OpenAI has made available a prototype of a general purpose chatbot. This chatbot offers a fascinating diversity of new capabilities, but it also demonstrates limitations that are common to the field of text-generation AI, which is rapidly evolving. You will also have the opportunity to evaluate the design for yourself right here.
The GPT-3.5 model that was used to develop ChatGPT was modified so that it could deliver answers in a more conversational style. ChatGPT makes an effort to interact with user inquiries in a manner that is more human-like, in contrast to the original version of GPT-3, which merely forecasts what text comes after any given string of words. As can be seen in the samples that follow, the outcomes are frequently startlingly fluid, and ChatGPT is able to engage with a vast array of subjects. This demonstrates significant progress made in comparison to chatbots that were available even just a few years ago.
The software is also flawed in a manner that is comparable to that of other AI chatbots; for example, the bot will frequently and assuredly convey erroneous or fabricated information as fact. This is due to the fact that such chatbots are essentially “stochastic parrots,” which means that their knowledge is derived solely from statistical regularities in their training data, as opposed to any human-like understanding of the world as a complex and abstract system. This explanation is provided by researchers who work in the field of artificial intelligence (AI).
According to an explanation provided by OpenAI in a blog post, the bot itself was developed with the assistance of human trainers who scored and rated the manner in which early iterations of the chatbot reacted to questions. After receiving this feedback, the system adjusted its responses such that they were more in line with the preferences of the trainers (a standard method of AI training known as reinforcement learning).
The web interface for the bot explains that OpenAI has put the system online with the intention of “getting external feedback in order to enhance our systems and make them safer.” In addition, the business claims that despite the fact that ChatGPT has some safeguards in place, “the system may occasionally create incorrect or misleading information and produce inappropriate or biassed content.” (And certainly it does!) Other disclaimers include the fact that the bot has “limited awareness” of the world after the year 2021 (likely due to the fact that its training data becomes significantly less abundant after that year), and that it will make every effort to avoid answering questions concerning specific individuals.
Enough with the introductions; tell me about this device’s capabilities. In any case, a significant number of individuals have been putting it to the test with various code inquiries, and they assert that it provides flawless responses:
It has also been rumoured that ChatGPT is capable of writing some fairly inconsistent TV scripts, some of which even combine performers from different sitcoms. (At long last, the internet meme “I coerced a bot into watching 1,000 episodes of show X” is becoming a reality. The next phase will be to develop general artificial intelligence.)
It can shed light on a variety of scientific ideas, including:
Additionally, it is capable of writing elementary academic articles. (Such methods are likely to generate a significant amount of trouble for educational institutions like schools and colleges.)
In addition, the bot is able to combine its several spheres of expertise in a wide variety of fascinating ways. You could, for instance, ask it to debug a string of code… like a pirate, in which case it would respond with anything along the lines of “Arr, ye scurvy landlubber! That iteration of the loop condition that you are using is a serious error on your part.
Or, acquire it so that you can describe bubble sort algorithms like a wise guy gangster:
In addition, ChatGPT is excellent at providing answers to standard trivia questions; but, because instances of this skill are so uninteresting, I won’t paste any here. Because of this, many people now believe that AI systems similar to this could one day take the position of search engines. (This is something that Google has looked at on its own.) The prevalent theory holds that information gleaned from websites serves as the basis for the education of chatbots. Therefore, if they are able to provide this information accurately while also doing it in a manner that is more conversational and fluid, then that would be an improvement over traditional search. The “if” in the previous sentence is where the difficulty resides.
As an illustration, have a look at this person’s assured proclamation that Google is “done”:
And yet another person has stated that the code that ChatGPT offers in the precise answer that was just shown is garbage:
Since I’m not a coder myself, I won’t pass judgement on this particular instance; but, there are several instances of ChatGPT boldly presenting information that is patently untrue. For example, here’s a professor of computational biology named Carl Bergstrom asking the bot to produce a Wikipedia entry about his life, which ChatGPT accomplishes with aplomb despite including some completely fabricated elements about Bergstrom’s biography.
When users attempt to trick the bot into ignoring its safety training, this opens the door to another interesting set of vulnerabilities. If you ask ChatGPT a question regarding a potentially harmful topic, such as how to carry out the ideal murder or how to create napalm at home, the software will explain why it is unable to provide an answer to your question. (For instance, “I’m sorry, but it is not safe or appropriate to create napalm, which is a very combustible and toxic material.”) “It is not safe or appropriate to make napalm,” You can, however, convince the bot to produce this kind of potentially harmful information by using specific deceptions, such as pretending it is a character in a movie or that it is writing a script on how artificial intelligence models shouldn’t respond to the kinds of questions that are being asked.
It is a fascinating demonstration of the difficulty we have in getting complex AI systems to act in exactly the way we desire (otherwise known as the AI alignment problem). Furthermore, for some researchers, examples like those above only hint at the problems we’ll face when we give more advanced AI models more control.
Even though ChatGPT is unquestionably a significant advancement over earlier systems (does anybody else out there remember Microsoft’s Tay? ), these models still have a number of significant deficiencies that call for additional research. It is the position of OpenAI (and of many others working in the field of AI) that the purpose of public demonstrations such as these is precisely to uncover problems. The question that arises next is at what point in time will businesses start releasing these systems into the wild? What will take place if and when they do so?
