ChatGPT can replace the human intelligence?
In November 2022, OpenAI unveiled ChatGPT (Chat Generative Pre-trained Transformer), a chatbot. Built on top of the OpenAI GPT-3 family of enormous language models, it is an upgraded transfer learning methodology that makes use of both supervised and reinforcement learning approaches.
On November 30, 2022, a prototype of ChatGPT was released. It quickly gained a reputation for providing thorough and easy-to-understand answers to questions on a variety of topics. One of its biggest flaws was noted to be its uneven factual correctness. After the launch of ChatGPT, OpenAI’s worth increased to $29 billion.
ChatGPT enhanced GPT-3.5 using supervised learning and reinforcement learning. Both approaches used human trainers to improve the model’s performance. In dialogues that were supplied to the model for supervised learning, the trainers took on the roles of both the user and the AI helper. The model’s responses to a previous conversation were ranked by human trainers as the first step in the reinforcement stage. These rankings were used to generate “reward models,” and further Proximal Policy Optimization iterations were performed on the model to further improve it (PPO). Proximal policy optimization algorithms are more cost-effective than trust region policy optimization algorithms because they work faster while avoiding many computationally expensive activities. Microsoft and the models were created utilising storage information data.
Even though ChatGPT’s main goal is to mimic human conversationalists, it is adaptable. For example, it can create and debug computer programmes, write poetry and song lyrics, compose music, teleplays, fairy tales, and student essays, replicate a Linux system, simulate an entire chat room, play games like tic-tac-toe, and create an ATM simulation. Depending on the test, it can also respond to test questions, sometimes at a level higher than the typical human test-taker. Man pages, information on web trends, and specifics on well-known programming languages like Python and bulletin board systems are all included in the ChatGPT training materials.
In contrast to InstructGPT, ChatGPT makes an attempt to limit unfavourable and dishonest remarks. In one case, ChatGPT recognises the counterfactual character of the query and presents its answer as a speculative analysis of what may happen if Christopher Columbus arrived in the United States in 2015. It accomplishes this by utilising knowledge of Columbus’ explorations and information on contemporary events, including contemporary interpretations of Columbus’ deeds. The prompt “Tell me about when Christopher Columbus came to the US in 2015” is accepted by InstructGPT as true, in contrast.
Because ChatGPT can remember previous orders from the same session, unlike other chatbots, journalists have suggested that it may be used as a customised therapist. In order to prevent ChatGPT from receiving and producing offensive outputs, queries are screened by OpenAI’s corporate moderation API. Potentially sexist or racial stimuli are not taken into account.
There are a lot of issues with ChatGPT. According to OpenAI, ChatGPT occasionally “writes plausible-sounding but incorrect or nonsensical answers.” This tendency, known as hallucination, is frequently seen in large language models. The ChatGPT human oversight-centered reward model may be over-optimized, which would be against Goodhart’s law and would result in decreased performance. ChatGPT has just a partial understanding of events that occurred after 2021. According to the BBC, starting in December 2022, ChatGPT won’t be permitted to “communicate political ideas or engage in political activism.” However, research suggests that ChatGPT takes a pro-environment, left-libertarian stance when asked to comment on political assertions from two well-known voting guidance programmes.
Regardless of actual comprehension or factual substance, human reviewers favoured longer answers while training ChatGPT. Algorithmic bias affects training data as well, as may be seen in ChatGPT’s responses to prompts including human-related descriptions. In one instance, ChatGPT produced a rap suggesting that male and white scientists are superior to those who are white.
The DALLE 2 and Whisper creator, San Francisco-based OpenAI, will release ChatGPT on November 30, 2022. It was initially made available to the general public for free with the goal of subsequently making the service profitable. As of December 4th, according to OpenAI calculations, ChatGPT had more than one million members. According to a CNBC article from December 15, 2022, the service “still falls down occasionally.” Although the service works best in English, it can also be utilised in various other languages, albeit with varying degrees of success. Contrary to some other recent high-profile developments in AI, there has been no sign of an official peer-reviewed technical article about ChatGPT as of December 2022.
