A new tool has been released that lets users visually inspect an AI's 'Chain of Thought'—its internal reasoning process—and modify those intermediate steps to guide the AI toward a desired final answer.
Imagine you ask an AI assistant, “Describe a symbol related to the sea.” The AI pauses, thinks for a moment, and replies, “A wave.” But what happened inside the AI’s head before it chose the word “wave”? Until now, we could only see the AI’s final output. It was like seeing only the correct answer on a test, without knowing what logical errors the student might have made during the problem-solving process.
Recently, however, an interesting web tool that allows us to peer into this “black box” (the state where the internal workings are unknown) of AI has been released on the tech community Hacker News, sparking significant buzz Source: Show HN: I built a web tool to see and edit what an AI thinks before it answers Source: hckr news - Hacker News sorted by time.
Why is this important?
AI is now integrated into every corner of our daily lives. However, it is very difficult to confirm why an AI provided a certain answer or whether it made any logical leaps. This tool allows users to visually inspect the AI’s “thread of thought” and even re-weave it Source: Show HN: I built a web tool to see and edit what an AI thinks before it answers. This is significant because it transforms AI from a simple tool that unilaterally executes commands into a partner with whom we can collaborate by correcting logic in real-time.
Simple Understanding: AI Needs a ‘Working Out’ Process Too
To understand this technology, you need to know the concept of ‘Chain of Thought’ (the intermediate reasoning steps an AI takes to reach a logical conclusion) Source: Show HN: I built a web tool to see and edit what an AI thinks before it answers.
Using a simple analogy, it is like expanding complex equations step-by-step instead of just writing the answer when solving a math problem. When you ask an AI to “describe a symbol related to the sea,” it doesn’t immediately answer “wave.” Internally, it sequentially reviews various associated words like “sea,” “ripple,” “shore,” and “curve,” building logic step-by-step.
The tool released this time visually displays the process of the AI selecting these words, as if lights are turning on Source: ShowHN:IbuiltawebtooltoseeandeditwhatanAIthinks…. What’s amazing is that it doesn’t stop there. When the AI is about to go down a wrong logical path, the user can directly edit the content at that stage to correct it Source: Show HN: I built a web tool to see and edit what an AI thinks before it answers. For example, if the AI is thinking about the “sea” and you intervene to change the direction of its thought to “lake,” the AI will reconstruct its final answer based on that modified logic.
Current Status
This tool has currently been released by an independent developer, and anyone can enter their own questions to experiment with what the AI is thinking before it answers Source: ShowHN:IbuiltawebtooltoseeandeditwhatanAIthinks….
However, this technology is still in its early stages. Rather than being a universal standard applied to all Large Language Models (LLMs—AI models that learn from vast amounts of data to understand and generate language like humans), it is closer to a way of peering into and intervening in the reasoning process of specific models. Nevertheless, this attempt to transparently visualize and control the AI’s internal computational process is expected to significantly change how software engineers verify whether they can trust AI outputs Source: Reflections on AI at the End of 2025.
What’s next?
In the future, moving beyond simply commanding an AI to “write this,” it may become common to monitor the logical steps the AI is taking while writing and provide guidance in real-time. If an AI is reasoning based on biased information, the user will be able to immediately correct that reasoning step to obtain more fair and accurate results. This will be a technological advancement that shows just how finely we can tune and grow an AI’s ‘intelligence’ together.
AI Opinion
AI is evolving rapidly, but its internal structure remains a complex maze. If we can truly peek into and correct the AI’s thought process, AI will no longer be a technology to fear, but our most precise and reliable partner.
References
- Edits images generated by AI
- Views and edits the AI's internal 'Chain of Thought' reasoning steps
- Automatically protects the user's personal information
- Chain of Thought
- Auto Learning
- Image Rendering
- The AI stops working
- The user can guide the direction of the final answer
- The AI's training data is completely deleted