Text is a Projection of the World: OpenAI’s Sutskever

OpenAI’s co-founder Ilya Sutskever believes that ‘text is a projection of the world’ and contrary to popular perception, ChatGPT is doing much more than just the surface-level learning of statistical correlations. Sutskever, during a fireside chat with Nvidia founder and CEO Jensen Huang at GTC, discussed ChatGPT and GPT-4 at length.

After an extensive discussion on the need for scaling to improve the performance of these models and laws surrounding scaling, Huang set the ball rolling when he said, “There’s a misunderstanding that ChatGPT is a large language model, but there’s a system around it”. He then went on to ask Ilya to explain the complexity and the working of ChatGPT.

Sutskever explained that in an attempt to accurately predict the ‘next word’, lots of different texts from the internet were consumed and in the process, AI has learned a world model. He said, “On the surface, it may look like we are just learning statistical correlations in text. But, it turns out that to just learn it to compress them really well, what the neural network learns is some representation of the process that produced the text. This text is actually a projection of the world; there is a world out there and it's as a projection on this text.”

He continued, “And so what the neural network is learning is more and more aspects of the world, of people, of the human conditions, their hopes, dreams, motivations, and their interactions, and the situations that we are in right now. And the neural network learns a compressed, abstract, and usable representation of that. This is what's been learned from accurately predicting the next word. Further, the more accurate you are at predicting the next word, the higher with fidelity, the more resolution you get in this process.”

Are LLMs Truly Intelligent?

Interestingly, the statement comes just weeks after celebrated linguist Noam Chomsky took clear shots at the vastly popular LLMs, namelyOpenAI’s ChatGPT, Google’s Bard and Microsoft’s Sydney, calling them ‘marvels of machine learning’ yet stating that they merely generate ‘statistically probable outputs’ that has ‘seemingly humanlike language’.

This aligns with a new theory that the scientific community is intrigued by, which suggests LLMs can contribute to the development of good theories of human cognition. Conversely, in a paper titled, ‘Sparks of Artificial General Intelligence: Early experiments with GPT-4’, a group of researchers from Microsoft wrote, ‘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. 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.”

Early version of AGI

According to the researchers, GPT-4 could be viewed as an early version of AGI, “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.”

The research paper aimed to prove that an initial non-multimodal version of GPT-4 displayed several attributes associated with intelligence. These findings were based on the description of intelligence given by a group of 52 psychologists, who defined it as a broad mental capability involving skills such as reasoning, problem-solving, planning, abstract thinking, understanding complex concepts, rapid learning, and experience-based learning.

The researchers aimed to prove that GPT-4 could tackle challenging and original tasks across a range of fields without requiring any special prompts, in addition to its expertise in language. Amongst the many tests, GPT-4 successfully passed the famous Sally-Anne False-Belief test from psychology. To ensure that the answer was not memorised from the training data, the test was tweaked and modernised.

However, experts like Yann LeCun, chief AI scientist at Meta, disagree with Ilya’s assessment. LeCun believes that, “Large language models have no idea of the underlying reality that language describes," he said, emphasising that a majority of human knowledge is nonlinguistic. Gary Marcus and other AI scientists too agreed with the notion that “LLMs don’t reliably understand the world”.