Today, we launched the second product of vibe coding: Hotspot Analysis
Let AI analyze the hottest topics in the crypto world, and you can grasp the hottest discussions in just 5 minutes.
This feature is quite common now, so what makes my effort different?
- An Interesting AI that Speaks Human Language
I made the reports written by AI more relatable and readable. This is the result of my careful tuning, which went through three iterations. The first version was created when I first started doing AI analysis, and like most people, I focused on functionality.
After completion, I found that the reports generated by AI were mediocre and filled with many obscure technical terms. So, I made improvements and began testing a dozen different large models and various prompt engineering.
I discovered that models have different personalities; some produce nuanced text, some are rigorous, and some prefer to be more creative. During this process, I also changed my perspective, shifting from a focus on professional depth to prioritizing the reading experience.
When I was about to finish, I read the second version of the report and felt that while the content was good, the expressiveness was lacking. So, I started adjusting parameters like temperature and fine-tuning prompts, which was a cautious attempt. It was necessary to allow AI to have its own personality while preventing it from going too far.
AI's analysis needs to be based on the provided data, not its own knowledge base. At the same time, it should be allowed to express the emotions in the tweet data. For example, when the market is filled with fear, AI's text shouldn't suggest that everyone is happy.
The final result was quite good; AI has a human touch. For instance, in its analysis yesterday, it humorously concluded:
"Don't think about 'When others are fearful, I am greedy'; it might just be 'When others are fearful, you go bankrupt.'"
- Carefully Selected Top 1000 KOLs
We all know that the quality of data determines the results. For information summarization, the choice of KOLs is also crucial. This is where KOL-Lens demonstrates its value. Based on its data, we selected influential KOLs in the crypto space with good track records and excellent data performance.
- Analysis Strategy
When the number of analyzed tweets reaches thousands, it cannot be fed to AI all at once. On one hand, it involves exceeding the maximum input token limit; on the other hand, thousands of tweets represent very chaotic information, and AI's ability to integrate and process this information is weak. In practice, the accuracy is not high. Therefore, I adopted a new analysis method. The specific steps are as follows:
Step 1: Data cleaning to remove useless junk data and improve AI's accuracy.
Step 2: Perform entity recognition analysis on each tweet for classification.
Step 3: Classify tweets that do not hit any entities by theme.
Step 4: Analyze the classified reports.
Step 5: Comprehensive deduplication.
The solution processed through these five steps can achieve comprehensive analysis in the context of large data volumes.
Website: http://riyuexiaochu.vip
To register an account, an invitation code is required. Please leave a message here. (Note: The KOL-Lens account is universal and does not require separate registration.)
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
