imagine having a tool that can dig through the internet and synthesize the best research on any Topic in seconds sounds like magic right but what if I told you that perplexity has just released deep search and unlike Google's version it's absolutely free a few weeks ago I covered Google's deep research feature and while it's powerful it's Lo behind a pay just like open ai's deep research and just yesterday newly released grock 3 now perplexity is bringing a gamechanging alter ative to the table offering five deep research queries per day so today we'll run some serious research prompts see how well it soures information compared to Google's deep research and most importantly find out if it actually delivers on the promise of better deeper insights I've run over 100 research prompts to prepare for this video so stay tuned till the end where I share all the main findings can this AI researcher replace hours of manual searching let's dive in and find out so let me start with introducing perplexity to you so complexity was released I think middle of 2023 they came out as actually the best alternative to Google searches just because whenever you search in Google you will have a lot of things to scroll down through because there will be sponsored ads there will be some in relevant searches and you basically need to click through the different links to find the right answer which is annoying and takes a lot of time in perplexity came with a novel approach that they would use a large language model to actually analyze all the top results and then we were looking at something like you know first for example like five links uh without any actual ads and they will summarize and extract the key information from those services and will provide you just with the information you need so instead of you going and scrolling through the ads and clicking through the links you will get the answer straight away so this is what the perplexity was initially about and now they introduced deep research feature and the main difference between the standard kind of llm search feature is that deep research actually goes much much deeper it's not only kind of scraping the surface like first five or 10 lengths this is actually going much deeper and analyzing more like tens or even hundreds of different web sources over time yes that takes a little bit longer and the average response is about from 2 to 5 minutes but it's actually analyz this topic much much broadly and if I think myself analyzing like hundreds of services this will take me actually days to do so and here with perplexity deep search I can literally do this on many topics in just minutes right so here is the blog post that you can read and here they describe what this feature is about and they even compare the feature to the regular answer and we're going to test it ourself but in order to go there and access the feature we need to go to perplexity website and this is very simple the URL address is perplexity doai and here you will have a regular chat window where you can ask anything and in order to use deep research you will need to sign in with your Google account and once you're in you will have the same chat box where you can enter information you want to search for and here you can select different models and by default you will have Auto model but you can use deep research for your analysis here you can also select what sources you want the model to search from is it web or academic or social it's up to you and we're going to discuss this along the way and by the way we will have more like seven different prompts that we're going to test real quick and we're going to launch them now and see how it performs because it takes time to do the research I will be showing some of the main features or quirks of the model within you know different prompts that we are testing because I've gone through this process like many times already and I have gathered some main insights by empirically like testing this tool myself so let's start with very basic prompt research information about myself just to see if it can gather relevant information and we can test the model basically if it can find the relevant information about me so that's going to be the first promt as we can see for example here what it does it has this Chain of Thought thinking and we're going to get back to the prompting so you can actually go yourself and see what it does what what it searches for so it searches for my name and here it got some information then it searches for maxim SI biography ltia it found some information the LinkedIn etc etc so it goes and searches actually 55 Services already so the next uh prompt that we're going to test is research the most popular axle influencers around the world include all social media that is available Tik Tok Facebook Twitter LinkedIn YouTube Instagram and other social networks the size is defined by cumulative following across all social media platforms so here in particular I want to test if the model can actually go to a social network like Tik Tok or Facebook or LinkedIn and actually gather the number of the follower count and then summarize it in a table format so I don't need to do that so the reason I'm asking about Excel influencers just because first of all I've worked in the industry and I'm familiar with some faces and names and secondly we have tested this prompt on actually the Gemini deep research and here we have some results and we going to compare the results later what is really more important here is that you can select different search types here you can search from VAP you can search from academics papers or from Socia so first we're going to be testing searching on web second we're going to be testing on searching on social and then we're going to compare the results because the results will be significantly different and you will see that and that's why I want to show what are the biases actually if you use different strategies so that's going to be the next one and the third that we're going to use we're going to select both the VAP and the social and the prompt is exactly the same and then we're going to run it like three times the same prompt with different settings and see and compare the results what would be the best outcome okay so let's actually set this up so the next I want to test how a model is actually good at searching academic articles on a particular topic recently I found out a video about this guy leam Motley who is actually saying why you're better at prompt engineering and how to fix it and he tries to provide empirical evidence from recent academic studies of what kind of prompt engineering techniques you need to use to build efficient AI systems and here he goes with ro prompting Chain of Thought emotion prompt Etc and he provides scientific backed research paper for each of those arguments which is really ni and imagine now you can create such videos and do the research instead of you going and searching for those papers you can ask AI to do that so what I wanted to do is actually I wanted AI to find recent studies about you know prompt engineering techniques that I might use and maybe create a video on prompt Engineering in the future and here we have a little bit different structure for my prompt so again what I wanted to emphasize here that prompt engineering techniques are still relevant and they still needed to be here here so for example here I would use HTML formatting for my prompt to specify to the model specifically what is the role of the model what is its task what are the specifics that I'm looking at and how I'm expecting my output to be and I'm saying exactly what I'm expecting to to find so I'm asking it to research the recent academic studies on different prompt engineering techniques and I'm looking for the studies from 2023 to 2025 also I'm looking to prioritize our Hive PR pre prein just because this is the place where the most studies on AI are actually published and we're looking from 2023 so just Recent research okay the next one I wanted to use to actually show that the model has cap on how many warts or actually characters it outputs per message and if we're giving to Broad of a prompt like for example here I will be asking to design an optimal 30-day erary for a solar traveler The Prompt itself asking for for a lot