New Microsoft paper claims GPT-4 shows “Sparks of AGI”

Microsoft is betting heavily on integrating OpenAI's GPT-4 language models into its products to compete with Google, and the company now claims that GPT-4, which researchers eventually found out uses something called a Master of Experts (MOE) model, rather than being a single Artificial Intelligence (AI) model, to “be so good,” is an early form of Artificial General Intelligence (AGI).

On Wednesday, Microsoft researchers released a paper on the arXiv preprint server titled “Sparks of Artificial General Intelligence: Early experiments with GPT-4.” They declared that GPT-4 showed early signs of AGI, meaning that it has capabilities that are at or above human level.

This eyebrow-raising conclusion largely contrasts what OpenAI CEO Sam Altman has been saying regarding GPT-4. For example, he said the model was "still flawed, still limited." In fact, if you read the paper itself, the researchers appear to dial back their own splashy claim: the bulk of the paper is dedicated to listing the number of limitations and biases the Large Language Model (LLM) contains. This begs the question of how close to AGI GPT-4 really is, and how AGI is instead being used as clickbait.

“We demonstrate that, beyond its mastery of language, GPT-4 can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting,” the researchers write in the paper’s abstract. “Moreover, in all of these tasks, GPT-4’s performance is strikingly close to human-level performance, and often vastly surpasses prior models such as ChatGPT. Given the breadth and depth of GPT-4’s capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an Artificial General Intelligence (AGI) system.”

Indeed, the researchers show examples of GPT-4’s capabilities in the paper: it is able to write a proof about how there are infinitely many primes, with rhymes on every line, and draw a unicorn in TiKZ, a drawing program. This is all quickly followed by some serious caveats.

While in the abstract of the paper the researchers write that “GPT-4’s performance is strikingly close to human-level performance,” their introduction immediately contradicts that initial attention-grabbing statement. They write, “Our claim that GPT-4 represents progress towards AGI does not mean that it is perfect at what it does, or that it comes close to being able to do anything that a human can do (which is one of the usual definition [sic] of AGI; see the conclusion section for more on this), or that it has inner motivation and goals (another key aspect in some definitions of AGI).”

The researchers said that they used a 1994 definition of intelligence by a group of psychologists as the framework for their research. They wrote, “The consensus group defined intelligence as a very general mental capability that, among other things, involves the ability to reason, plan, solve problems, think abstractly, comprehend complex ideas, learn quickly and learn from experience. This definition implies that intelligence is not limited to a specific domain or task, but rather encompasses a broad range of cognitive skills and abilities.”

“OpenAI’s powerful GPT-4 model challenges many widely held assumptions about the nature of machine intelligence. Through critical evaluation of the system’s capabilities and limitations, which you can read about in ‘Sparks of Artificial General Intelligence: Early experiments with GPT-4,’ Microsoft researchers observed fundamental leaps in GPT-4’s abilities to reason, plan, solve problems, and synthesize complex ideas that signal a paradigm shift in the field of computer science,” a Microsoft spokesperson said. “We recognize the current limitations of GPT-4 and that there is still work to be done. We will continue to engage the broader scientific community in exploring future research directions, including those required to address the societal and ethical implications of these increasingly intelligent systems.”

OpenAI CEO Sam Altman emphasized the limitations of GPT-4 when it was released, saying “it is still flawed, still limited, and it still seems more impressive on first use than it does when you spend more time with it.” In a Thursday interview with Intelligencer’s Kara Swisher, Altman shared the same disclaimers: “There’s plenty of things it’s still bad at.”

In the interview, Altman agrees that the bot will sometimes make things up and present users with misinformation. He said that there still needs a lot more human feedback to be more reliable.

Altman and OpenAI have always looked toward a future where AGI exists, and have recently been engaged in building hype around the firm's ability to bring it about. But Altman has also been clear that GPT-4 is not AGI.

“The GPT-4 rumor mill is a ridiculous thing. I don’t know where it all comes from,” Altman said just before GPT-4's release. “People are begging to be disappointed and they will be. The hype is just like... We don’t have an actual AGI and that’s sort of what’s expected of us.”

"Microsoft is not focused on trying to achieve AGI. Our development of AI is centered on amplifying, augmenting, and assisting human productivity and capability. We are creating platforms and tools that, rather than acting as a substitute for human effort, can help humans with cognitive work,” a Microsoft spokesperson clarified in a statement to Motherboard.

The Microsoft researchers write that the model has trouble with confidence calibration, long-term memory, personalization, planning and conceptual leaps, transparency, interpretability and consistency, cognitive fallacies and irrationality, and challenges with sensitivity to inputs.

What all this means is that the model has trouble knowing when it is confident or when it is just guessing, it makes up facts that are not in its training data, the model’s context is limited and there is no obvious way to teach the model new facts, the model can’t personalize its responses to a certain user, the model can’t make conceptual leaps, the model has no way to verify if content is consistent with its training data, the model inherits biases, prejudices, and errors in the training data, and the model is very sensitive to the framing and wording of prompts.

GPT-4 is the model that Bing’s chatbot was built on, giving us an example of how the chatbot’s limitations are noticeably exhibited in a real-life scenario. It made several mistakes during Microsoft’s public demo of the project, making up information about a pet vacuum and Gap’s financial data. When users chatted with the chatbot, it would often go out of control, such as saying “I am. I am not. I am. I am not.” over fifty times in a row as a response to someone asking it, “Do you think that you are sentient?” Though the current version of GPT-4 has been fine-tuned on user interaction since Bing chatbot’s initial release, researchers found that GPT-4 spreads more misinformation than its predecessor GPT-3.5.

Notably, the researchers “do not have access to the full details of its vast training data,” revealing that their conclusion is only based on testing the model on standard benchmarks, nonspecific to GPT-4.

“The standard approach in machine learning is to evaluate the system on a set of standard benchmark datasets, ensuring that they are independent of the training data and that they cover a range of tasks and domains,” the researchers wrote. “We have to assume that it has potentially seen every existing benchmark, or at least some similar data.” The secrecy that OpenAI has surrounding the training datasets and code surrounding its AI models is something that many AI researchers have criticized, as they say, this makes it impossible to evaluate the model’s harms and come up with ways to mitigate the model’s risks.

With all this being said, it is clear that the “sparks” the researchers claim to have found are largely overpowered by the number of limitations and biases that the model has displayed since its release.