I paid for a $200 yearly subscription to Tess AI, so you don't have to. In this video, I'm going to go over my experience so far using it and why I wouldn't really recommend it right now. So if you're unfamiliar with Tess AI, it's another multi-agent or, excuse me, multi-LLM platform similar to you.
com or perplexity and what these tools do from the ground up is they allow you to use multiple models multiple llms so you're probably familiar with chat gpt but there's several other companies like meta or facebook who release models clod and tropic you probably heard of deep seek what these platforms allow you to do is use all of those models in the same platform and usually for the same monthly price. So here you can see we have our one from DeepSeek, Sonar, which is Perplexity's custom model and a reasoning model as well. And this is all in the free Perplexity Labs playground.
For Tess, we have DeepSeek, O3 Mini, and Gemini, which each of these DeepSeek is free, but OpenAI costs 20 euro per month and Gemini costs 20 euro per month. So if platform you are kind of getting these for the same cost so their their landing page looks very you know slick i don't really like the ai imagery on the landing page but it does describe what it does right um where it integrates it has a knowledge base all this stuff it has an ai studio and it looks like it's more designed for teams but in my experience um i'll kind of go over how it wasn't super useful for me. So this is the plan I subscribe to, the individual plan.
And I believe the company is based out of Brazil, which links to the first reason I wouldn't recommend buying this product is because many of the support pages I landed on and some kind of weird quirks in the interface were in Portuguese and I don't speak Portuguese. And so I'm not quite sure that this is ready and has been tested fully in international markets. I don't mind that, of course, but if I am paying the same price as I would pay for u.
com or Perplexity, I do kind of expect a certain level of polish around the product. The other major thing is this is an annual credit system versus on u. com and perplexity.
I've never run out of credits and I've never hit a limit and I use these products for eight plus hours most days. But I do like the pricing table. It is a little bit complex but I like a little bit how you know how many features they list here.
It's part of what inspired me to buy the product. Let's go to the actual product and kind of go over my feedback of it. So the first thing, I left this window up for a reason because it really bothered me during onboarding to even get into the product.
After you pay, you have to fill out three screens of full-page pop-ups with company information, a domain, everything like that. So unless you're an established company, it doesn't feel like this product is meant for you. So if you don't have a website, If you're not using this for a company, it feels like you can't even go through onboarding unless you have that.
So I put in some random website, but it just kind of sends the message that it's not for me. The other thing, since I've been using the product over the last week or so, I keep getting random pop-ups like this, which is, it's great to hear about new features, but for me, it's quite intrusive. I kind of have a double pop-up here, and if I count the chat as another indicator, I kind of have three things going on.
So we'll close out of these here. So the cool thing here is the interface is pretty easy to understand, I would say. You can see I really haven't played with this too much, just because I'll go over why actually in a minute here, but it's because it's a little bit, the AIs fell a little bit flat for me.
So I have the individual plan here. you can see that this is from less than two hours of actual use, I would say. So that concerns me a lot.
I don't think 36,000 credits would last me more than a month max. So in the onboarding documents or kind of the website copy, I think I was led to the impression that Tess AI is their own in-house developed AI, kind of like sonar is for perplexity. But it turns out it's like an AI picker.
So I guess when you select these models, it helps determine who to send it off to among all of these. And I think this is probably okay for someone who's really new to AI, but I think anyone who's been using any LLM systems for more than a week, you're going to probably prefer specific models for specific tasks. So I'll just go through kind of how to use the interface here, and I have a couple of prompts prepared.
So here is a kind of good thinking and research prompt that I like to use. So it's make a list of all country independence days, who they gained independence from, and their current GDP. And we'll just use the Test 5 Pro, and we'll say use the internet.
A quick note here, just again, in case you're considering buying it, I have tried the Instagram and Facebook scraping. That was one of the main reasons that I purchased this product, and it wasn't quite able to do what I needed. So I'll give an example in a minute here, but just so you're aware.
All right, and then we'll do a systemic mode. Again, I don't really like when so many things are being passed to my prompt, but we'll get this running. So from my understanding, test AI is going to say, what tool should I use and which model should I use?
And then it's going to pass it off to that. For reference, perplexity deep research and open AI deep research is able to one-shot this question pretty easily in the last few months. So I'm going to let this keep running.
On the side here, you can see we have image generation as well, which is pretty standard in most models. So we'll just use their default prompt here. Oh, okay.
Maybe. And the other thing I don't, I feel like was maybe hidden from me is that image generation costs differently depending on the model. So you can see here it's like 44 coins or credits.