of things to be done and that's why I will receive sort of very high level responses not very deep not very detailed just because it has a cap on messages and I'm going to discuss it a little bit later so let's actually see what we got for the first one okay research information about myself Maxim s comprehensive analysis professional contributions Etc so emerged as a pival figure in intersection of corporate finance and artificial intelligence okay see Flow by thei and here we can actually test where this actually comes from so there is the squins website where I have one publication I guess and there is like win parters that's my uh business partner that I work with it actually says finished the mini MBA Stockholm scho of economics and I've actually finished receiv received my diploma just yesterday so congrats to me okay in 2024 transition to decisively toi that's totally correct there is a YouTube video I actually U take have taken this from my YouTube video that's that's really nice so it actually can watch the Youtube videos well it actually doesn't watch the videos it probably takes the transcripts of the videos but still pretty cool that it actually found the information on on the video I think I mean it's really straight to the point I would say you know educational Outreach a automation strategies second best assistance I think generally it's actually precise was so far so s research I driven has published in journal such as Journal of economics and management research this is although this is true the research that is published here let me check okay so there is like another Maxim so I'm Maxim sich and there is like random guy Maxim ianov that has published This research which is not me I have actually published uh Research In This Journal but on the different economic uh paper so it actually got this off so you definitely have to check the sources because as you can see here it's just just a different person right and for example here collaborations with institutions like Institute of electronics and computer has never done that generally it's very good it's very to the point describe my company my previous experience what I do what the company does all that stuff was there different sources how many sources it searched let's actually see 55 sources so again you can see how it thought Chain of Thought here right you can see if what terms it was searching for what it got you know analyze different SES summarize them Etc here you can go through different SES and see if this SCE looks like it shouldn't be there you can actually go for example this was there was there Maxim COV if I can find it right now um let me see okay so here we go so so this source for example we don't we know that this source is not relevant so we can select this source and say remove source and if we do so what happens next is it's just going to rerun the whole research for me well on the one hand you might think okay so this might be a little bit better output because I can sort of manage but then again it just go again and do the whole research for you and I don't really find it very very valuable of course you can that way kind of filter out some information but it actually takes a lot of time move on to the next one so we needed to research the most popular axle influencers actually searched for 161 search and it still working where dat says he says 2. 7 million you so the problem is that sometimes this actually gets stuck it it's been more than 5 minutes already and I think it got stuck let's wait um a couple minutes if it doesn't move from there we'll need to rerun it again okay so this one was the Deep research it searched for 137 sources so the first one was using web searches across the entire internet and the second was just using social and here comes the big difference when we have socials if you look at the references right here that's basically social for deep research is actually meaning Reddit I would consider socials being like Facebook LinkedIn Tik Tok and many other social media networks here in fact if you press social it means you're actually searching on Reddit and you have to take this into account well the problem with the social though is that sometimes some reddits have some you know jokes and information I mean it's un regulated thing if you're looking to search for different Trends or discussions on specific topic that might be a golden information but if you're looking for the most popular axle influencers that's not always the case but you are looking for exact and precise numbers that are actually checkable on the internet you're not looking for an opinion right and that's why here for example if you're looking for social you they're saying okay there is like this Miss Excel dominance the fast growing Exel influencer there is this Mike giren Leila gani if we were to see what's the result on the VAP that's going to be a totally different result let's actually check because the last query we were asking to search it on the web and on social so there are 120 SES there are some reddits but there are also like decent web pages that was gone through and that's probably would be the the best analysis so far so this one definitely struggled so let's actually rerun this again so when I refresh the page well the problem is when I refresh the page it actually drops the search deep research query for me so you need to rerun it again and that's kind of a bad bad sign so instead here we go let's do um sorry this was for the web search only so let's rerun and while we are rerunning let's actually check what we got so um this guy is actually saying that has web search plus social that there are cumulative followers engagement rates blah blah blah so we've got the Chi-Chi being on the first place and this is actually true they have 28. 7 million followers action dictionary being very very popular that's probably t T Tok I guess or uh this might be Tik Tok plus Instagram I'm not sure so there is like a comparative growth viral educational content so yeah we've got the actual dictionary 1.
8 million and 4. 3 it might be on Instagram I'm actually not sure we need to check that okay so there is like Miss Excel Excel dictionary uh Grant uh then there is like LinkedIn influencer cat Norton 620,000 followers s chort that's actually Excel dictionary lady uh so uh that's this lady yakob Carrom 310,000 Chichi 8. 7 on Instagram XL dictionary again so what we actually see here it provided four different SES so Instagram Tic Tac LinkedIn YouTube which is really nice I'm like I imagine myself going and looking for this this is so annoying and now for example it provided all of these and calculated the cumulative number of followers really nice really fast so if I'm looking to work with like the biggest followers or just generally some of the followers in the industry um this makes it so much easier even it has like Revenue diversification uh descriptions well okay so that's not what I asked for but but okay emerging Trends platform saturation ethical concerns fine it actually provided just with the three it's very very little and the problem with this one is just because it has so little context output window and that's the big the big problem actually right now because it provided a lot of of other information that I didn't need to I just needed it to actually go and find like the the top people and I don't need anything else I don't need ethical concerns and stuff like that so it might be the case that you would need to prompt it a little bit more specifically but the bigger problem is that it has just limited context window and here is just three people that's that's the problem I wanted also to quickly compare it to Google deep research and Google deep research Who provided me with like uh again Kate noran Leila garani John M kodas some other people actually like Paul burher is not an Excel guy but it's still there I mean the others are and it tried to kind of gather some statistics didn't find the Facebook or Twitter handles it did find though Leila garani which is one of the biggest actually followers on Excel and interestingly like uh for example perplexity search didn't which is again went weird right it provided with yakob Carrom for his 310,000 followers it didn't provide with was Lea garani which has 2.
7 million on YouTube and 1.
