The AI research tool that saves me hours every week

And why it might revolutionize the search industry.

I've come across dozens, if not hundreds, of AI tools at this point, and there are only a few that I still use on a regular basis. Today I want to highlight one that I personally use for learning and research - Perplexity.

You might have heard of Perplexity before - I saw it mentioned on Twitter for months before I finally decided to take a closer look.

But as I used it more, I realized it's an impressive productivity tool. And over the past few months, it's helped immensely in my research.

What is Perplexity AI?

Perplexity is a new kind of search engine, built with large language models in mind. The whole experience revolves around having conversations, not clicking links.

When you're searching with Perplexity, it returns full-length explanations. These explanations are based on actual search results - they're fetched from an index, then passed to an LLM to analyze and write the final output. Free users can only access the default models, while paid users can use GPT-4, Claude 2.1, and Gemini Pro. But the most useful part might be the citations - each answer comes with inline links back to the original sources.

Of course, this system isn't completely infallible. The LLM may be citing bad internet data, or might misunderstand the question or the answers. But this setup goes a long way towards avoiding hallucinations, and I'm more likely to trust Perplexity's answers over ChatGPT's. It also helps that sources are ranked by trustworthiness - the Washington Post is going to be more heavily weighted than a random Substack, for example.

And in case you were wondering about the name: "perplexity" is a term from natural language processing (NLP) that measures how well a probability model predicts a sample.

How to use Perplexity

If you're researching a new topic or idea, the best way to start is by asking a question. Once you ask, you'll see the model processing and analyzing its sources. And after a few seconds, you'll get a concise answer, with a list of citations baked in (if you copy the response, the full list of sources will be copied too).

You'll also get a set of follow-up questions - the same as with Google. I've found that these are often more useful than my first question - I'll use them to quickly explore related ideas and different perspectives.

Each query/conversation can be put into a Collection, which is helpful for bucketing different topics and projects. I've found myself using one bucket for AI-specific research and another bucket for life admin/planning

Copilot

Let's talk about some of Perplexity's other features for a minute. A few months ago, on top of their normal search interface, they introduced Copilot.

As an aside: I'm getting pretty tired of the name Copilot for AI-powered features. We've got at least two distinct Copilot offerings from Microsoft and GitHub, more if you count each individual product Copilot, like Excel and Word and Outlook and Bing, as a separate entity. We need to find new names. It's too confusing.

At any rate, Copilot takes the search experience one step further and adds an assistant veneer on top of it. If you're asking about a trip to Japan, Copilot will ask additional questions to refine the search before giving an answer. You can also customize your profile (much like ChatGPT) to bake in extra context for your answers.

The free plan of Perplexity gives you a handful of Copilot-powered searches a day, while the paid plans offer effectively unlimited usage (600 searches per day). But Copilot isn't needed for simple searches - it's more helpful with planning complex responses.

Focus Modes

Perplexity also offers Focus Mode, which narrows your search into a specific domain. You can think of this like adding the "site:" parameter to a Google search. For example, you can restrict your search to Reddit, and Perplexity will "read" through dozens of Reddit comments for you. A great use case is product recommendations or opinionated answers - things I use Reddit for already.

There are other areas of focus, like Yelp reviews, YouTube videos, or Shopify products. These seem to be in beta at the moment, so you may not see all of them. There are also non-searching modes, which is pretty much a straight competitor to ChatGPT. We'll talk more about that comparison later.

But it's interesting to me how Perplexity's Focus Mode somewhat mirrors GPTs. In both cases, the systems are less effective by being general purposes. And users can get better results by paring back the given context that the model works with.

pplx

One last feature is Perplexity's API, which is a bit of a strange offering. The API offers hosted versions of open-source LLMs like Mixtral, Mistral and Llama. Since its launch, the company has also added its internal models (pplx-7b and pplx-70b) which offer answers based on grounded results.

At a high level, it seems like Perplexity is trying to 1) expose its own infrastructure for open-source models, and 2) offer a competing LLM that is more trustworthy and less prone to hallucinations. But it's still unclear to me exactly what kind of company is using this API, and for what use case. Some might want to "embed" Perplexity's assistant in their own application, but you can't yet add your own data/documents.

Perplexity vs ChatGPT

On the surface, Perplexity might seem like another ChatGPT competitor. Which is fair - there's a decent amount of capability overlap between the two. Ultimately though, they target two distinct use cases, and will likely keep coexisting.

ChatGPT is a general-purpose assistant - it can brainstorm ideas and draft essays. With its expanded toolkit, it can also create images and execute code. But it's not good at working with up-to-date facts or searching the web. It's not designed to do so - and as a result, it's internet browsing capabilities feel a little bolted on. (For what it's worth, Bing Chat is actually much better at this - and it feels much more internet-native than ChatGPT).

Unlike ChatGPT, Perplexity doesn't really have a "knowledge cutoff" - all the knowledge it's working with is as old as the last internet crawl it can access. It’s really great for diving down internet rabbit holes, and asking questions that I would have thought to ask before. But it's a little awkward outside of that domain. It might be a personal thing, but I don't go to Perplexity to try and write code.

So there's room for both - and based on the type of work you do regularly, it may even be worth paying for both. Though I'd suggest trying the free tier first, since it's pretty generous.

Perplexity’s threat to Google (and the internet as a whole)

There's one more reason why I wanted to talk about Perplexity, beyond the productivity boost it gives me. Perplexity is a peek at the future of search.

It feels like what Google Bard should have been. Instead, Bard feels like yet another ChatGPT competitor. Google has access to the world’s information, and is incredibly well-positioned to be build the ultimate AI assistant. Of course, to do so would mean cannibalizing its multi-billion dollar money printer: ads.

Tobi Lutke, CEO of Shopify.

When ChatGPT first made it big, plenty of pundits were debating whether it could kill Google. I was pretty suspicious. At the time, I wrote:

We're still finding our bearings when it comes to using LLMs for search. And while there are exciting signs, there are also discouraging ones. ChatGPT and Bing have a hallucination problem - you can't trust their answers to always be true. Generally speaking, your search engine straight-up lying to you is a pretty serious issue. While Google may be serving up links to the internet's misinformation, at least it isn't creating the misinformation.

Why ChatGPT won't kill Google

​

I stand by that today - I think any narrative of Google being killed by an AI upstart should be met with suspicion. The company is far too entrenched to be dethroned in a matter of months or even years.

But tools like Perplexity can force Google's hand. They can push Google down a path that they wouldn't otherwise go. Because if all my internet searches run through an AI assistant, it leads to a lot of thorny questions without easy answers. How are advertisers getting in front of their target audiences? How are publishers expected to make money from ad revenue when 90% of the traffic going to their website are bots and search engine crawlers? If my factual content is coming via an LLM, is it susceptible to prompt injection or other malicious text?

Like it or not, serving free content alongside ads is a critical piece of the internet. It hinges on the assumption that commanding attention will ultimately be monetizable. However, if that equation changes - if machines are reading the content and ignoring the ads - then some of the internet’s fundamental incentives are now upside-down.

I'm not going to get into the weeds on how that might play out - that's a topic for another day. But it’s clear that some of our simplest AI tools are already laying some big foundations - and helping us work more efficiently along the way.