And the reason that matters is because once we get over to video, a lot of the video models cost so much that I would run out really quickly. So you see here, just for even using their in-house model to choose, it's $140. But when we start using some of the others, you know, $210, the cost goes up quite a bit, especially compared to using something like fall.
ai. So for me, if I, you know, had pitched to my boss, hey, we should buy this tool. We can generate social media videos and everything else like that.
And then we can generate, how many is that? 36,000 goes into 400. So we can generate about 90 videos using one of the decent video generation models.
And of course, not all of those videos are going to be correct. So we're going to have to do multiple videos for each one until we get models where the fingers aren't falling off or kind of typical other video model mistakes. So, you know, we're talking about 90 videos for social media and then we're out of credits and I have to ask a finance department for more.
So it's a little bit important if you plan to use the image or video generation. Here, I don't really know that code should be its own model. I played with this a little bit.
What I was looking for is something like artifacts in Claude where it actually shows or canvases in GPT where it actually shows the development in another panel, but I haven't been able to get that working and playing around with it. Where did we just go? Okay.
And then VoiceOver is fairly standard. It uses 11 labs models as well as open AI models, which is like, you know, state of the art, I would say best in class. The other thing, there's a couple new voice models that have gotten released this month.
and last month it's March 2025 and I don't know how quickly they add models to their interface but this is something that I've really appreciated with Perplexity over the last year is almost as soon as a company releases an open source or even a closed source model they get it into the interface for ChatGPT 4. 5 for example I think Perplexity had it in the interface within I want to say under a week after it came out which I really appreciate and here you know I'm checking my internet speed here and my internet is stable and you can tell I clicked transcription and I'm just having some loading issues and this this has kind of continued to happen I don't know why it it seems to kind of get stuck in a couple of different places and again I'm not I'm not trying to come at the team I think they built an awesome product but there's just a lot of a lot of little things that you kind of see me struggling with that for me makes perplexity or you. com even maybe stand out a little bit more.
So let me grab this here. Okay. So actually this is kind of an example.
I sent the chat here and then I wanted to show the rest of the interface and I guess it's not going to work on it in the background. So I clicked away and, okay, maybe it was working out in the background. I can't really tell.
There's no UI indicator that I saw. Okay. So this got a whole bunch of information here.
And what I'm looking for is, yeah. Okay, so this is part of the challenge for me of using this. Because every company wants to reduce their cost of compute to the least amount possible.
and this is a great list but I don't really have sources I don't know what it went and searched for but the most important part for me is that I use test to choose my model which is kind of a claim to fame on this platform and it uses the cheapest model like I can run this on my my current computer right now for free and to use this on the web is usually free as well you can even find deep seek models to use completely for free via API on things like open router or you can use llama a more powerful llama version for free via together AI so they could have actually passed this to an even more powerful model and I get that maybe they determined okay if it didn't need that level of compute but I guess for me as the end user it it kind of it doesn't give me a great feeling that they're using just you know any the cheapest possible model like and even it's even llama 3 not llama 3. 3 which is the most recent so for me that doesn't give me the uh the best feeling so let's go ahead and just use kind of a fictional example i'm gonna start a new chat i guess i can go over here too and then i'll use test 5 again and this is more of a give me the steel man or straw man case along with a swat for changing my company name from bulldog coffee to Ox Cafe. So this should be more of kind of like, I would personally pass this to something like DeepSeek R1 for a thinking model and have it like really think it over.
But I think this just passed it to a random model. Let's see here. Yeah.
So Gemini 2 Flash, again, a very, very, very cheap model. This is even cheaper than I think GPT-4-0 if I recall. So let's acknowledge I mean, the answer is good, right?
But it kind of defines things that don't need to be defined, which I would say is a trait of 2-flash. And I can actually demonstrate the difference here. Let's go ahead and do the same thing, but I'm going to choose the model this time.
And if I show on O3 mini high, it's going to probably give a more comprehensive answer, and it's not going to be as verbose. One of the big mistakes that I see, or I guess I think it's a mistake, maybe it's not, is that more text is better thinking or more verbosity is better thinking. But what I found is the shorter the answer, the more succinct or the smarter the model is.
So this is, you know, OpenAI deep or O3 high. So a pretty smart one. And you see here, it's not defining the steel man case in like, it's all it's saying is best case scenario to give people a reminder.
And then the straw man case. And it's making so many different assumptions. I would say it's about the same length, but it's a little bit more snappy.
I don't know if you can see a major difference here, but I would say even the structure, I appreciate more. Again, more text for me means usually a worse reply. The smarter the model, the more succinct it can be, much like people.
