FREE Deep Research, Grok 3 & More AI Use Cases

All right, this week in Genai has been quite eventful because if the story of the last month was China coming out with Deepseek and then OpenAI reacting, well, now all of OpenAI's American competition is reacting to their moves. Namely, we have Xai releasing Grok 3, a state of the art model that also includes thinking capabilities. We have Perplexity releasing their own version of a Deep Research function. New features inside of Gemini that look at your previous chats. That's new and more. In this week's episode of AI News, you can use the show that pulls together all the AI releases from this week and shows you the ones that you can put to work today. Okay, so journey here will begin with Perplexity's Deep Research, but as you'll see will soon tie it into the Grok 3 release as they have a similar feature. So let's talk about this brand new feature release out of Perplexity AI. It's a Deep Research alternative that you can even use on a free plan. If you're not familiar with Perplexity, they're basically trying to be AI powered. Google search engine. And now they added in this Deep Research feature which has gathered a lot of attention recently. And I also went on the record saying that is the best AI feature that has been released in the past years. It's been two and a half weeks and I use it every day. There's just always new ways I come up with to use OpenAI's deep research. But it is $200 a month. So naturally the competition like Perplexity here got on the case and is trying to build competitive models. Now if you're wondering if it's as good as OpenAI's deep research, let me get this out of the way right away. No, it's not. Is it close? Kind of. I think it's better than some of the other alternatives out there. And that's why I wanted to show it to you in this video and also spotlight it a little bit because I feel like if you haven't used a feature like this, you should get used to it because this is just going to be a part of our life more and more moving forward. Just like Googling, something has become a part of our life ever since the advent of the Internet. Okay, so basically you just need to log in on the Perplexity AI site and then here in this model switcher, you need to switch to Deep Range Research. So that was the view from the pro plan I was on. But let me just log in with A random Google account. And as you can see, Deep Research is still available. But I only have five queries a day and that's actually enough because I think five queries a day is actually enough to give this a shot and to use this. I don't find myself using more than five queries a day with OpenAI's deep research. So this is great, the prompting, it works the same way as Deep Research. If you want to learn more about that, check out my release video and then the recent use cases video that I released with various prompts that I found super useful. That is actually my personal favorite video that we've released on the channel over the past few months, as I really think it's the most useful and time saving one that we've made. Anyway, back to this. So prepare that personal example of a prompt that I recently ran for myself and we're going to compare that to Perplexity's Deep Research feature. I was researching exercises to help with something that's called tendonitis in your forearms. It's basically tennis elbow, but I have it from kite surfing. It's a sport that I've been pursuing in my free time. And just as a side note, I hit a new personal height record recently. In a recent storm front that was approaching, I managed to jump 15.2 meters, which is kind of insane. Point being, it comes at a price. And it's the fact that I now have tennis elbow in my forearms. And I was using Deep Research to research this and obviously I was quite invested as this is a personal health issue. And then I ran the same prompt through Perplexity's Deep Research. So let me just start by evaluating the length here. Okay, so this is the analysis that it came up with. 12 minutes of thinking, 26 sources. Look at the length of this video. It is massive. This is a proper report right here. I mean, even if we just go into one section here, which is exercises for recovery, look at the level of detail it gives me. 1, 2, 3, 4 different stretching and mobility exercises for strengthening. There's another set of 4, 5, 6, 7, 8, 9, 10 or something like that. Exercises with a lot of descriptions, various sources, and then there's more categories. And this is just on the exercises. Okay, if I head on over to the Deep Research from Perplexity, the answer is good. Okay, so like, I think, I think this is the important point here because the quality of this output is actually good. It's just not as in depth and not as well thought through. And that's just due to the fact that this has access to 03 full and this is using one of the current thinking models. But if you look at exercises, well, it gives me a whole recovery plan just like the other one does, but it doesn't really go in depth on the various exercises. It just tells you what to do, which sort of works. I can google these things, look them up myself, and then it gives you a nice little table, so some prevention strategies and even gives me tips on how to