In the past month, at least three top-tier LLM companies launched "show your work" features—the AI black box era is being forcibly ended. From OpenAI's o1 to Anthropic's new Claude, to various fine-tuned models in the open-source community, everyone is doing the same thing: making AI lay out its scratchpad for you to see.
What this is
This discussion stems from the open-source community's focus on the <thinking> tags in model outputs. Simply put, previously when we asked AI a question, it just spat out the answer; now, before giving the final response, it first outputs an "inner monologue" wrapped in tags. This is the so-called Chain of Thought (the intermediate steps the model takes to derive a conclusion). It's like in math class: the teacher doesn't just give the final answer, but also writes the derivation process on the board. Exposing the intermediate process directly to the end user is an interaction paradigm shift that has only just happened.
Industry view
Most believe this is a necessary path to building trust. When AI's answers involve complex logic or high-risk decisions, a transparent reasoning process significantly boosts user trust—we only dare to use conclusions whose logic we can understand. But we must pay attention to the opposing voices: some point out that this not only significantly increases Token (model billing unit) consumption but also introduces a new risk of "fake reasoning." That is, to appear "deep in thought," the model might generate a bunch of seemingly thoughtful but practically useless nonsense, yet users have to foot the bill for this redundant compute. Transparency does not equal accuracy; this is a contradiction the industry has yet to resolve.
Impact on regular people
For enterprise IT: API cost structures must be re-evaluated. The Token fees generated by "thinking" may far exceed the final output, posing new challenges for budget management.
For individual professionals: The way we collaborate with AI shifts from "reviewing results" to "auditing processes." We need to cultivate the ability to quickly scan AI drafts and judge their logical trajectory, rather than blindly trusting conclusions.
For the consumer market: Users need to accept that "waiting for AI to think" is the new normal. AI assistants that reply in seconds often represent a compromise on depth of thought.