Awesome. So we kind of went through a couple of different models. We talked about TESS 5 is, I guess, a model selection model that, hey, maybe it works great.
Maybe I'm just really bad at choosing models, but I haven't particularly been impressed. We talked about image generation, video generation, and the cost. Again, using some of the more expensive video models, you're only going to get 90 videos out of this, which I do feel like compared to the pricing page here, that's not really signifying on here that it feels like I missed that point.
Now, one of the most important parts, and once again, the reason that I started using it, is because supposedly it can get into the walled garden of Facebook and Instagram. So what that means is my pitch to my company, if I wanted to buy this tool would be, Hey, um, all right, I'm going to use this again. Hey, we need something for our social media managers.
And so right now, if you want to see the trending towns on Instagram, it is a little bit challenging to find, which I think is a great LLM use case. So I'm going to say, what are the trending sounds or clips in the last week on Instagram Reels. So here I am expecting it to find sources.
I am expecting it to give me specific links just because I've been trained to expect that with OpenAI's deep research or perplexity or even u. com occasionally. So I can see it's using Instagram search.
And again, you know, this going to blink when I mouse over it is one of those things that I kind of question a little bit of the UI occasionally. While this is searching, I do want to talk about this. I have gotten, I think, one or two prompts to start a course on some sort of AI thing, which again leads me to believe they're targeting folks who are very, very, very new to what AI is, I guess, and what it can do.
So so this might not be the best product Okay All right. Yes, I want info on all of them So here I search for one of the trending sounds or clips in the last week on Instagram reels It replied and selected again a very cheap model of Gemini to flash. It supposedly went and scraped Instagram It says it used the tool Instagram search and it came back with the answer I've searched for trending sounds on reels, which this is the reason why it should not be using, in my opinion, Gemini 2 Flash.
It should be using a smarter model to understand the implied intent of my search. So when a user searches, what are the trending sounds? What I wanted to say is I wanted to think like an intern or an employee and say, oh, you know, my boss just asked me for what the trending sounds are.
I'm going to go find trending sounds. I'm not going to search for trending sounds because that's not going to really get any results. I'm going to say, okay, let me see this creator and this creator.
What sounds are they using? All right, so this is what it was doing last time as well. So I'm happy this was, I guess, replicated on video.
And I'm sorry I keep complaining about this, but again, we're getting a model that costs pennies and less to run. GPT-40 mini here. And yeah, I'm using it for search, so I guess we don't need expensive models here.
But this trend of using the cheapest model does bother me a lot because, you know, if you're charging 20 euro a month for something, you need to compete with U. com or Perplexity in my opinion. So find and use trending audio on Instagram.
Yeah, so this is just summaries of posts and it doesn't give me any. So this is probably the most useful one. Maybe there's like an account that gives away trending stuff.
But I would say yes. So maybe two of eight of these are useful so I can see the top reels. But I have to click to another platform.
And this to me, it doesn't feel like it actually went and scraped inside of the walled garden of Instagram. Which does concern me because the whole reason I would buy this tool is because no other AI tool I've used has been able to see inside of Instagram and pull information from it or see inside of Facebook or LinkedIn and pull information from it So that is more or less the summary I'm going to run one last prompt here And Just to just to kind of see which model it picks we'll use test 5 Pro Just in case that performs a little bit differently, but we're asking for a blockchain solution for seamless rebooking during multi airline cancellations Given the speed this is writing. I can tell it's using a cheaper model so Maybe it's a really easy question, but I would really rather Send it to a smarter model.
Oh wow, okay, Claude So not I'm I am pleasantly surprised because this is a model I use almost every day I use the extended version, which I don't know if they can grab through the API. Let's check. Yeah.
So they're not using the thinking version here, but they're using a state-of-the-art recently released model. So that's great to see. Given that, let's go ahead and try another prompt here.
And we'll pass it to the pro model again. And let's use LinkedIn search. so here I'm hoping it scrapes LinkedIn oh alright okay so this seems a little bit similar to OpenAI deep research where it says what it's going to do and then it goes back and scrapes the website I did want it to only scrape LinkedIn in so maybe they're supposed to be that maybe there should be a check box beside this this output is really not oh okay yeah so so here I asked it a question of critique the welfare implications of substituting GDP with biometric well-being indices using a contrast and Pareto mental model so it did some analytical thinking it did use a web scrape tool and it said let me gather some recent scholarly perspectives which is like that's something i would love if every model does right grab some recent studies but it doesn't i'm not really sure how it used them because like reading the rest of this it's i would say this is like very basic non-internet connected model thoughts um Yeah, it's not really giving me anything I'm surprised by, and most importantly, I asked for it to use LinkedIn, and it hasn't used LinkedIn for this response at all, from what I can tell.
