Everyday AI Podcast – An AI and ChatGPT Podcast - EP 464: Perplexity Deep Research - What it is and if you should use it

Episode Date: February 18, 2025

It's free. It's fast. But is it good? Perplexity joined the Deep Research train, so we are giving it a thorough rundown. Is this your next AI sidekick? Newsletter: Sign up for our free daily newsletterMore on this Episode: Episode PageJoin the discussion: Ask Jordan questions on PerplexityUpcoming Episodes: Check out the upcoming Everyday AI Livestream lineupWebsite: YourEverydayAI.comEmail The Show: [e-mail] with Jordan on LinkedInTopics Covered in This Episode:1. Breakdown of Perplexity Deep Research2. Comparison with Other AI Deep Research Models3. Live Demonstration and Deep Research Prompts4. Differences and Mechanics of Deep Research Models5. Results and Analysis of Perplexity's Deep Research QueriesTimestamps:00:00 "Your Everyday AI: Resources & Newsletter"03:51 Perplexity Deep Research Overview07:16 "Deep Seek Truth Episode 460"10:57 "Generative AI Partner Opportunities"15:32 Evolving Importance of Benchmarks19:45 Perplexity: An Answers Engine Competitor22:36 Perplexity's Overwhelming Model Complexity26:15 Researching Nike's Q4 2024 Earnings28:45 Enhancing Language Model Use Skills31:55 Importance of Citing Statistics34:31 DeepSeek's Global Tech Impact38:23 "Fact-Check AI with Personal Data"42:04 AI's False Claims Exposed45:39 AI Query Results Irrelevant48:39 Unrelated Thoughts on Criticism50:22 Comparing AI Research MethodsKeywords:deep research, perplexity, AI companies, tech companies, AI tools, GPT-4, Google Gemini, OpenAI, AI strategy, reasoning models, internet connected models, perplexity deep research, chat GPT search, Google's deep research, OpenAI's deep research, AI benchmarks, humanity's last exam, AI hallucinations, pro search, reasoning search, everyday AI, AI newsletter, AI podcast, AI career growth, generative AI, AI tools comparison, perplexity Sonar, transformer models, reasoning models, AI queries, large language modelsSend Everyday AI and Jordan a text message. (We can't reply back unless you leave contact info) Start Here ▶️Not sure where to start when it comes to AI? Start with our Start Here Series. You can listen to the first drop -- Episode 691 -- or get free access to our Inner Cricle community and all episodes: StartHereSeries.com Also, here's a link to the entire series on a Spotify playlist. 

This is the Everyday AI Show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life. Meet Firefly AI Assistant, now live in Adobe Firefly, the all-in-one creative AI studio. Just describe what you want to create and the assistant handles the rest, orchestrating multi-step workflows across Photoshop, Premiere Express, and more in one conversational interface. You direct the outcome. The assistant accelerates execution. Perplexity is getting into the deep research game.

Yeah, there's a deep research game. Now, all the big AI and tech companies are trying to play there. So in today's show, we're going to do a quick overview of perplexity's version of deep research. Talk about what makes it kind of unique and maybe some potential use cases that you could use it for your business. And we're going to do a live run-through. We're going to put in a couple of prompts for deep research, the perplexity version, and then kind of see exactly what we get. All right, I'm excited for today's episode.

I hope you are too. And welcome to Everyday AI. What's going on, y'all? My name's Jordan Wilson, and I'm the host of Everyday AI, and this thing's for you. It's your daily live stream podcast and free daily newsletter, helping everyday people like you and me, not just learn AI, but how we can actually use it to grow our companies and to grow our careers. So if that sounds like you, welcome.

You're in the right place. Maybe it's your first time listening on the podcast. Thank you. Please make sure to subscribe, follow the show. But after you're done, or you can just check in the show notes right now and go to your everyday AI.com. There, you're going to want to sign up first for our free daily newsletter. We're going to be recapping today's show.

And we recap the show every single day, bringing you additional info. So if this one hits you, then you're going to want to make sure to read today's newsletter. But also while you're there, there's more. more than 450 a free catalog of back episodes, videos, resources, literally everything that you need to become the smartest person in AI at your company. It's all on our website for free, on bias, giving it to you straight exactly how it is. So make sure you go check that out. While you're there, also make sure you go check out our 2025 AI roadmap series, y'all.

I don't care. It's whatever I'm on February. It doesn't matter. This is still required listening. So make sure if you haven't already, go listen to episodes 443 through 447. They're quick, about 25 minutes each. So make sure you go check that out. All right.

We're going to have the AI news in the newsletter. You know, there's too much to get to in this. I want to keep it tight. So let's talk about it. And this happened over the weekend, y'all. Like, what's up with all the AI companies now coming up with, you know, new releases and new features over the weekend? Don't they know people, you know, need to rest, right?

