OpenAI has released a powerful new image- and text-understanding AI model, GPT-4, that the company calls “the latest milestone in its effort in scaling up deep learning.”
Pricing is $0.03 per 1,000 “prompt” tokens (about 750 words) and $0.06 per 1,000 “completion” tokens (again, about 750 words). Tokens represent raw text; for example, the word “fantastic” would be split into the tokens “fan,” “tas” and “tic.” Prompt tokens are the parts of words fed into GPT-4 while completion tokens are the content generated by GPT-4.
Other early adopters include Stripe, which is using GPT-4 to scan business websites and deliver a summary to customer support staff. Duolingo built GPT-4 into a new language learning subscription tier. Morgan Stanley is creating a GPT-4-powered system that’ll retrieve info from company documents and serve it up to financial analysts. And Khan Academy is leveraging GPT-4 to build some sort of automated tutor.
GPT-4 can generate text and accept image and text inputs — an improvement over GPT-3.5, its predecessor, which only accepted text — and performs at “human level” on various professional and academic benchmarks. For example, GPT-4 passes a simulated bar exam with a score around the top 10% of test takers; in contrast, GPT-3.5’s score was around the bottom 10%.
OpenAI spent six months “iteratively aligning” GPT-4 using lessons from an internal adversarial testing program as well as ChatGPT, resulting in “best-ever results” on factuality, steerability and refusing to go outside of guardrails, according to the company. Like previous GPT models, GPT-4 was trained using publicly available data, including from public webpages, as well as data that OpenAI licensed.
OpenAI worked with Microsoft to develop a “supercomputer” from the ground up in the Azure cloud, which was used to train GPT-4.
“In a casual conversation, the distinction between GPT-3.5 and GPT-4 can be subtle,” OpenAI wrote in a blog post announcing GPT-4. “The difference comes out when the complexity of the task reaches a sufficient threshold — GPT-4 is more reliable, creative and able to handle much more nuanced instructions than GPT-3.5.”
Without a doubt, one of GPT-4’s more interesting aspects is its ability to understand images as well as text. GPT-4 can caption — and even interpret — relatively complex images, for example identifying a Lightning Cable adapter from a picture of a plugged-in iPhone.
The image understanding capability isn’t available to all OpenAI customers just yet — OpenAI’s testing it with a single partner, Be My Eyes, to start with. Powered by GPT-4, Be My Eyes’ new Virtual Volunteer feature can answer questions about images sent to it. The company explains how it works in a blog post:
“For example, if a user sends a picture of the inside of their refrigerator, the Virtual Volunteer will not only be able to correctly identify what’s in it, but also extrapolate and analyze what can be prepared with those ingredients. The tool can also then offer a number of recipes for those ingredients and send a step-by-step guide on how to make them.”
A more meaningful improvement in GPT-4, potentially, is the aforementioned steerability tooling. With GPT-4, OpenAI is introducing a new API capability, “system” messages, that allow developers to prescribe style and task by describing specific directions. System messages, which will also come to ChatGPT in the future, are essentially instructions that set the tone — and establish boundaries — for the AI’s next interactions.
For example, a system message might read: “You are a tutor that always responds in the Socratic style. You never give the student the answer, but always try to ask just the right question to help them learn to think for themselves. You should always tune your question to the interest and knowledge of the student, breaking down the problem into simpler parts until it’s at just the right level for them.”
Even with system messages and the other upgrades, however, OpenAI acknowledges that GPT-4 is far from perfect. It still “hallucinates” facts and makes reasoning errors, sometimes with great confidence. In one example cited by OpenAI, GPT-4 described Elvis Presley as the “son of an actor” — an obvious misstep.
“GPT-4 generally lacks knowledge of events that have occurred after the vast majority of its data cuts off (September 2021), and does not learn from its experience,” OpenAI wrote. “It can sometimes make simple reasoning errors which do not seem to comport with competence across so many domains, or be overly gullible in accepting obvious false statements from a user. And sometimes it can fail at hard problems the same way humans do, such as introducing security vulnerabilities into code it produces.”
OpenAI does note, though, that it made improvements in particular areas; GPT-4 is less likely to refuse requests on how to synthesize dangerous chemicals, for one. The company says that GPT-4 is 82% less likely overall to respond to requests for “disallowed” content compared to GPT-3.5 and responds to sensitive requests — e.g. medical advice and anything pertaining to self-harm — in accordance with OpenAI’s policies 29% more often.
There’s clearly a lot to unpack with GPT-4. But OpenAI, for its part, is forging full steam ahead — evidently confident in the enhancements it’s made.
“We look forward to GPT-4 becoming a valuable tool in improving people’s lives by powering many applications,” OpenAI wrote. “There’s still a lot of work to do, and we look forward to improving this model through the collective efforts of the community building on top of, exploring, and contributing to the model.”