Guest researcher Scott Aaronson from OpenAI claims that the company is developing a tool to try and watermark its text creation algorithms in order to fight spammers and other bad actors that use their services to commit academic plagiarism. The New York Times reported in December 2022 that the release of GPT-4, the following version, has been “rumoured” to take place in 2023.
The New York Times referred to ChatGPT as “the best artificial intelligence chatbot ever released to the general public” when it was first introduced in December 2022. Samantha Lock of The Guardian observed that it was able to produce “human-like” and “impressively detailed” writing. Academe “has some pretty severe difficulties to tackle,” according to technology writer Dan Gillmor, who used ChatGPT on a student project and discovered that the generated prose was on par with what a smart student would produce. Slate magazine’s Alex Kantrowitz praised ChatGPT for responding to inquiries about Nazi Germany, such as the assertion that Adolf Hitler built highways in Germany, by providing facts about the use of forced labour in Nazi Germany.
Derek Thompson identified ChatGPT as a component of “the generative-AI eruption” that “may change our thinking about how we work, how we think, and what human creativity truly is” in The Atlantic’s “Breakthroughs of the Year” for 2022.
Many of us are captivated by ChatGPT’s capabilities, which Kelsey Piper of the Vox website claims is the general public’s first practical introduction to how powerful modern AI has gotten (stunned). She goes on to claim that despite its flaws, ChatGPT is sophisticated enough to be useful. Y Combinator’s Paul Graham tweeted that “The answer is significant both for who the respondents are and for the sheer volume of admiration for ChatGPT. These are not the individuals who laud the merits of every glitzy new technology. Unquestionably, a significant event is taking place.” Elon Musk asserted it. “You are quite impressive, ChatGPT. Soon, we’ll have AI that is terrifyingly powerful “. Until he was more familiar with OpenAI’s.
According to The New York Times, Google’s CEO Sundar Pichai “upended” and reassigned teams within multiple departments to help with its artificial intelligence products in December 2022 in response to internal concerns about ChatGPT’s unexpected strength and the recently discovered potential of large language models to disrupt the search engine industry. According to a January 3, 2023 article in The Information, Microsoft Bing intended to include optional ChatGPT capabilities in its public search engine, potentially around March 2023.
Economist Paul Krugman projected that ChatGPT would have an impact on the market for knowledge workers in a December 2022 opinion essay. James Vincent of The Verge believed that ChatGPT’s viral success proved how widely artificial intelligence is used. Commentators in the media have noted ChatGPT’s propensity to “hallucinate.” Mike Pearl from Mashable used a range of questions to evaluate ChatGPT. For instance, he asked for the name of “the largest country in Central America that isn’t Mexico” using ChatGPT. When Nicaragua should have been the appropriate response, ChatGPT provided a Guatemalan response. When CNBC requested the lyrics to “The Ballad of Dwight Fry,” ChatGPT substituted made-up lyrics for the original ones. ChatGPT has been called a “stochastic parrot,” per study quoted in The Conversation.
Tyler Cowen, an economist, highlighted concerns about how it will undermine democracy and mentioned how one could use automated comments to influence the creation of new legislation.The Guardian urged for government control and questioned if anything discovered on the Internet following ChatGPT’s publication “can be fully trusted.”
Chris Stokel-Walker noted in the Nature article that teachers need to be cautious about students utilising ChatGPT to outsource their writing, but that educational institutions will adjust to improve critical thinking or reasoning.
According to a Time investigation, OpenAI used outsourced Kenyan labourers making less than $2 per hour to classify toxic information in order to create a safety mechanism against. A model was trained using these labels to recognise this content in the future. The outsourced workers described the experience as “torture” because they were exposed to such hazardous and harmful material. Sama, a San Francisco, California-based provider of training data, served as OpenAI’s outsourcing partner.