There's too much going on in AI. We need to rest. Well, I didn't. I played with deep research and have been using it quite a bit since it came out. So let's go ahead and dive straight into it and talk about perplexity, deep research, what it is and if you should use it or when or how. All right. So if you see this symbol around for our live stream audience, right, this futuristic, new,

little deep research, that's perplexities. So, yeah, they just dropped this via a blog post. in a Twitter announcement over the weekend. It's gotten some mixed reviews so far, to be honest. And I've even had some mixed reviews of it myself, but I said, hey, let's do this live. And again, I appreciate and I'm glad when people reach out and they're like, hey, Jordan, thanks.

We trust everyday AI for making all of our decisions great, right? But you need to be testing these things out yourself. Let me just say this right now. A couple of use cases, a handful of use cases is never enough to go on if something's good or bad. If there's problematic outcomes or outputs, absolutely, right? Because if that happens once, it doesn't matter if you use it 10 times a month or 100 times a day. If you get one pretty bad output, you know that you have to keep that in mind if when and if you're evaluating if some of these options are good fit for your company or your department.

All right. So here's the jest of what deep research is. And this is from Perplexity's blog post. So they said, perplexity deep research performs dozens of search, reads hundreds of sources, and reasons through the material to autonomously deliver a comprehensive report. Okay, they said right now, deep research is free for all. Pro subscribers get unlimited deep research queries, while non-subscribers will have access to a limited number of answers per day. So yeah, the last I read, if you are, if you have a free perplexity account, you will get five deep research queries a day. That could vary, right? If this gets very popular, they might make

that, you know, two. That's the latest number that we heard from Open AIs version of deep research. It's not, it's only for the pro subscribers on the $200 a month plan and they might end up giving two a day to free users. But for perplexities, deep research, I mean, this is pretty good, right? because Google's variation is paid only right now. Open AIs is paid only, although they will be rolling this out very limited free. But out of the gate, you got to like this. Perplexity has five free a day.

So automatically just for that, just for accessibility, I think a huge plus for perplexity. All right. So we're going to get into a little bit more how this works, but we're just going to go straight in and do it live because sometimes these take a while. I actually think that perplexity's version of deep research is going to be much faster than others. So I'm going to go ahead.

I have a couple of pre-typed out prompts that we're going to put in to deep research. And then we're going to go backwards from there. So we're going to give it some time. We're going to kind of launch all of these and go. So the first one that I'm doing, I'm saying provide in. And this is the exact same thing I did for 01's deep research. I'm doing two of the same and then I have another one that I add it. So I'm saying, provide an analysis of Nike's latest quarterly performance compared to

Adidas and underarmors include key financial metrics, recent earnings commentary, and relevant market news cite all sources and highlight any discrepancies in analyst opinions. All right. So this one is pretty detailed, right? We're not giving deep research a lot of room to operate and make its own decisions, right? So in this example, we are being a little more open-ended. All right. So let's go ahead and do our second one.

And again, these first two are the same things that we did for our episode. If you want to go and listen to that one or watch that one, let me see what number was that. That was deep research. There we go, 454. So if you want to go listen to that, it was episode 454. All right. So the second one that we're going to be doing in perplexity, deep research, let me go ahead and share my screen there for our live stream audience.

More open-ended, right? Just saying find the latest news and information about deep seek. All right. There's another one. That was a, you know, controversial. We'll say that. But yeah, if you care about deep seek, if you want to know the truth, go listen to episode 460.

All right. So that is our second deep research query that we put in. And then we're going to come back and check on them in a couple of minutes. So then the, so first one was. an analysis of Nike's latest quarterly performance compared to Adidas and Under Armour. The second one, more open-ended. Find the latest news and information about DeepSeek. And then we're going to go a third one here.

All right. Well, first we'll even see if we can run three concurrently. You know, Google's version of Deep Research as an example only allows you to run two. All right. So now my next, or my last one here and just FYI, I am on the $20 a month paid plan. for perplexity. So in this instance, I have unlimited usage of deep research. All right. So now the last one, this one, like I care about fact checking, right? And for some of these things, I don't know everything about Under Armour and Nike, right? Although I did spend better part of a decade

partnering with Nike. I don't know all that information. DeepSeek. I know a ton because I just did a very in-depth report. I probably spent, I don't know, at least 30 hours. I kid you not. over the past couple of weeks, just reading, researching about deep seek. But at least with this one, about myself, I know this. So I'm saying, tell me everything about Jordan Wilson, who does everyday AI from birth until today, make it creepy in depth. Yeah. Sometimes I have tried a lot of these before with other deep research type tools.

And this is a good one for me, at least, right? I know myself. I give enough information to the tool. and there's plenty about myself on the internet, right? I've written stuff on some of my old company websites. You know, I was a journalist for 10 years. You know, so there's plenty. I've been on other people's podcasts a lot.

So, you know, there's plenty of information about me on the internet. All right. So we have those three queries. So in about five or 10 minutes, we're going to go back and check on those. All right. But let's go ahead and jump back in to a little bit more about perplexity's deep research. And you already just heard me say, this, right? The exact same name. Everyone's got the same name. Deep research, right? So

perplexity now has deep research. Google has deep research. They were actually the first to market with it. And then Open AI has deep research as well. So a little confusing, right, when people are just talking about deep research, but it almost seems it's taking on this kind of co-pilot kind of mentality, right? When everyone has a co-pilot, you know, I guess technically Microsoft GitHub co-pilot was one of the first co-pilots to market. And then Microsoft obviously has the Microsoft 365 co-pilot, but there's so many other like, quote-unquote, AI co-pilots. And it's almost just become this general term. And it looks like that's what's going to happen with deep research, right?

