Original video in Chinese.
Key Takeaway
- Perplexity’s search advantage: it pulls authoritative financial APIs in real time (FMP, Morningstar, FactSet, SEC, etc.), attaches source links to every sentence, breaks questions into sub-questions, and uses strict system instructions, completely avoiding outdated training data and polluted web content. It is far superior to general-purpose AI like ChatGPT and Grok, and is suitable for serious work involving money.
- The essence of Perplexity Computer: a cloud-based agentic workflow that can automatically break down tasks, create sub-agents to run in parallel (up to 19 models collaborating), generate PDF reports, or continuously monitor tasks (for example, inflection points in U.S. and Chinese government bond yields). It supports custom Skills and is more professional and safer than OpenClaw.
- Usage advice: ordinary agents like OpenClaw have too much permission and insufficient accuracy, making it easy for major problems to occur. Perplexity Computer is the real serious tool. I strongly recommend it for investment decisions, research, and other high-risk scenarios. Ordinary people should not casually bet their money on outputs from general-purpose AI.
If you want to use AI to handle some very serious work, such as using AI to assist with investment decisions, then I recommend Perplexity Computer. Remember: do not use those general-purpose AI search tools, and absolutely do not use something like OpenClaw.
I’ll explain this from two levels: Perplexity and Perplexity Computer. Let me start with the first one.
I previously saw in an investing group that someone used OpenClaw to build a set of investment intelligence analysis systems, letting Lobster help collect materials and tag investment opportunities, generate daily briefings, and so on.
It looks very plausible, but to be honest, this whole thing is wrong at the root. Because general-purpose AI search simply should not be used in areas like financial investing. If you use it to collect news and take a casual look, that’s fine. But if you use it for investing, it really is not rigorous enough and not accurate enough.
These AI answers come from two sources: first, training data; second, information scraped from the web. But that is exactly where the problem lies:
Answers based on training data are bound to be outdated; answers based on web-scraped information often have not been cross-validated enough, and even the source material may already be contaminated.
And that is where Perplexity is strong.
It does not rely on the LLM’s “memory” to provide key numbers like stock price, EPS, or revenue. Instead, it pulls them in real time from specialized financial APIs. It integrates authoritative data sources such as Financial Modeling Prep, Morningstar, FactSet, and the U.S. Securities and Exchange Commission, among others. These core data points are fresh to the second and come from official sources or paid data vendors. So they are definitely rigorous and accurate, with an extremely low risk of source contamination.
Perplexity has also built its own index. Combined with the authoritative data sources just mentioned, it far exceeds ChatGPT and Google in accuracy, timeliness, and depth.
In addition, when Perplexity generates each paragraph or even each sentence, it attaches footnote-style source links. This mechanism forces the model, at the algorithmic level, to find solid evidence before it can output content. And for complex questions, it will also break the initial question into multiple sub-questions, perform multiple independent searches, and then synthesize, compare, and summarize the results.
Finally, the system instructions Perplexity gives the LLM in the background are extremely strict. It suppresses the AI’s tendency to go off in random directions as much as possible, ensuring that every word in the output stays as close as possible to the retrieved original text.
So if the work you’re doing has anything to do with money, especially your own money, you have to be careful. If AI scrapes a few websites and gives you an answer, and you dare to trust it, then if you lose money, who else is going to lose it? And data pollution is very serious right now, especially with AI-generated erroneous content. I’ve even seen some domestic broker reports get contaminated. Everyone really needs to be careful.
OK, what I just talked about was the foundation: accuracy and rigor. Perplexity is absolutely stronger than any peer in this regard. On that basis, they released Perplexity Computer.
Simply put, it’s something like Manus.
You only need to describe the final result, such as “Help me produce a deep-dive analysis of Nvidia’s fiscal 2026 Q4 earnings report,” and it will:
automatically break it down into subtasks; create sub-agents to run in parallel; run asynchronously, for example for hours or even months in the background, while you go do other things; and finally deliver a complete result.
Look at that — it actually generated a report in PDF format in the end. If you’re working at a company, this report could be handed in as-is.
What I just demonstrated was a one-off task. If you want to create a persistent task, that’s also possible.
For example, I had it create a monitoring task for “the synchronized inflection points of 10-year government bond yields in China and the U.S.” Because China and the U.S. are the main valves of global liquidity, only when both countries turn toward easing at the same time is there enough market liquidity for asset prices to surge.
These are all basic uses. Perplexity Computer can do the job beautifully, and besides the solid search fundamentals I mentioned earlier, there are two things I especially like and appreciate.
First, it uses a multi-model orchestration design, with up to 19 models working together.
At the core is the currently strongest Opus 4.6 model, which is responsible for analysis, decomposition, and orchestration. Other models, such as GPT, Gemini, and Grok, which are all top-tier industry models, are called by Opus depending on the situation.
Not relying on just one model, but using whichever is strongest and whichever is most suitable—I think this is the basic idea behind building multi-agent orchestration systems at this stage. I build my own newtype CLI the same way.
Second, Perplexity Computer supports Skills.
Although it is a cloud virtual machine, this Computer allows you to create Skills and give the agent specific abilities to complete specific tasks. Or you can use the official prebuilt Skills as well. I tried a few, and they were all very professional.
Looking at Perplexity Computer from today’s perspective, it may not seem especially dazzling. But I made this video specifically to emphasize one thing:
Starting with OpenClaw, the pace of agent adoption is accelerating. More and more people are starting to use them in daily life and work. The problem is that most people are far too trusting of these agents—for example, giving them the highest privileges and placing a great deal of trust in the results they produce.
If you’re just casually looking at news, that’s one thing. But if you’re using that intelligence to make investment judgments or conduct research, you really need to be careful.
When Perplexity does search, it has always pursued accuracy and rigor. Now, they are scaling up that core advantage to the level of an agentic workflow.
Compared with a year or more ago, Perplexity’s voice has almost disappeared. But this company is not dead, and it has brought out Perplexity Computer. This is not an AI toy; it is a serious tool.
OK, that’s the content for this episode. If you want to understand AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next episode!