Can OpenAI stand to “reason”?

OpenAI has unveiled its latest AI model, once code-named Strawberry and now officially dubbed o1. The company behind ChatGPT claims that this model represents a significant leap forward in artificial intelligence capabilities, specifically that it can perform human-like reasoning and tackle complex problems in ways that previous models, such as GPT-4, could not.

But can we really call what o1 is doing “reasoning,” or is that simply marketing-speak for more sophisticated pattern-matching?

According to OpenAI, o1 uses a novel processing approach, similar to the chain-of-thought technique, for prompting chatbots. Chain-of-thought is essentially a set of instructions aimed at getting previous generations of large language models to process questions step-by-step rather than all at once. Large language models are best at guessing the next word in a sequence, and they’re not great at answering complex questions accurately, so this technique helps it break down the task into simple steps to minimize error.

Gadjo Sevilla, senior analyst for technology at the market research firm eMarketer, told GZERO that o1 “integrates human-level reasoning that can more carefully analyze prompts and requests and generate more analytical responses.” It’s better at things like complex mathematics, chemistry, and computer science, he said.

Sevilla noticed that o1 takes longer to respond than previous models, and it notes how long it’s been “thinking” and “formulating a solution,” along with other indicators that it’s answering more carefully. Sevilla says the reasoning function still feels very early, which is why OpenAI has released it as a “preview.”

“More than anything, the timing of the release seems to coincide with OpenAI's current bid to increase its valuation,” he noted. OpenAI is reportedly seeking a monster valuation of around $150 billion in its upcoming funding round, making it one of the most valuable private companies in the world.

But by using terms like “thinking” and “reasoning,” OpenAI is employing human language for its models, a common marketing technique for AI companies that seem to want to suggest that their products are capable of human levels of intelligence — even when they’re just really, really good guessers.