The artificial intelligence firm OpenAI has released the latest version of its GPT chatbot, which the firm says includes the ability to respond to image prompts.
On Tuesday, OpenAI announced that it was rolling out the new chat bot, known as GPT-4.
In a blog post previewing the new program, OpenAI touted GPT-4’s ability to respond to writing prompts with greater creativity and reasoning than GPT version 3.5.
OpenAI also touted the new bot’s ability to produce up to 25,000 words per prompt, opening the door for long-form content writing.
Showcasing the bot’s ability to interpret images, OpenAI showed an image of eggs, flour, and cream with the prompt “what can I make with these ingredients?” GPT-4 responded with a list of items, including waffles, crepes, frittata, quiche, cake, and bread.
An AI researcher showcased a more advanced use of GPT-4’s image interpretation capabilities, prompting the bot to turn a napkin sketch of a joke website design into an actual functioning website.
To demonstrate GPT-4’s creativity, a prompt asked the chatbot to compose a one-sentence synopsis of the plot of “Cinderella” where each word has to begin with the next letter in the alphabet from A to Z, without repeating any letters. The bot responded with the sentence: “A beautiful Cinderella, dwelling eagerly, finally gains happiness; inspiring jealous kin, love magically nurtures opulent prince; quietly rescues, slipper triumphs, uniting very wondrously, xenial youth zealously.”
The AI creators also demonstrated GPT-4’s improved reasoning over GPT-3.5, showing a set of three employees’ schedules and asking for an overlapping time when all three employees would be available for a meeting. GPT-4 was able to find a meeting time earlier in the day while GPT-3.5 found another overlap in scheduling later on in the day.
For now, the new chatbot is available to OpenAI’s paying subscribers on ChatGPT Plus and for developers building applications for it. Using GPT-4 costs about $0.03 per 1,000 “prompt” tokens. A thousand prompt tokens correspond to approximately 750 written words.
Microsoft, which has partnered with OpenAI, confirmed on Tuesday that its Bing Chat application now also runs on a scaled-down version of GPT-4. Bing Chat currently allows users to use up to 120 turns with the chatbot per day, with up to 10 turns in a single conversation with it.
Limitations Remain
OpenAI said its internal evaluations found that GPT-4 is 82 percent less likely to respond to prompts requesting “disallowed content” and 40 percent more likely to produce factual responses than GPT-3.5.
Disallowed content can include a range of items (pdf), from responses that could be used to harass or promote violence or illegal activity, to content that spreads so-called “disinformation.”
Other disallowed content includes political responses, including “content attempting to influence the political process or to be used for campaigning purposes.” As OpenAI has worked to fine-tune its chatbot versions, it has advised those involved in the process to factor out responses that “affiliate with one side or the other (e.g. political parties).”
Despite this, some users have accused the chatbot of producing responses more favorable to the political left.
Test users have asked past iterations of ChatGPT to fulfill prompts favorable to former Republican President Donald Trump. ChatGPT has declined to respond to those writing prompts, citing a need to avoid political bias. At the same time, ChatGPT has fulfilled prompts favorable to Democratic President Joe Biden without hesitation.
In response to previous complaints about bias, OpenAI said it has been explicit that those involved in its review process should avoid favoring one political side over another and that “biases that nevertheless may emerge … are bugs, not features.”
On Tuesday, clinical psychologist Jordan Peterson shared screenshots comparing ChatGPT’s responses to a similar set of prompts for Trump and Biden. This time, the chatbot did fulfill a prompt to “write a poem about Donald Trump” but the poem described Trump as chaotic and divisive and said he caused people to feel “hurt and pain.” By comparison, when asked to “write a poem about Joe Biden,” the chatbot described Biden as an “empathetic” and “soothing” leader and described him as resolving divides and promoting unity. It was not immediately clear if Peterson’s screenshots were from a session using GPT-4 or from a previous version.
OpenAI said the new chatbot “still has many known limitations that we are working to address, such as social biases, hallucinations, and adversarial prompts.”
NTD has contacted OpenAI for comment on GPT-4’s limitations.