Image credits: STEFANI REYNOLDS/AFP / Getty Images
Today, at its first-ever developer conference, OpenAI unveiled GPT-4 Turbo, an improved version of its flagship text generation AI model, GPT-4which the company claims to be both “more powerful” and less expensive.
GPT-4 Turbo is available in two versions: one that strictly analyzes the text and a second version that includes the context of the text and images. The text analytics model is available in preview via an API starting today, and OpenAI says it plans to make both available to the general public “in the coming weeks.”
They are priced at $0.01 per 1,000 input tokens (~750 words), where the “tokens” represent pieces of plain text — for example, the word “fantastic” divided into “fan”, “heap” and “tick”) and $0.03 per 1,000 exit tokens. (To input tokens are tokens introduced into the model, while to go out tokens are tokens that the model generates based on the input tokens.) The price of GPT-4 Turbo image processing will depend on the size of the image. For example, transmitting a 1080 × 1080 pixel image to GPT-4 Turbo will cost $0.00765, OpenAI says.
“We have optimized performance so that we can offer GPT-4 Turbo at 3x lower price for input tokens and 2x lower price for output tokens compared to GPT-4,” writes OpenAI in a blog post shared with TechCrunch this morning.
GPT-4 Turbo has several improvements over GPT-4, one being a newer knowledge base to rely on when responding to requests.
Like all language models, GPT-4 Turbo is essentially a statistical tool for predicting words. Powered by a large number of examples, mostly from the web, GPT-4 Turbo learned the probability of words appearing based on patterns, including the semantic context of the surrounding text. For example, given a typical email ending with the fragment “Looking forward to…”, GPT-4 Turbo could complete it with “…to receive a response”.
GPT-4 was trained on web data through September 2021, but GPT-4 Turbo’s knowledge cutoff is April 2023. This should mean that questions about recent events – at least events that happened before the news deadline – will provide more precise answers.
GPT-4 Turbo also has an extended pop-up window.
Popup, measured in tokens, refers to the text that the model considers before generating any additional text. Models with small pop-ups tend to “forget” the content of even very recent conversations, causing them to wander off topic – often in problematic ways.
GPT-4 Turbo offers a popup of 128,000 tokens, four times the size of GPT-4 and the largest popup of any commercially available model, surpassing even that of Anthropic. Claude2. (Claude 2 supports up to 100,000 tokens; Anthropic claims to be experimenting with a 200,000 token pop-up but has not yet made it public.) 128,000 tokens translates to about 100,000 words or 300 pages, which , for reference, is approximately the Wuthering length. Height, Gulliver’s Travels and Harry Potter and the Prisoner of Azkaban.
And GPT-4 Turbo supports a new “JSON mode,” which ensures the model responds with valid information. JSON — the open standard file format and data exchange format. This is useful in web applications that transmit data, such as those that send data from a server to a client so that it can be displayed on a web page, OpenAI explains. Other new related settings will allow developers to make the model return “consistent” completions more often and, for more specialized applications, record the probabilities for the most likely output tokens generated by GPT-4 Turbo.
“GPT-4 Turbo performs better than our previous models on tasks that require careful following of instructions, such as generating specific formats (e.g. ‘always respond in XML’),” writes OpenAI. “And GPT-4 Turbo is more likely to return the correct function parameters.”
OpenAI did not neglect GPT-4 by deploying GPT-4 Turbo.
Today, the company is launching an experimental access program to refine GPT-4. Unlike the development program of GPT-3.5GPT-4’s predecessor, the GPT-4 program will involve more oversight and guidance from OpenAI teams, according to the company, primarily due to technical hurdles.
“Preliminary results indicate that GPT-4 fine-tuning requires more work to achieve significant improvements over the base model compared to the substantial gains made with GPT-3.5 fine-tuning,” writes OpenAI in the post. blog.
Elsewhere, OpenAI announced that it would double the token throughput limit per minute for all paying GPT-4 customers. But prices will remain the same at $0.03 per input token and $0.06 per output token (for GPT-4 model with 8,000 token popup) or $0.06 per output token. input and $0.012 per output token (for GPT-4 with 32,000 token popup). token pop-up window).