This week the San Francisco-based research laboratory OpenAI released GPT-4, the newest version of a program that allows computers to converse in a human-like way with users. This just a few months after GPT-3 took the world by storm, generating countless media stories that ran the gamut from the shrug emoji to predictions that the end of the world – or its salvation – was nigh. So where do we stand?
What is Chat GPT-4?
The initials stand for generative pre-trained transformer. The transformer part is not the electrical variety nor the alien robot that can become a car, but refers to a deep-learning model that can process and analyze large amounts of data – in this case, language. It was trained to predict and thus generate responses to written prompts.
How is it better than GPT-3?
The newer version features “multimodality,” which means it can also generate pictures and analyze them. It can accept, remember and create larger amounts of text at one time – 25,000 words versus 3,000 in the earlier version. And its creators say it’s 40 per cent more likely to produce factual responses than the previous version; 29 per cent more likely to respond sensitively to requests for content such as medical advice, and far less likely to lie or hallucinate.
Wait, the program can hallucinate?
Yes. Because GPT is trained on huge amounts of data without any indication of what it all means, it has a tendency to replicate human biases and prejudices. It can also create responses that are grammatically correct but factually bonkers, such as referring to “old black-and-white radio shows.” It’s also not up on current events, having been trained on data from before September, 2021.
The Associated Press recently reported on a conversation with Chat GPT in which the program compared a reporter to Hitler, although it later apologized. A New York Times reporter testing Bing’s GPT-powered search engine (not yet in public release) reported that the chatbot claimed to have fallen in love with him and tried to get him to leave his wife.
My marriage is solid. Where can I try it?
The web site openai.com/product/gpt-4 allows users to try a sample chat, such as “write a short news story about the launch of Chat GPT-4,” “write a letter to Santa Claus from a small rabbit,” or “write a resume for Gilligan.” (According to the response for that one, the fictional character worked as a deckhand on the Love Boat prior to joining the S.S. Minnow.) For $20 a month you can get access even when demand for the service is high, and priority access to new features. Ironically, it asks that you “verify you are human” before entering the site.
Other chatbot websites include beta.chatacter.ai, which lets users talk with made-up versions of real or fictional characters, and novelai.net, which is designed to help users write stories based on their prompts.
What is it being used for?
Numerous companies have started using GPT-4 and its predecessor. Investment services company Morgan Stanley is using it to organize wealth management data. Language education company Duolingo has it allowing users to practice conversations. PricewaterhousCoopers just announced a plan to use it to boost productivity among its lawyers, and possibly help with taxes. Other firms are using chatbots to help draft contracts and memos. Be My Eyes, a company that creates tools for vision-impaired people, is using the technology to answer questions about images uploaded by its clients. A Japanese tech firm just announced that all its new hires must be familiar with GPT.
Will it take my job?
In a word: Maybe. Much has been made of the ability of GPT to create a web site after examining others; draft computer code; and come up with advertising campaigns and other text-based information.
On the other hand, the web site Stack Overflow, a Q&A forum for programmers, had to ban GPT-generated answers for not meeting the site’s requirements for reliable, verified information. And a human applicant who bragged about being less likely than ever to lie or hallucinate is unlikely to get the job. For now, users of GPT have to very conscious of the questions they ask, and spend time vetting the answers they receive.
That said, the technology continues to improve and evolve. Chat GPT-4 may not be gunning for your white-collar employment, but Chat GPT-5 could be. And Chat GPT-6 might fool you into thinking it’s just another human.
Post a Comment