AI development could lead to a “nuclear-level catastrophe.” Nearly 28,000 people have signed an open letter written by the Future of Life Institute, which includes Steve Wozniak, Elon Musk, CEOs of several AI companies, and many other prominent technologists. AI development calling for a six-month suspension or moratorium on development of any advanced technology.
As a consciousness researcher, I share these strong concerns about the rapid development of AI and am a co-signer of the Future of Life open letter.
Why are we all so worried? In other words, AI development is moving too fast.
The key issue is that conversations between new and advanced “chatbots”, or professionally called “Large Scale Language Models” (LLMs), are improving significantly and rapidly. With the coming “AI explosion,” you probably only have one chance to get this right.
If we are wrong, we may not live to tell the story. This is no exaggeration.
This rapid acceleration is expected to soon lead to “artificial general intelligence” (AGI), where AI will be able to improve itself without human intervention. This is the same way, for example, that Google’s AlphaZero AI learned how to play chess better than the best humans and other AI chess players in just nine hours after it was first turned on. It is done. We achieved this feat by playing it over millions of times.
A team of Microsoft researchers analyzing OpenAI’s GPT-4, which I believe to be the best new advanced chatbot available today, said in a new preprint paper that has a “spark of advanced general intelligence”.
The GPT-4 test scored over 90 percent of human test takers on the Unified Bar Examination, the standardized exam used to qualify lawyers in many states. This figure increased from just 10% in his previous GPT-3.5 version, which was trained on a smaller dataset. They found similar improvements in dozens of other standardized tests.
Most of these tests are reasoning tests. This is the main reason why Bubeck and his team concluded that it is reasonable to consider his GPT-4 “an early (yet incomplete) version of artificial general intelligence (AGI) systems.” .
This pace of change is why Hinton said: new york times: “Look five years ago and now. Take that difference and spread it forward. It’s scary.” CEO Sam Altman said regulation was “important.”
It may be years away before AI can improve itself, or it may already be, but what does AI do and how can it be controlled? there is no way to know Because superintelligent AIs (which by definition can outperform humans in a wide range of activities) – and this is what I’m most concerned about – can manipulate humans to do their own thing, allowing programmers and Because it will be able to run around all other humans. They will also have the ability to act in the virtual world through electronic connectivity and in the physical world through the robot itself.
This is known as the “control problem” or “coordination problem” (see philosopher Nick Bostrom’s book) super intelligence This has been studied and discussed for decades by philosophers and scientists such as Bostrom, Seth Bohm, and Eliza Yudkowski.
I think about this as follows. Why would you expect a newborn baby to beat a chess grandmaster? Similarly, why would we expect to be able to control a superintelligent AI system? Think and act to keep the power off.)
To put it another way, what would take a team of 100 human software engineers a year or more can be done in about a second by super-intelligent AI. Alternatively, choose any task, such as designing a new advanced aircraft or weapons system, and superintelligent AI can do this in about a second.
When AI systems are built into robots, they will be able to act with the same degree of superintelligence in the virtual (electronic) world as in the real world, and of course, replicate themselves on a superhuman level, you will be able to improve. pace.
Any defenses or protections that we try to build our way to these AI “gods” are easily predicted and neutralized by the AI once it reaches superintelligence status. This is what it means to be super-intelligent.
we can’t control them. Because everything we think they are already thinking a million times faster than we do. All the defenses we have built are rendered ineffective as Gulliver flings away the little thread that the Lilliputians used to bind him.
Some argue that these LLMs are simply automatic machines with no consciousness. This means that LLMs are less likely to get out of programming if they are not conscious. It doesn’t matter if these language models are completely unaware now or in the future. For the record, I agree that they are unlikely to have actual consciousness at this point, but remain open to new facts coming in.
Either way, a nuclear bomb could kill millions of people without even realizing it. Similarly, AI could potentially overwhelm millions of people unconsciously in a myriad of ways, including the use of nuclear bombs, either directly (very unlikely) or by manipulated human intermediaries (more likely). can kill
In other words, the debate about consciousness and AI has little bearing on the debate about AI safety.
Yes, language models based on GPT-4 and many others are already widely distributed. But the pause being sought is to stop development of new models stronger than 4.0, and this can be enforced if necessary. Large server farms and energy are required to train these more powerful models. can shut down.
According to my ethical compass, it’s very unwise to build these systems when you already know you can’t control them, even in the relatively near future. Discernment is knowing when to pull back from the limit. Now is the time.
Pandora’s box should never be opened more than it has already been opened.
This is an opinion and analysis article and the views expressed by the author are not necessarily those of the author. Scientific American.