So, all in all, I just wanted to show kind of how a few different prompts work in different models. Let's go ahead and I'm just going to drop this to end the video. I'm just going to drop this into Sonar Reasoning Pro, which is a perplexity-developed model, And we can just kind of see the difference.
This is, again, labs. perplexity. ai.
It's free and it's very experimental, but I feel like this will probably outperform what we saw in the model selection there. So right away, and what Sonar is, by the way, sorry, oh, while this is writing. Sonar, from my understanding, is the web search AI or web search kind of LLM or tool that's custom developed by Perplexity.
I think they trained it in-house. And then the reasoning, I think, is a R1 or DeepSeek model stacked on top of that or any other type of reasoning model. And you can see all their reasoning up here.
On their actual go-to-market product, this is all hidden. But it's really nice to be able to see that. And that's something I felt like I was missing out on a little bit in tests.
I wanted to see how it reached these points. And the critical part here is, I'm sorry I keep mentioning it, but it uses an ultra-cheap older model, the Claude Haiku, which is, you know, that's the equivalent of Gemini 2 Flash, that's the equivalent of GPT-40 Mini. It's just like a very cheap model, and it just makes me question if I'm being asked to cut costs while also having a credit system and also paying 20 euro a month.
Those three things combined. If this was an unlimited tool, I would be so much kinder to it, but given that I will run out of credits probably within a month at my rate of usage, that's why I'm a little bit harsher on it. Okay, so back to perplexity here.
We can see already something important to me is that the text format is completely different. We have multi-level bullet points, we have bolding, and we have sources. This is part of why I continue coming back.
We can see we have analysis from Brookings Institute, and it looks like there was another BDE, IMF. Okay. So we have analyses from some pretty specific places, and then we can see biometric, GDP, contrast, Pareto efficiency implications, synthesis and challenges.
And then it kind of gives us like, here's how we could move forward. we should move forward with the dashboard approach, but it's actually quite easy. It's pulling quotes from specific papers here.
So it's saying, you know, GDP captures, quote, the main being of well-being or the main component of well-being. So you can kind of tell, even with a free model here, blowing this out of the water, I hate to say it. We can, let's try this one more time and then this video will be over.
Let's try this one more time with, Let's see what the best model they have right now is. Yeah, so I think 03 Mini High is the best model they have. And actually, this is kind of interesting because on the top here, I see, okay, so DeepSeek R1.
That came out two months ago. 03 Mini came out at least a month ago. And Gemini Pro, I guess it depends on which version they're using, but that came out a little while ago as well.
Because I think they have Thinking Pro out as of recording this video. Okay, so let's try a 4. 5 preview and see how it does.
I'm leaving LinkedIn scrape on just because I'm curious. Yeah, and it appears to just hop into it without really thinking about what to search for. It doesn't appear to use, I left LinkedIn scrape on, it doesn't appear to want to use that at all.
Even though I find, I feel like that would be useful to go and grab some people who've worked at the IMF or Brookings Institute and find some of their LinkedIn posts about these topics. But I will say, you know, better formatting here. It's not two levels of bullet points, it's just one level of bullet points.
So slightly better. But missing sources and missing quotes. And it seems to have only used these two models.
So it did follow the instructions a little bit better. But I'm chalking this more up to GPT 4. 5.
Last point before I close this video. If you like this type of video, kind of a long form, exploring models, especially the gated ones behind paywalls, please drop a comment with which tool you want me to test. The reason I do this is because from the landing page, without payment, you cannot tell how good a tool or platform is.
I've paid for a year of Perplexity. I have a year of U. com.
I have a year of Claude. I have paid JetChat, ChatGPT. So I'm paying probably 100 euros per month for all of these different AI tools just to test them and use them in my daily life.
And if you have one you want me to try, drop it in the comments. I'll make another video like this. The thing that I'm really frustrated by here is just the marketing feels so different than the end product.
And then I don't want to complain about support, but that is something else I would test before your purchase. In their favor, I haven't tried their AI studio features as well. Maybe there's something in here that really would take me to the next level.
But again, with the credit limitation, that's going to make me very wary to do something that repeats every day, like checking Instagram or LinkedIn for certain posts, because I'm going to run out of credits fairly quickly. So thanks so much for watching. Yeah, appreciate it.
Have a great month and week ahead.
