OpenAI’s most powerful AI model ever can tell jokes and write bar examinations, but can it also inflict harm?
GPT-4, the latest release from OpenAI, is the most powerful and stunning AI model ever from the firm that brought you ChatGPT and the Dall-E AI artist. The system can pass the bar exam, answer logic problems, and even offer you a recipe for leftovers based on a snapshot of your refrigerator – but its inventors warn that it may also propagate false information, incorporate toxic beliefs, and even manipulate people into performing duties on its behalf. Here’s all you need to know about our most recent AI ruler.
What is GPT-4?
GPT-4 is primarily a text-creation machine. Yet, it is a very excellent one, because being extremely good at writing language is basically equivalent to being very good at comprehending and reasoning about the universe.
So, if you give GPT-4 a question from a US bar exam, it will write an essay demonstrating legal knowledge; if you give it a medicinal molecule and ask for variations, it will appear to apply biochemical expertise; and if you ask it to tell you a fish joke, it will appear to have a sense of humour – or, at the very least, a good memory for bad cracker jokes (“what do you get when you cross a fish and an elephant? Swimming trunks!”).
Is it the same as ChatGPT?
Not exactly. If ChatGPT is the automobile, GPT-4 is the engine: a strong universal technology that can be tailored to a variety of diverse applications. You may have already encountered it, as it has been powering Microsoft’s Bing Talk – the one that went insane and threatened to harm people – for the previous five weeks.
Yet, GPT-4 may be used for more than only chatbots. Duolingo has built a version of it into its language learning app that can explain where learners went wrong rather than simply telling them the correct thing to say; Stripe is using the tool to monitor its chatroom for scammers; and assistive technology company Be My Eyes is building a tool that can describe the world for a blind person and answer follow-up questions about it using a new feature, image input.
What distinguishes GPT-4 from the previous version?
GPT-4 outperforms its elder brothers on a wide range of technological tasks. It can answer maths better, gets deceived into providing false answers less frequently, can perform quite well on standardised examinations – except for English literature, where it is comfortably in the bottom half of the league table – and so on.
It also has a stronger sense of ethics than the previous version: ChatGPT took its original engine, GPT-3.5, and placed filters on top to try to prevent it from answering hostile or damaging inquiries. Now that such filters are included into GPT-4, the system will respectfully reject to execute duties like ranking races by beauty, delivering sexist jokes, or offering sarin synthesis recommendations.
So GPT-4 isn’t dangerous?
OpenAI has clearly attempted to accomplish this. The business has published a lengthy report outlining examples of damages that GPT-3 might create that GPT-4 can mitigate. It even offered an early version of the system to third-party researchers at the Alignment Research Center, who sought to make GPT-4 to act like an evil AI from a movie.
It failed at the majority of those tasks: it couldn’t explain how it would duplicate itself, obtain new computer resources, or launch a phishing assault. But, the researchers were able to recreate it by using Taskrabbit to persuade a human worker to pass a “are you human” test, with the AI system even determining that it should lie to the worker and claim that it was a blind person who couldn’t see the photos. (It is unclear whether the experiment involved an actual Taskrabbit worker).
Yet, others are concerned that the better you teach an AI system the laws, the better it will learn to break them. The “Waluigi effect” appears to be the result of the fact that, while grasping the full nuances of what makes ethical conduct is difficult and complex, the answer to the question “should I be ethical?” is a much easier yes or no. Trick the system into thinking it is not ethical, and it will happily perform everything you ask of it.
