OpenAI shipped GPT-4 today, the much-anticipated text-generating AI model, and it’s a curious piece of work.
GPT-4 improves upon its predecessor, GPT-3, in key ways, for example giving more factually true statements and allowing developers to prescribe its style and behavior more easily. It’s also multimodal in the sense that it can understand images, allowing it to caption and even explain in detail the contents of a photo.
But GPT-4 has serious shortcomings. Like GPT-3, the model “hallucinates” facts and makes basic reasoning errors. In one example on OpenAI’s own blog, GPT-4 describes Elvis Presley as the “son of an actor.” (Neither of his parents were actors.)
To get a better handle on GPT-4’s development cycle and its capabilities, as well as its limitations, TechCrunch spoke with Greg Brockman, one of the co-founders of OpenAI and its president, via a video call on Tuesday.
Asked to compare GPT-4 to GPT-3, Brockman had one word: Different.
“It’s just different,” he told TechCrunch. “There’s still a lot of problems and mistakes that [the model] makes … but you can really see the jump in skill in things like calculus or law, where it went from being really bad at certain domains to actually quite good relative to humans.”
Test results support his case. On the AP Calculus BC exam, GPT-4 scores a 4 out of 5 while GPT-3 scores a 1. (GPT-3.5, the intermediate model between GPT-3 and GPT-4, also scores a 4.) And in a simulated bar exam, GPT-4 passes with a score around the top 10% of test takers; GPT-3.5’s score hovered around the bottom 10%.
Shifting gears, one of GPT-4’s more intriguing aspects is the above-mentioned multimodality. Unlike GPT-3 and GPT-3.5, which could only accept text prompts (e.g. “Write an essay about giraffes”), GPT-4 can take a prompt of both images and text to perform some action (e.g. an image of giraffes in the Serengeti with the prompt “How many giraffes are shown here?”).
That’s because GPT-4 was trained on image and text data while its predecessors were only trained on text. OpenAI says that the training data came from “a variety of licensed, created, and publicly available data sources, which may include publicly available personal information,” but Brockman demurred when I asked for specifics. (Training data has gotten OpenAI into legal trouble before.)
GPT-4’s image understanding abilities are quite impressive. For example, fed the prompt “What’s funny about this image? Describe it panel by panel” plus a three-paneled image showing a fake VGA cable being plugged into an iPhone, GPT-4 gives a breakdown of each image panel and correctly explains the joke (“The humor in this image comes from the absurdity of plugging a large, outdated VGA connector into a small, modern smartphone charging port”).
Only a single launch partner has access to GPT-4’s image analysis capabilities at the moment — an assistive app for the visually impaired called Be My Eyes. Brockman says that the wider rollout, whenever it happens, will be “slow and intentional” as OpenAI evaluates the risks and benefits.
“There’s policy issues like facial recognition and how to treat images of people that we need to address and work through,” Brockman said. “We need to figure out, like, where the sort of danger zones are — where the red lines are — and then clarify that over time.”
OpenAI dealt with similar ethical dilemmas around DALL-E 2, its text-to-image system. After initially disabling the capability, OpenAI allowed customers to upload people’s faces to edit them using the AI-powered image-generating system. At the time, OpenAI claimed that upgrades to its safety system made the face-editing feature possible by “minimizing the potential of harm” from deepfakes as well as attempts to create sexual, political and violent content.
Another perennial is preventing GPT-4 from being used in unintended ways that might inflict harm — psychological, monetary or otherwise. Hours after the model’s release, Israeli cybersecurity startup Adversa AI published a blog post demonstrating methods to bypass OpenAI’s content filters and get GPT-4 to generate phishing emails, offensive descriptions of gay people and other highly objectionable text.
It’s not a new phenomenon in the language model domain. Meta’s BlenderBot and OpenAI’s ChatGPT, too, have been prompted to say wildly offensive things, and even reveal sensitive details about their inner workings. But many had hoped, this reporter included, that GPT-4 might deliver significant improvements on the moderation front.
When asked about GPT-4’s robustness, Brockman stressed that the model has gone through six months of safety training and that, in internal tests, it was 82% less likely to respond to requests for content disallowed by OpenAI’s usage policy and 40% more likely to produce “factual” responses than GPT-3.5.
“We spent a lot of time trying to understand what GPT-4 is capable of,” Brockman said. “Getting it out in the world is how we learn. We’re constantly making updates, include a bunch of improvements, so that the model is much more scalable to whatever personality or sort of mode you want it to be in.”
The early real-world results aren’t that promising, frankly. Beyond the Adversa AI tests, Bing Chat, Microsoft’s chatbot powered by GPT-4, has been shown to be highly susceptible to jailbreaking. Using carefully tailored inputs, users have been able to get the bot to profess love, threaten harm, defend the Holocaust and invent conspiracy theories.
Brockman didn’t deny that GPT-4 falls short, here. But he emphasized the model’s new mitigatory steerability tools, including an API-level capability called “system” messages. System messages are essentially instructions that set the tone — and establish boundaries — for GPT-4’s 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.”
The idea is that the system messages act as guardrails to prevent GPT-4 from veering off course.
“Really figuring out GPT-4’s tone, the style and the substance has been a great focus for us,” Brockman said. “I think we’re starting to understand a little bit more of how to do the engineering, about how to have a repeatable process that kind of gets you to predictable results that are going to be really useful to people.”
Brockman and I’s conversation also touched on GPT-4’s context window, which refers to the text the model can consider before generating additional text. OpenAI is testing a version of GPT-4 that can “remember” roughly 50 pages of content, or five times as much as the vanilla GPT-4 can hold in its “memory” and eight times as much as GPT-3.
Brockman believes that the expanded context window lead to new, previously unexplored applications, particularly in the enterprise. He envisions an AI chatbot built for a company that leverages context and knowledge from different sources, including employees across departments, to answer questions in a very informed but conversational way.
That’s not a new concept. But Brockman makes the case that GPT-4’s answers will be far more useful than those from chatbots and search engines today.
“Previously, the model didn’t have any knowledge of who you are, what you’re interested in and so on,” Brockman said. “Having that kind of history [with the larger context window] is definitely going to make it more able … It’ll turbocharge what people can do.”