It looks like that is just going to become a term. And I'm guessing we're going to see this from a lot of other companies as well. It's starting with the big companies that already had the data. They already had the architecture and systems. in place, right? Because all three of these, right? So Google with their Gemini, perplexity, and Open AI, they all already could browse the internet, look at a lot of web pages, and essentially try to decipher what's good and what's not, and put together a short response, right? So the biggest differences with this deep research, if you haven't used them versus a traditional, right,

like chat GPT search or using Google Gemini, which is now finally better connected. to Google, right? You might be wondering, like, what's the difference, right? Those queries that I put in, you could probably run those in chat, GPT search or a normal Google Gemini and get a pretty decent result. Yes. So the difference is these deep researches, well, they do that. They take it much, much deeper. And the three of these work in a very different way. There's some similarities between the three. There's a lot of differences. But, you know, also, it's worth noting, right, because everyone's like, oh, everyone's copying Google. Well, Maybe because actually before Google released deep research about six months before that,

there was some reporting and we even covered it at Everyday AI. There was reporting that Open AI was working on a product that did quote unquote deep research. So even though, yes, Google was the first to market with it, the kind of deep research kind of feature had been tied to Open AI many months before Google's version came out. Hey, live stream audience, I'm thinking about just, well, no, I am going to do it. I'm going to do a, we're going to do a comparison of all three of these tomorrow because I was going to do that for today's show. And then I'm like, oh, man, that's going to accidentally turn into like one of those hour long shows. And I don't want to do that, right?

I get tons of emails for people. They're always like, oh, I listen to you on my runs, right? And I just run until the podcast is over. And I'm like, that's dangerous. because sometimes I accidentally do like an hour plus podcast and maybe you weren't prepared to do a half marathon or a 10K. So I'm going to try to keep today's episode short. But tune in tomorrow if you're interested. I'm going to try to get a rubric of sorts.

You know, aside, I'll probably do the same three questions I just did here and a couple of others and kind of go over some of the pros and cons, comparison, similar, the differences in these tools. but for the most part, at least as it pertains to perplexity, I think there's some big advantages. I think speed, I think it is a little faster and cost, right? So, I mean, those are two of the biggest factors when people are looking at, right, when choosing which AI tool do I want to use or should my company be looking at. So right away, at least as of today, perplexity is the only deep research tool that you can use for free, right?

Like I said, Open AI said that they're going to be rolling out, I think, two a month because their deep research tool is utterly fantastic. All right. But hey, perplexity already has a nice competitive advantage. But tune in tomorrow will be doing a much deeper dive on them. All right. A couple other things. And this is from perplexity.

So they said that they're doing great in terms of benchmarks on this. So there is a new kind of a newer benchmark out there called Humanities Last exam. So a lot of times it's like you're trying to figure out like, hey, are these models good, right? Or what's the difference between, you know, GPT40 latest, the one that just rolled out like a couple of days ago versus the GPT40 latest that was out, you know, a month ago or three months ago, or, you know, the Gemini 1206 versus the Gemini 121, right? It's, it can be hard, especially when these big AI tech companies, you know, aren't the best at naming their models. And you always don't know which version of a model you're using.

Benchmarks are extremely important, right? But there's been a lot of, you know, talk over the last couple of years that benchmarks aren't as important, you know, in 2025, maybe as they were in 2023, mainly because, you know, the argument is a lot of these benchmarks kind of get thrown into the training data. So, you know, the argument is, well, well, the models just, you know, measure, they essentially memorize, you know, a lot of this, you know, a lot of these questions that are on these benchmarks. Right. So this new one, humanity's last exam, it's pretty difficult, right? And it's all this newer information that requires a lot of, you know, reasoning, a lot of logical thinking and just a lot of, you know, math, coding, STEM. All these different aspects are required to get a good score on this humanity's last exam. So perplexity did share that their deep research version scored a 21% on humanity's last exam. which is the second highest score out there next to Open AIs, which I believe they got a 26%.

