Original video in Chinese.
Key Takeaway
- Perplexity is a phenomenal AI-native Q&A engine that has the potential to replace traditional search engines. Its core value lies in directly providing organized answers rather than web links.
- Perplexity expands search keywords through “Copilot” mode and offers a “Focus” option for specific types of searches, improving the user experience.
- Perplexity emphasizes rigor in its answers, provides clear source attribution, and supports multi-turn interaction and “Collection” features.
- Although it has been criticized as a “wrapper,” Perplexity has shown its technical strength by fine-tuning GPT models, using other large language models, and building its own open-source models.
- On top of search, Perplexity integrates powerful RAG technology and also has knowledge base functionality, with the potential to launch more products in the future.
Recently I saw a very forceful line:
A wrapper product with one hundred thousand users is more meaningful than having your own model but no users.
If you have a friend around you who invests in large language models or works on large language models, remember to pass this line along to him.
The person who drew so much hate is Perplexity’s CEO. They just completed their Series B round, with a valuation of $520 million. The participating companies include Nvidia, as well as big names like Bezos.
Perplexity’s product is a phenomenal AI-native application, a Q&A engine that has the potential to replace traditional search engines.
What is a Q&A engine?
Up to now, search engines have returned web pages. But are web pages really the result we want? What we want is the content inside those web pages. This is where the value of large language models shows up:
They help us go through all the pages we find, extract the relevant content, organize the logic, and ultimately present the result in one step.
This is something traditional search engine technology cannot do. That is also why search is a clearly defined track that will inevitably be completely transformed by large language model technology.
Over the past year, I’ve tried many AI applications. But the only two that I’ve kept using continuously, the ones I can’t do without, are:
- GitHub Copilot
- Perplexity AI
I strongly recommend everyone try Perplexity. It is an AI application that helps every single person. After using it, chances are you won’t need Google anymore, let alone Baidu.
I’ll do a demo on the web version. This product also has phone and iPad versions, which are very convenient.
Once you turn on “Copilot,” it provides more accurate and deeper answers, at the cost of being a little slower. The free version seems to give you 5 credits every four hours, and the subscription version gives you 300 credits per day, which is basically enough.
The “Focus” option is easy to understand: it lets the large language model focus its search on a certain type, such as academic papers, Reddit discussions, or YouTube videos. If you choose Writing, then it won’t go online, which is basically the effect of using the large language model directly.
Perplexity’s subscription price is $20 per month. From a practical standpoint, I suggest you can skip subscribing to ChatGPT Plus, but you should subscribe to this one. After all, search is a high-frequency need. Perplexity’s search is stronger than ChatGPT’s. And if you need GPT-4 to generate directly, just choose Writing mode.
Let’s do something simple. For example, search for “GitHub Copilot.” The large language model will first understand the question or keyword, and then expand on it based on that understanding.
Since we only entered “GitHub Copilot,” which is fairly broad, the large language model judges that the user most likely wants a preliminary understanding, such as what it is, what it is used for, what its pros and cons are, and so on. So it helps us do a series of expansions, then uses those to search, finds a bunch of sources, and finally gives the answer.
After the first interaction, Perplexity will guide the user to either query related questions or continue asking follow-up questions.
A multi-turn interaction expanded from one question or keyword becomes a collection, archived in Library, which is basically a history record that can be looked up again later or continued from. This is also one of the things I really like about this product.
“Collection” is the newest feature. You can make more detailed settings for a certain topic in the form of a prompt, and you can also share it with other people.
As for Discover, it’s the official trending feed. You can check it out when you’re bored.
Perplexity is widely recognized as the AI Q&A engine with the best user experience and the highest result accuracy.
Let me start with the user experience.
“User-centered” is not just a slogan to them; they really believe in it. I’ll give two examples.
First, why does Perplexity help users expand search keywords first?
Because most users don’t know how to ask questions.
Just like in the demo above, I only gave it one keyword. If it were a traditional search engine, because the user input is too little or too inaccurate, the results often wouldn’t be very good.
So is it the user’s fault?
The user is not at fault. It’s your technology problem, your product design problem. This is the reality application builders have to face.
Let me add one more point: I think this wave of large language model technology breakout is bringing not natural-language interaction between humans and machines, but intent interaction. A lot of projects are moving in this direction; it just depends on who gets there first. Back to the main point.
Second, Perplexity has already provided the final answer, so why list the sources?
Because users are always worried.
They worry both about the authority of your answer and whether the large language model hallucination will kick in.
Especially if some viewpoints in the answer don’t match what I expected, I will definitely go look through the source web pages or videos.
Perplexity is in the business of building products; technology is just the means of implementation. But that doesn’t mean they don’t have technical strength.
The reason their CEO said the provocative line at the beginning is because in the early stage, Perplexity, like many other projects, used OpenAI’s large language models, and then got labeled as a “wrapper.”
But is it enough to just hook up GPT-3.5 or GPT-4 and call it a day?
First, the GPT-3.5 that Perplexity uses is a version fine-tuned by themselves, with significantly improved performance, but at a lower cost than GPT-4 and with faster speed than GPT-4.
Second, besides GPT, they also use other large language models, such as Claude, because it supports longer context and is especially suitable for meeting the need for users to upload documents.
Finally, Perplexity knows it can’t rely on OpenAI forever. So they use open-source large language models for fine-tuning and built two large language models: pplx-7b-online and pplx-70b-online. The former is based on mistral-7b, and the latter is based on llama2-70b. These two large language models are specifically used to handle real-time data on the internet. And the fine-tuning work will continue as well, constantly improving performance. The training data was also prepared by them, and it is high-quality and diverse.
I estimate that when the performance of open-source large language models fully matches GPT-4, Perplexity will definitely use open-source large language models as its foundation and completely break away from dependence on OpenAI.
Having a large language model customized for search is not enough. To do this well, you also need very strong RAG technology.
So Perplexity is absolutely not a wrapper project; their technical strength is definitely not weak. At the same time, Perplexity is also not the kind of purely technical project that only focuses on technology; they know how to use technology to satisfy demand.
And search will definitely not be their only product. As large language model technology develops, this team will absolutely come out with more new products later on. That is also one of the reasons I will keep paying attention to them.