Yesterday, a friend recommended a search tool to me: perplexity. This is a search engine based on large language models.
After trying it out, I asked some relatively simple questions, and it felt much more accurate and faster than Google.
Google can only lead me to articles and websites written by others, but this can directly tell me the answer. However, this answer is not directly generated by the large model; it combines retrieval results, analyzes existing websites, articles, and videos, and then infers the answer through the large language model.
What surprised me the most is its ability to retrieve video content and perform inference, which is something Google cannot do at present.
For a question, Google can only recommend some videos, but when faced with two or three videos that could total ten minutes or even an hour, it is very difficult to quickly find the answer.
Perplexity can answer based on the content of the videos. I guess it must have transcribed the audio and subtitles of the videos into text and then processed it as natural language.
This is a very intuitive engineering technique, but it is very useful and functional. This is a feature I have always wanted, which YouTube has not implemented, while perplexity has. YouTube may have used this technology for video recommendations, but it has not allowed users to ask questions based on this technology.
After using it for a day, the most outstanding feature of perplexity is its speed; the responses are very quick. In contrast, GPT-4 sometimes really takes a long time to ‘think’. But this also made me wonder, if it is always this fast, can the accuracy be high?
Regarding this question, I looked at some evaluations, and the characteristic of perplexity is that it provides fast and comprehensive answers. It seems this is the product design feature: fast, comprehensive, and detailed.
About Accuracy
Regarding accuracy, this is a common issue with large models that cannot be solved for now. Large models will inevitably produce hallucinations, meaning they can make things up. This is something one must remain vigilant about. Good large models will have a much lower hallucination rate and high usability.
I compared the free and paid versions of perplexity. For simple questions, both versions have acceptable accuracy. However, the free version requires high vigilance for more in-depth professional questions.
I asked it, “In Unreal Engine, can modules written in Python (called by blueprints) be used at runtime?”
After several attempts, the free version consistently gave incorrect answers. In fact, Python can only be used in UE to write editor-related processes, and cannot be used at runtime (excluding third-party). Here, the free version confidently told me it could be used at runtime and even fabricated a lot of details.
Then I selected the copilot (paid version), and it provided the correct answer, and indeed it was comprehensive and detailed, as it thoughtfully listed which runtimes cannot use Python.
In addition, the paid version allows for deeper and more detailed Q&A. Intuitively, this is a very useful feature for content creators who need depth. It can even guide users and refine questions based on user needs.
Due to the much higher probability of the free version making erroneous statements compared to the paid version, it may be essential for professional and in-depth content creators to opt for the paid version. Whether to choose perplexity as the main tool is something each person needs to evaluate for themselves.
About Shelling GPT
Perplexity, along with the emergence of some other apps recently, has changed my perspective on shelling GPT.
Previously, OpenAI’s CEO Sam Altman said that shelling GPT would inevitably fail, and I agreed with that. However, my recent use of various AI products has made me feel that product design is also very important, independent of model capabilities.
Generally speaking, GPT-4’s capabilities are still the strongest, but sometimes it does not provide the best results and experiences. Instead, products that are deeply optimized for specific tasks are more useful, even if the underlying model capabilities are not as strong as GPT-4.
For example, tools like pdf.ai and perplexity.
In summary: the experience is good, but the old problem of large models remains—accuracy cannot be guaranteed, and vigilance is necessary. Creators should treat them as acceleration tools, not substitutes for their own thinking.