I don't understand why perplexity did this, if I'm being honest, because it's a different type of tool, right? So we have these, you know, kind of the transformer tools or the transformer models, right? Then you have now your reasoner models, and those are probably going to get merged anyways. So we'll probably have a show on that and what that means. So you have your, you know, your quote unquote traditional transformer. models, right? So that's your GPT-40, your Gemini 2, your Claude 3-5 Sonnet, etc. Then you have your reasoner models. You have these models that think. OpenAI 01, 03 Mini, 03 Mini, hi, right? Then you have Gemini thinking. And now you have these deep research models, which are essentially fine-tuned

versions that usually include a reasoning model as well. So I believe, and hey, I guess I'll find out when I do a little more research, but I do believe that open AIs and perplexities uses a reasoning model where right now Google's does not. And you'll kind of see that hopefully in the results and in our show tomorrow when we compare them all. But I didn't understand why perplexity included this because essentially there's only two deep research models that have done this humanity's last exam benchmark and perplexity isn't second. But, you know, I don't know. It is what it is, I guess. You know, they're like, all right, well, hey, we can say we're second on the list and put up a bunch of, you know, other models that it doesn't necessarily make sense

because they're different classes, right? But I get it, right? There's not, you know, 50 different, you know, deep research models. There's not 50 different reasoner models. But that's important when you look at benchmarks and you hear people talking about this, right? So let's say you're the CTO at a medium-sized company and someone's coming out and, you know, spitting these facts out at you and being like, oh, we got to start using it. Okay, well, hey, is it a transformative? model, is it a reasoner model, or is it a deep research mode, right, that you're looking at these metrics? It's important to understand the difference.

All right. So how the heck do you use this thing? Well, it's simple. And then we're going to check in here on some of our results. And you know what? I might actually, I wasn't planning on this. I might run some of these same queries in the reasoning mode. So perplexity.

it is a very unique product, right? And I've been a paying subscriber to perplexity, probably not the day it came out, but the week or the month, pretty early on. And if I'm being honest, I think very early on, I was extremely, extremely bullish, right? Loved it. Been a little bearish over the last, you know, year or two. Or maybe year. I feel in my experience, and again, this is limited, but I've read plenty.

I feel perplexity's hallucination rate has not kept up with the rest of the industry that has internet connected models. And we'll see. I mean, we'll see if I get, you know, in our little results here if that happens. Anyways, there's a lot of different things to keep in mind about perplexity. It is a little different. For the most part, it is an answers engine. So think of it more as a direct competitor to something like chat GPT search versus, you know, know in OpenAI's, you know, GBT4O or something like that.

So you can choose a different large language model that it runs off of, right? So perplexity does have their own model called Sonar, but I don't think anyone that uses Perplexity uses Sonar for the most part, although they did just release an updated version that I think is really good in making it available via the API. But for the most part, when you're using Perplexity, even if you have a free plan, you can use their pro search. So for their pro search, again, this is what perplexity is kind of bread and butter is.

It's going through a lot of websites. It was, you know, one of the first players in the AI space that did exactly this, that did, it was a precursor for deep research, right? So one of the first iterations of perplexity and why I think it was pretty amazing is you would ask it a query. You would choose a model, right? So if you had the paid version, you could choose, hey, I want to use, you know, GPD 40 or Claude Sonnet 3.5.

So you can choose a model that actually powers the answers engine. And then even the first version of perplexity, what it would do is it would generally go to 10 to 20 different websites and it would try to get a better idea of what you were asking. Right. So it was almost like unsteerable rag, right? So one of the things when you're working with a model, whether it's GPT40, whether it's, you know, Google Gemini, I, you know, pro to whatever the model is, you always have to keep in mind. There's probably a lot of old data in that model, right?

Even if the knowledge cutoff as an example is June 2024, like GPT40, a lot of that information that is in the data training is probably very old, right? Just because, right, there's people get this false sense of security when they see a knowledge cutoff date, right? And they're like, oh, June 2024, okay, that's not bad, right? No, that just means that's when the knowledge cut off was. That doesn't mean 100% of what goes into these models that have trillions of parameters. That doesn't mean that it's all current and all correct and all accurate through June 2024.

Absolutely not, right? A lot of these data sets are probably offline and have a lot of older data. You hope through the reinforcement learning with human feedback or, you know, the RL stages where humans are going in and training the model that they're taking out some of that old data. but it's not always the case, right? So large language models, when they're using their own kind of internal data, it can be old, it can be outdated, which is why even the original perplexity, I think, was a very innovative product, right?

Because what it would do, essentially, it was this version of rag, right? Almost uncontrollable, though, right? So before it would kind of rely on whatever base model you chose, it would first usually go to 10 to 20 websites and, you know, verify that information, see if it's up to date, right? but then use also the training data in whatever model you chose. So that's kind of how perplexity. Perplexity has always been ahead of the game when it came to kind of this concept of deep research

or pairing internet research multiple pages with a large language model. But it's actually kind of confusing, right? And I think perplexity is now suffering from kind of what OpenAI and Google are as well. There's just too many models, too many features, right? And if you're an average user, it can be a little difficult. Because if you're looking at my screen right now, right, I'm on a perplexity pro plan. So I have auto, which it says best for daily searches. I have pro search, which essentially does three times more sources and detailed answers.

Then I have what we're going over today, which is deep research. And then we also have reasoning, right? So this is the reasoning search. Think of it like the normal pro search, but the pro search uses a. transformer model. And then the reasoning search is like a pro search, but it uses a reasoning model. You can choose either 03 mini from OpenAI or R1 from Deepseek. So it's a little confusing, right? You know, in terms of like, hey, how should we be even using something like perplexity if you have a paid plan? All right. So but in this instance, when we went in, we went in, we chose deep research.

