@tjrobertson52

So yesterday I signed up for Perplexity Enterprise for my entire team and today I canceled it. I know a lot of you love perplexity, so I wanna talk about what made me sign up for it in the first place and why I don't think it's a good fit for me. Hopefully this video will help other people decide if perplexity works for their case.

So this started a week ago where I made a video about all the large language models. And in it I said, I don't know why anyone uses perplexity anymore. And a bunch of people in the comments told me how perplexity is amazing.

And upon giving another look, I realised they added a bunch of new features. So if you don't know, perplexity is kind of like a large language model rapper, although they do have their own large language model and a bunch of cool features. There were two features in particular that made me immediately think, yes, this is what I need and I'm switching right now.

The first is that you can use most of the popular models inside of perplexity. For someone like me that was already paying for Claude, ChatGPT and Gemini for my entire team, this was very appealing. Each model has different strengths so being able to use all of them in the same platform is a big plus.

And there are other rappers that do this. Kicker for me in perplexity was the spaces Feature. Now this is nothing new.

It's the exact same thing you'll see in custom GPT's for Claude Projects or Gemini Gems. 90% of my team's LLM usage happens inside of one of these projects. Most commonly we'll create a brand ambassador project for each of our clients.

We'll load it up with everything there is to know about that particular client and then we can use that project anytime we're creating content or developing strategy or brainstorming for that particular client. This process by far has been the biggest advantage I've seen from AI. And the great thing about Perplexity Spaces is that we could share them across our team, which you can't do in Claude and Gemini.

So everything just seems perfect, right? That's why I bought enterprise. That's why I made the video yesterday about how excited I was about perplexity.

But over the last 24 hours, I spent a good amount of time with it and the output just wasn't hitting right. The first thing I noticed is that the brand ambassadors I made just weren't doing a good job representing the client. They were still giving me generic output instead of output that sounded like it was made specifically for that client.

So I did some research and I realized that Perplexity's context window is only 4,000 tokens by default. Now that can go up to 32,000 tokens on their most expensive plans, but that's still nothing compared to the hundreds of thousands or millions Of tokens on the other models. Not only is their context window smaller, but they're not doing as good of a job utilizing that context window.

And that might seem like a small thing, but this context window and how it's used is the No. 1 reason that models have become so much more useful in the last three months. It's the reason our brand ambassador projects in Claude are so amazing at what they do.

In Claude, I can load up hundreds of thousands of words of content about each client. And every time I perform a prompt, it has all of that context and it utilizes it really efficiently. After doing some research on perplexity, my understanding is that it is retrieving only the content that it thinks is relevant for your query.

And to save on money, it really skimps on how much it returns. Unfortunately, that alone makes it unusable for 90% of my use cases. The second problem is that even though perplexity is calling these other models, it's calling it through the API.

And so you're losing a lot of the infrastructure that's set up within the models themselves. Now, I knew this was the case, but I didn't know how much of a difference it would make. Again, after doing some research, my understanding is that within Claude, for example, instead of just sending your prompt to the API, it's running additional prompts to check and reformat its response.

The most significant example, I found of this, I asked both the Cloud Platform and perplexity calling cloud to write headlines and descriptions for Google Ads. Not only were the headlines and descriptions better inside the cloud platform, but all of them came out just underneath the maximum character limit. However, perplexity calling the exact same cloud model went over the character limit on half of the headlines and more than double the max character limit on all of the descriptions.

My understanding is that inside of the platform, Claude is doing something to double check its work. Feel free to correct me if I'm getting any of the technical details wrong, but at least for the way I'm using Claude, unfortunately, perplexity is just not gonna work. Most of the use cases I found where people rave about how great perplexity is, is where they're trying to learn something and that makes sense to me.

I do think perplexity is probably a really good teacher, but if you need to create content or do anything where large context windows are important, for now, I think it's worthwhile to at least pay for Claude, if not all three platforms.