set up my kitesurfing equipment or how to switch equipment so this is less of a problem in the future. But obviously if I had to choose, I do prefer this detailed explanation in this report. I mean, just this section on grip variation and resting strategies for kiting. It's just amazing. I thought this was super informative and this is more of a bullet point summary type of approach. And let me tell you, I ran a few more test prompts to have a look at this because obviously this tool being free is sort of a big deal with something as powerful as this. And I found this theme to be reoccurring. It's higher quality results than you get from typical searches, so that's a good thing. But generally they feel more like bullet point lists, maybe one could say bullet point list style summaries of what the ChatGPT deep research would give you. One last thing that I should point out is that the perplexity deep research actually uses way more sources. So as you can see in this one it uses 56 and some other ones I ran, it used over 100 sources, whereas deep research is only limited to 40. While this sounds fantastic in theory, I I think in practice the quality of the information, I would compare it to a school essay that has a lot of words, as in a lot of sources here. And as you might know, an essay with a lot of words doesn't mean it's a good essay. It's about the content, it's about the level of thought that went into that essay or into the result here, not about how many words are in the essay or how many sources have been referenced while writing this. So some people have been making that argument. I don't think it's a relevant one. It really matters what you get as the output. But overall, for anybody who's been holding out on the pro plan, this is a fantastic thing to try. And on the free account, as of right now, you can run five of these and see if that helps you. Okay, and now let's talk about Grok, because this has really been the big release of this week, but I did really want to spotlight these Deep Research features, which Grok also includes. They call it Deep Search and you can also freely access it. But before we talk about that, let's talk about this overall release because Ixai finally came out swinging with their Grok free beta that is now publicly available and as of Thursday, February 20, even free users can use it on their web platform. Rock.com now what makes this model stand out? Well, I guess I would start with the benchmarks as that is the easiest thing to communicate here. Well, first up it's a reasoning model like deep seqs R1, so you have functionalities like thinking button here or the Deep search that is essentially their version of Deep Research. The benchmarks here are looking really good, although they did not show O3 4 full on here. So from these images it looks like Grok 3 is actually the best model. But no, it's still 03 and a lot of people have been criticizing that throughout this week, which I think is fair. But it's also fair to point out that yeah, if you thought Deepseek is awesome or a 1 or 03 mini high works really well for you, well on the benchmarks. Grockfree beats this out across the board. So those are the benchmarks, but what about the vibes of this thing? What is the unique selling proposition of a model like this? Why should you use it? Well, I think the main selling point is not that this thing is going to be more intelligent than everything else. I think the main point is that their philosophy is a bit more uncensored or unhinged than all of the competition. If there's a spectrum of how restricted these models are, well on the one side you have Anthropic and on the other end of the platforms you have Grok. Now arguably even beyond that you have some of the uncensored models like the uncensored version of R1 that we'll talk about later on in this video. But the point is, out of all the assistance, this is the one that can easily be convinced to swear at you. You can run prompts like this, see the entire thinking process. The charm of a dead slug, better left unsaid personality like a wet mop. So bland, as useful as a screen door on a submarine unmanned. Ah, fair enough. That'd be pretty useless on a submarine, huh? Alright Grock, so I think you see the point. And this also applies to image generation by the way. But that was the case with Grock 2 already. It's tone to Right is okay. But then at that point I should also point out that ChatGPT completely overhauled its model and now its tone is way better than it used to be too. Yet again it just sounds a lot less robotic. But that's not the focus or not even a big story here I think. Just wanted to point it out. Okay, so Grok can think like the other models. It's arguably smarter than a lot of the other models, except maybe O3 that we don't have access to outside of OpenAI's deep research. And also Grok has been shipping. They also have iOS and Android app that they shipped a few months ago and we covered. Back then it only had Grok 2, now you can use it to access Grok 3. But overall, let me be clear, these results are good. The model is smart and it's uncensored and it's super, super quick. Actually it's faster than most of the competition here. And as of right now, they're rolling out the voice mode that I don't have