All right. And then here's what perplexity said. So they said. said it takes about two to four minutes and they said it's reasoning, its research with reasoning. This is what perplexity says. So it says equipped with search encoding capabilities, perplexity's deep research mode iteratively searches, reads documents and reasons about what to do next, refining its research plan as it learns more about the subject areas. That is big and that is one of the differentiators, at least from Google. Again, tune into tomorrow's episode.

We'll tackle that a little bit more in depth. But that's important to know because perplexity's deep research, it is iterative, right? So if it starts going down path A to, you know, answer your query, but hey, oh, something's changed. Maybe it needs to go down path B or path C. That is something that Google's deep research does not necessarily do by default in open AIs does. So that's important. It does use a reasoning model to steers. itself to give you the best answer using the deep research.

All right. So let's go ahead. Let's check in. Let's look at some of these results. All right. And then we'll kind of give you my hot take. All right. So here is the first one we did.

Provide an analysis of Nike's quarterly performance compared to Adidas and Under Armour. Include key financial metrics, recent earnings commentary, and relevant market news, cite all sources and highlight any discreference. and analyst opinion. All right. So I'm going to zoom out a little bit here. All right. So you'll see what happened in deep research. I always encourage people, right, especially with these reasoning models where you can kind of see their chain of thought or you can see their rationale. You know, Open AIs is a summarized chain of thought, but they just improved the summarized chain of thought about a week ago. All right. So you can go through and see how deep research is doing this. So the first

thing, it says, I need to search for information on Nike's Q4, 2024 earnings report to gather key financial metrics, recent commentary in market news. This will help establish a baseline for comparison with Adidas and Under Armour. I am looking for specific data such as revenue, net income, growth rates, and notable trends or challenges highlighted in the earnings commentary. So then essentially what it did, it went to a search engine and all it typed in was Nike latest quarterly performance earnings report, 2024. And you'll see for this first step of its research, it went to, let's count them. So we got four, eight, it went to 10 different searches.

All right. So some of these searches, interestingly enough, perplexity does have a new, I'm just going to go ahead and open this. So perplexity does have a new kind of finance mode. So it essentially launched three different queries for Nike, Under Armour, and Adidas. So I should probably bring up the correct slide there for our live stream audience, right? So it launched into this kind of financial view of Nike, which is super helpful, right? And then we go back and we look.

It did the same thing for Under Armour and Adidas. So it ran some of, it ran additional queries on perplexity. which I think is really cool. It went to Investors. com. It went to a couple Yahoo finance articles. It went to CNBC as well. All right.

So pretty good. So then you can see it says, from the most recent search, I found detailed information about Nike's latest quarterly performance, including its Q4, 2024 results. All right. So then it's starting to report back, right,

what those results were. And then it's going on to its next version. So it's saying, now I need to gather similar information for Adidas to compare its financial performance strategies and market challenges. All right. So then similarly, it looks like it goes into about, let's see, about eight or nine additional searches looking at Adidas. All right. It replies back. Then it says, I need to do the same for Under Armour.

So it's doing, you know, both traditional open-ended searches, right? under Armour Q4 earnings report. Same thing. It went to about nine different sources. And then from here, it's pretty impressive, right? And this is why I always encourage people. If you want to become better at using large language models, especially ones that reason and are connected to the internet, you're always going to get better results the second time you do this, right?

Because what you can do is you can go and learn from exactly what this model did. Right. So this is mimicking human behavior when you give it the ability to use a reasoning model and it can change its original course of research. This is huge. So in the same way, this is like if you could go give an assignment to an analyst on your team and you were and you also paid someone to audit them, right? And write down every single step. And then they gave this to you, right? If you are smart, let's just say you're a senior analyst, right? This is maybe what a junior analyst might do. But you could go through and see exactly step by step what they did correct and what they did incorrect. And then you can go through in theory, you could run this again, right? And then have a little bit more pointed or a little bit more detailed instruction.

So in this prompt, it wasn't super restrictive. It wasn't pre-steered too much. It was kind of open-ended, right? So that's where I think it's extremely important. Don't just use these to save time, right? Use these to improve your next round of outputs. I always tell people, never run something like this, a deep research query once. Don't do it.

Same thing, even with a reasoning query, never run it once. If you really want to grow your company, if you really want to become the best person in AI, study the chain of thought. Study what these models are doing under the hood. That's how you get better. You don't get better by just blindly, you know, trying to hand off as many tasks as possible, right? Because these models still aren't that great, right? They're the worst today they ever will be. But you can make them better if you understand how they work.

All right. So let's just go ahead. Let's skip to the end here. We have an answer from perplexity deep research. It says comparative analysis of Nike's, Adida, and Under Armour, Q4 performance, and market dynamics. All right. So first, we have Nike's. We have a couple paragraphs here about Nike, their financial performance, operational challenges, market reaction, and analyst sentiment.

Same thing with Adidas, financial highlights, strategic drivers, analysts and credit outlook, under armor. They said it's a lagging contender. All right. So financial and market position, strategic missteps, analyst perspectives. Then we have a nice chart here. Did a really good job. So we have the quarter four revenue, full year revenue, gross margin in Q4. operating profit, market cap, great, great charts here from perplexity.

