OpenAI Unveils "Reasoning Model" o1: A "Humanized" Revolution in AI Thinking?

OpenAI Unveils "Reasoning Model" o1: A "Humanized" Revolution in AI Thinking?After weeks of speculation and anticipation, OpenAI has finally launched its first "reasoning model," o1. The product is touted as one of the company's most powerful AI offerings yet, showcasing unprecedented human-like thinking capabilities in its problem-solving abilities

OpenAI Unveils "Reasoning Model" o1: A "Humanized" Revolution in AI Thinking?

After weeks of speculation and anticipation, OpenAI has finally launched its first "reasoning model," o1. The product is touted as one of the company's most powerful AI offerings yet, showcasing unprecedented human-like thinking capabilities in its problem-solving abilities. At least, that's OpenAI's public claim. However, like many of OpenAI's previous research and product launches, o1 is being introduced with a degree of "teaser." OpenAI asserts that the model excels in handling complex tasks but provides minimal details about the model's training. Currently, o1 is only available in limited preview to paid ChatGPT users and specific programmers.

OpenAI confidently states that o1 has demonstrated PhD-level thinking depth in fields like physics, chemistry, and biology. This advancement is deemed so significant that OpenAI decided to start over from the existing GPT-4, resetting the model's numbering to "1" and even considering abandoning the widely recognized "GPT" brand. This brand not only defines its chatbot but also signifies the rise of the entire generative AI field.

OpenAI's research report and blog post published today showcase o1's remarkable abilities in tackling complex reasoning tasks. These tasks span higher mathematics, programming puzzles, code decryption, and even specialized problems from genetics, economics, and quantum physics. Numerous charts demonstrate that o1, in internal company evaluations, has significantly surpassed its top language model, GPT-4, particularly in programming, mathematics, and scientific domains.

The key to these advancements stems from a deep insight into childhood education "Think before you speak." OpenAI claims that OpenAI o1 takes more time to "ponder" before providing answers, mimicking human thought processes. The company calls this process "chain of thought," a term in AI research referring to a strategy of breaking down problems into multiple intermediate steps. This "chain of thought" mechanism enables the model to solve small tasks incrementally, self-correct, and optimize solutions.

When users pose questions to o1, the model displays "Thinking" and subsequently reveals some steps in its reasoning process, such as "tracing historical evolution" or "integrating evidence fragments." Finally, it indicates the duration of its thought, like "Thinking 9 seconds," before giving the answer. Although the full "chain of thought" behind o1's answer generation is not visible to users, simplifying the user experience, this also sacrifices transparency, making it difficult for users to understand how the model arrives at its final conclusion. This strategy also serves to protect the model's core technology from competitors.

OpenAI offers scant details about o1's construction, only mentioning that its training relies on "novel optimization algorithms and training datasets." Despite OpenAI's unprecedented marketing efforts, it remains uncertain whether o1 will usher in a revolutionary experience for ChatGPT or merely represent an incremental improvement to existing models. However, based on the company's showcased research findings and my initial testing, o1's outputs do appear more comprehensive and logically grounded. This reflects OpenAI's confidence in the scale effect: larger AI models, more data, and more powerful computing capabilities will drive leaps in AI performance.

The longer o1 trains, the better it performs. However, extended thinking comes with higher costs. OpenAI allows programmers to pay for using its technology, and o1's cost per word output is roughly four times that of GPT-4. The high-performance chips, electricity, and cooling systems required for generative AI are prohibitively expensive. To accommodate these massive computing demands, tech companies, energy companies, and other industries are expected to invest trillions of dollars. This has sparked concerns: Will AI become the new bubble, similar to the cryptocurrency or internet bubble eras?

As o1 requires more time to respond to questions, its resource consumption increases, further fueling the uncertainty around when AI technology might become profitable. This extended processing time, perhaps most significantly, impacts not just the technical or financial burden but the rebranding of the technology.

Compared to the obscure terms from past AI models like "transformer" and "diffusion," OpenAI's "reasoning model" and "chain of thought" sound closer to everyday language, carrying a "humanized" tone. This linguistic strategy is not unique to OpenAI. Startup Anthropic portrays its flagship model, Claude, as having "personality" and "mind," Google heavily promotes its AI's "reasoning" abilities, and AI search startup Perplexity claims its product "understands you."

OpenAI's blog directly states that o1 "thinks like humans," "works like real software engineers," and "possesses human-like reasoning abilities." While research lead emphasizes that OpenAI doesn't consider its product equivalent to the human brain, he also acknowledges that o1 does appear more "human-like" in certain respects compared to previous models.

For an industry with a yet-unclear product positioning, the "humanized" expression is undoubtedly a powerful marketing tool. The definition of intelligence itself is ambiguous, and the actual value of language models is difficult to assess accurately. The name "GPT" appears simple but fails to convey any real meaning. Although OpenAI's Chief Research Officer, Bob McGrew, believes that OpenAI o1 is the first step towards "more sensible naming," aiming for clearer expression of its products, these subtle differences in letter and number combinations often hold little significance for the average person.

However, marketing a tool that can "think like you" is entirely different from scientific lab terminology; it's more akin to a literary concept. Such descriptions, while no more precise, and perhaps even more vague, than other AI terms, possess a unique charm. An AI model claiming to "think like humans" leaves room for imagination, allowing each user to fill in the blanks and imagine a machine that "operates like me." Perhaps the key to selling generative AI lies in letting customers build and fill in the "magic" themselves.