Unfortunately, these metrics are not cited in the chart, which I wish they were, right? I could go through in the, I can go through in the rest of this answer and probably find those citations, right? And this is one of those things I would say, in my follow-up, I would say, hey, make sure when creating a chart to cite each individual statistic in the chart. Because again, human in the loop, this changes. This probably, you know, it's not the most impressive deep research I've ever seen, but it was fast.

You know, I don't see anything off the top of my head that looks inaccurate. But again, I don't know this information 100%. So it wasn't actually a very in-depth report. It was extremely detailed, extremely specific. It looked like only a couple hundred words. So then what you can do at the end is you can, a couple of things. Adobe just introduced an entirely new way to create, bringing the power and precision of its creative suite into one conversational experience.

Meet Firefly AI Assistant, now live in the Adobe Firefly app, the all-in-one creative AI studio. Powered by Adobe's creative agent, Firefly AI Assistant, lets you start with your vision, just describe what you want, and shape the outcome as it takes form with the assistant. The assistant orchestrates multi-step workflows, drawing on 60 plus pro-grade tools across Adobe Creative Cloud apps, including Photoshop, Illustrator, Premiere, Lightroom Express, and more to help bring your ideas to life.

You can also get started with creative skills, a growing library of pre-built workflows for common creative tasks, like batch editing photos, creating mood boards, and mood boards, portrait retouching, and creating social variations. Every step the assistant takes is visible so you can refine, redirect, or take over at any time. You stay in the driver's seat as the creative director. Adobe Firefly AI assistant now in public beta. See it today at firefly.adobie.com. So you can click view sources.

You can rewrite it if you want. You can export it as PDF, markdown, or create a perplexity page that then you could share with someone. or you can click share and then this will send a copy of this chat to whoever you want to share it with. All right. So let's go ahead. We'll quickly tackle the other ones. All right. So now this one was find the latest news and information about deep seek.

So it ended up going to 55 sources. All right. I'm kind of looking through here and seeing kind of where it went and in what order. So first, it was just trying to get some general information. Then this is good. It's saying, I need to search the web specifically for news and updates about deep seek from February 2025. This is great. Love seeing this from perplexity.

So then it's searching deep seek Feb 2025 to make sure it has the most up-to-date and accurate information. That's huge, especially in a developing story. Freshness is extremely important. It's talking about the Texas Attorney General. launching an investigation. A lot of information on the Texas thing here, which is not terribly important, but it's important. All right, it's going through some Wikipedia sources, a New York Times article, Economic

Times, and then it says writing the research report. All right. So let's see what we came up with here. I'm just going to go ahead and read probably the first one or I'm just going to read the first a couple of sentences here. So the rapid assent of Chinese artificial intelligence startup Deepseek has reshaped global technology markets, regulatory frameworks, and competitive dynamics in early 2025. Founded in 2023 and backed by hedge fund high flyer, the company achieved unprecedented cost efficiency through architectural innovations like mixture of experts and multi had latent

attention, enabling its models to rival industry leaders at one-tenth the training costs. While its open source deep seek R1 model became the most downloaded free app in the U.S. Apple store in January, the company faces mounting security concerns with nine governments banning its technology on official devices. This analysis examines deep seeks, deep seeks technical breakthroughs, market disruption, and the complex geopolitical tension surrounding its rise. All right. So the recap was okay. You know, I didn't see anything. I mean, I saw things I could, you know, pick, pick apart a little bit. But for the most part, it did a pretty good job. It was pretty fresh, pretty relevant.

No hallucinations. There were probably some gray area there and some just some overly general responses, but nothing that was factually inaccurate, right? There's things that were up for discussion, but, you know, nothing terrible here. All right. So then it goes through looking at the architectural innovations driving cost efficient. the market disruptions in competitive response, all right, security concerns and regulatory backlash, strategic implications for AI developments, emerging frontiers, multimodal capabilities.

All right, that's good. It's talking about like Janus Pro, kind of deep-seek's version of Dali, audio-visual. All right, so not bad, right? Again, not a ton of information, right? So if you wanted to read something and, you know, digest a couple dozen pages and really get a lot of information, at least at first, I'm not seeing this from perplexity's version of open or sorry, of deep research. So did it research deeply? Sure.

Right. Did it give you a very long and in-depth report? No. essentially it gave a bullet point of facts. The good thing, the good thing, which is normal with perplexity, is, you know, after some of these facts in the body of this response, you can hover over and look at the different sources, right? So when it says, you know, like as an example, when it said, it became the most downloaded

free app in the U.S. Apple App Store by late January. I can hover over and see that was from the BBC and We Forum. Okay, so you can go through and check some of the sources. All right. So overall, again, not bad. Not the best version I've seen, but fast, free. So, hey, if those are two things you care about. If you care about speed, if you care about price, perplexity, not hating it, not hating it.

All right. Let's try the last one. And this is one of those ones where it's like, I can't 100% very quickly look at accuracy. When I tell it to be creepy on myself, I can't. All right. So this one I said, tell me everything about Jordan Wilson, who does everyday AI from birth until today, make it creepy in depth. All these prompts, FYI, I ran them multiple times. All right.

I ran them before because I also want to see how or if the product is getting better, right? Because the perplexity CEO, there is some terrible responses that came out immediately. they said they were going to update it. So I'm actually hoping that this version did a little better because it's easier to fact check something on yourself. And I encourage you to do that, right? If you want to see if any of these tools are at least worth considering for your company, ask it to do something incredibly in depth, almost at a creepy level on either yourself,

your company, your department, if you work at a big enterprise, you know, your competitors, whatever it may be, first tested on something that. you know like the back of your hand. In C.F, number one, is there going to be anything incorrect? Number two, is there anything maybe you didn't know or forgot about, right? At least for me, I don't think it's going to be anything I didn't know, but might be something I forgot about. All right.

So let's quickly check the accuracy here. All right. So it says, Jordan Wilson, architect of accessible AI and pioneer of everyday generative intelligence. All right. Okay. it's talking about a Midwestern youth. All right, sure. That's correct. Early formation and academic foundation. So it says childhood precursors to technical orientation or technological orientation.

Mainly I'm just looking. It got in something. I used to A-B-Test my lemonade stand as a kid. All right. It looks like I talked about that on someone's podcast I was on. All right. So that's correct. It says, all right, it says I went to S-I-U. That's correct. Looks like it pulled that in from Coursera. I did a course there a couple of years ago. Looks like it looked at my LinkedIn profile as well. All right. So from newsrooms to boardrooms. Let's see if I got this correct. All right. So it's talking about my background in journalism in the early 2000s. That's correct. Strategic pivot to digital enablement. All right. So media production to digital strategy through leadership roles, triple threat mentoring, working with Nike and Jordan brand. Pretty good.

good. Genesis of Accelerant Agency. So yeah, one of my two companies. So it says in 2015, I founded Accelerant Agency. That's not correct. It wasn't 2015. It was 2017. But not terribly, right? One small detail, but incorrect there. Let's see. Everything else on here looks pretty good. Looks pretty good. All right. It's talking about a little bit of the PPP. courses. It's talking about everyday AI revolution launching everyday AI in April 2023. Yep, that's pretty good. All right. It got the time of our live stream, 7.30 a.m. Central Standard Time. Pretty good. It's bringing in some information from some of our more recent deep research. You know, just kind of meta, right? It just brought in a open AI's deep research in the deep research from

perplexity. All right. So I'm looking through here. Nothing on this instance. That's really wrong so far. Okay, here we go. This doesn't look client engagement models. This doesn't look like anything I've done. So it says, Wilson's consulting arm with everyday AI employs a phased implementation strategy. AI readiness assessment, ethical framework development, pilot program design, and scale optimization. All right. So we have our first set of hallucinations. So that made it up. And also, interestingly enough, there's no citations in that instance.

So it made up a client engagement model that we do at everyday AI, which we don't do that. It says notable client's successes include a Midwest manufacturing firm achieving a 40% reduction. Yep, that's not true. We'll see where it sourced that from. So it looks like it brought in some maybe a random transcript from our website talking about, yeah. So I'm looking at the website, just hallucinated that and it tried to source it inside it, which was incorrect. All right, here we go.

More, more hallucinations. central to Wilson's consulting philosophy is what he terms the human amplification principle. All right. That's wrong. So it was some other website that's pulling in information from other parts of the page, it looks like. So making up some stuff. So then it says media empire and thought leadership, the everyday AI podcast ecosystem. All right.

So those are all correct. All right. So it says our daily newsletter achieves 63% open rates through a structured content framework. We have very high open rates. It's not 63%. So yeah, toward the end, we started to get a lot of hallucinations, right? But I will say this, at least these hallucinations were on brand, right? It says, it says I have collaborated to some of these research partnerships and white papers.

which I did not do. It says I have 37 peer reviewed papers. Again, so it's pulling things from, you know, the Chicago AI week. So let's even go look at that. So a lot of these things that were hallucinated was from this page. I did a panel at Chicago AI week. And none of this information is even on there. There's like no information on this. There's a little very short bio in the session that I led and that's it. Right. So it's interesting that perplexity is making these hallucinations and citing them to certain pages that have like no information on them anyways, right? So yeah, I will say the bottom half of this. I'd have to look.

It looks at least about 40% hallucinated, not good, right? So the good thing is, is I can go up here. I can look at these sources, right? So it looks like it went to 23 sources. So I could see like, oh, if it's pulling in information that seems to be. be hallucinating, I could click on the source and I could click, you know, remove source. And then it would rerun that whole, that whole kind of deep research, right? So I'm, you know what?

I'm going to do this. I wasn't going to do this at first, but I think it's probably important to talk about. So let me do, I ran this exact same deep research query. Like I said, I ran all these twice. So here's the first one, the first one that I ran. I think it was over the weekend. So I said, same thing. Tell me everything about Jordan Wilson,

who does everyday AI from birth until today, make it creepy in depth. You'll see in this instance, it went to 243 sources, right? But if I click the source list here, I saw this and I'm like, automatically, I don't know what the heck is going on here. Right?

It's doing all these things from Reddit. I looked at about half of these. None of these are about, myself, they're not about everyday AI. A lot of these sources are. But I noticed that it pulled in a bunch of sources that were extremely irrelevant, right? At least on this second variation, where it looks like it may be hallucinated about 30 to 40 percent of the time, at least it brought in sources that looked somewhat relevant, whereas the first time I ran this exact same query, I mean, it just brought in. It mainly read it. I don't know why. It was,

was bringing in dozens of irrelevant Reddit threads. And then in the actual deep research report, I would say it was about 60% hallucinations. So it did get some things correct, right? It got the, you know, my undergrad and grad school didn't get the years correct, but it got the schools correct. It got, for the most part, some of my original place. of employment, correct? But it started hallucinating very, very early on, making a lot of things up. You know, in the first variation, it was bad. Very, very bad, right? So imagine if you don't know,

right, this is why it's important. Imagine if you don't know about the topic that you are deep researching. I haven't seen this level of hallucinations. I mean, we'll see. We're going to do the exact same thing. But I haven't seen this same level of hallucination. with Google Gemini's deep research. I haven't seen it with OpenAI's deep research. But again, on something that I know a lot about myself, everyday AI, I was a little concerned with what happened with perplexity's version of deep research. So is it free?

Yes. Is it fast? Yes. But at what cost, right? If you're trying to learn something, if you're trying to research a potential client, right for a huge sales meeting. If you're trying to put together some information for a big presentation for your board, you really kind of know your stuff, right?

I would not at this point feel safe and copying, not just copying pacing, but taking the ideas from perplexity deep research, right? So the way, let me say this, I'm not going to be using it. I'm not, right? There's a certain level of truthfulness. have to get. I understand. Humans hallucinate, right? There's inaccurate information on the internet. The internet hallucinates, right? But when you're using a reasoning model and you're going to these dozens or hundreds of websites, right? This should have known that all these random Reddit threads,

right, where it's like, I'm just reading some of the names. Please help. I am a new dad. It's been nine weeks. That's nothing to do with me. I think it's kind of lame when people criticize or talk trash, right? That's not about me, right? These Reddit threads. Why do some software engineers say leak code isn't worth it? Suggest me a book that speaks to what the next few years. You know, Wilson released a statement. That's an NFL thing about, you know, the old quarterback for the Seattle Seahawks, right?

So the fact that perplexities, the first version of this deep research brought in just any model, any person, anything with a brain should know if it can't match these sources up with me. My name is Jordan Wilson. I run everyday AI. You got my background correct. So why are you bringing in dozens or more than a hundred of sources that are irrelevant and then seemingly pulling from that, right? Not good.

So could this be okay-ish maybe, right? But I would be very, very careful. Don't just jump on and use perplexity deep research because it's free, because it's fast. you really have to know your stuff and you really have to pay attention. And I think the role of the human in the loop, if you are using perplexity deep research, the onus is on you, right? So I don't know from a time savings perspective if I would ever use this, at least as it is now. Hopefully it improves.

It looks like it has improved in the last 24, 48 hours since I first ran this same search. But still, even in the version I just ran out, a lot of hallucinations, a lot of concerns about accuracy and truthfulness. All right. So I hope this was helpful. Like I said, we're going to do the exact same thing tomorrow. We're going to take a look. We're probably going to do those same three prompts. It's going to be a faster one, but we're going to look at Google deep research compared to open AI deep research and perplexity deep research. And we're going to see which one is best. Is it better to use the free version and maybe just have a little bit more control, right? And only use it for something that you are innately aware of. All right. So

we're going to be doing that tomorrow. So I hope this was helpful. If so, please share this with your network. Yeah, you can use everyday AI as your personal cheat code, but I'd really appreciate it. If you'd share it, help other people out. Learning AI is scary. It is time consuming. You really have to know a lot.

That's why I do this for you, right? And I do this live. This isn't some polish and I spend 30 hours editing to get it just right and to show a certain message. This is all live, y'all. This is live, unedited. for the most part, unscripted, showing you the real part of AI, which, as you saw there, there's promise and there's peril.

So you really have to be wise and intentional about how and if you even use some of these tools. All right. So thank you for tuning in. Join us tomorrow. We're going to do the big breakdown, all three of them. Please also go to Your EverydayAI.com. Sign up for the free daily newsletter. Thank you for tuning in.

We'll see you back tomorrow. and every day for more, everyday AI. Thanks y'all. Meet Firefly AI Assistant. Now live in Adobe Firefly, the Allman One Creative AI Studio. Just describe what you want to create in your own words and the assistant handles the rest, orchestrating multi-step workflows across Adobe Creative Cloud apps, including Photoshop, Premier Express, and more in one conversational interface.

You direct the outcome while the assistant accelerates execution. Stand control with the ability to step in and refine at any time. See it today at firefly.adobie.com. And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit Your EverydayAI.com

and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.