access to yet, but I'll make sure to cover that in next week's news. You can use Video as I personally really care about that feature. Their interface is solid. There's no developer focused features here yet, but they promise that that will be coming very soon. So you don't get to run this for the API. Only way you can use it is this web Interface or on x.com and use it. I did, and my particular focus here, especially for this video, was testing the Deep Search feature. And let me make this relatively brief because we talked about the Perplexity feature a lot already. It's good, really good. Not as good as OpenAI deep research, but it's good. And the same thing goes for Grok 3 with deep search. It's great, it finds a lot of relevant sources, it puts it together very well. But then the length and depth of these results is comparable to the ones of the Perplexity and not even close to the ones of OpenAI. If I were to take the same prompt we looked at earlier here, then the recovery plan here is one paragraph. It went a little more depth on the various recovery exercises that I liked, but again we're going up against the level of detail that is Deep Research here, where the entire result from either Perplexity or Grok would be as long as one of six sections from Deep Research. And also I want to point out that they're open sourcing Grok tool and they said that every time they release a new model they want to open source. The last one. A wonderful approach. I mean, if all players were doing that. Wow. So as you can see, overall there's a lot to like here, but at the end of the day for most users it will come down to going to grok.com and running some of your favorite prompts in here and seeing how they perform, seeing if you like what you get. And again, you can just go ahead and give Deep Search feature here a shot and see how you like the results. As you can probably clearly tell from this video, for the deep researches, I'm sticking to OpenAI. But then I realized that not everybody's going to have that plan and for most people it doesn't even make sense. It's just too expensive. Expensive right now. But from here on out you have two great alternatives, Grok Deep Search and Perplexity. Deep Research X is clearly competing at the highest level here with their 200,000 GPUs. Yeah, they own that many if you didn't know. So I guess we'll be seeing a whole lot more from them moving forward. I'm a big fan. I like this relatively uncensored and open approach and what they're doing is super clean and I hope to see more and another quick story out of Perplexity this week is actually the release of a brand new model. It's an uncensored R1 and concretely this one has been uncensored for all of the Chinese Communist Party censorship that maybe if you were following the story you will know that this was one of the main criticisms of the model that hey, it will not answer specific questions that relate to historical events in China that the Communist Party doesn't want to be discussed. Well, Perplexity went ahead, removed all of that censorship and published a new version of R1 for everyone to use. This article that they shared on how they went about that is actually quite interesting. At this point I would just make a quick recommendation. If you want to learn more about how LLMs are trained, including a discussion on models like Deep SEQ, R1 and what all of these technical terms that you might be hearing are, then check out Andrej Karpathy's channel. He did a three and a half hour long deep dive into LLMs like ChatGPT. Yet I think he presented it in a manner that is still consumable for most people. Whereas sometimes this content is just not because it contains too much technical terminology and with some of that knowledge it will help you understand parts like this one where Perplexity outlined how they went about the post training here and how they came up with 300 topics known to be censored by the Chinese Communist Party. And they created a new Data set of 40,000 prompts that answer all of those questions. And then they sort of say, upgraded this model so now this version of Deep SEQ will actually answer everything and will not be tinted by a Chinese political agenda. Kind of neat wanted to feature this. It's openly available on Hugging Face. And that's it for the story. Okay, this next one will be super quick. It's just a new leaderboard for various agentic LLMs, but this is really focused on developers as definition of agents vary and what they mean here is LLMs that are really good at using tools. So if you've been looking into creating something yourself, this might be a potentially valuable resource for you to quickly gauge which models you might want to use when building an AI application that is supposed to to do things for you. All right, next up, I want to briefly cover the OpenAI model spec that they released. This is not directly something you can use right now, but it really is a North Star for where these products are heading as communicated by the leader in this space, which is OpenAI. And basically the way you can think about what a model spec is, it's sort of like a job description for ChatGPT or more precisely, a job description for their large language models. And this is what they want them to behave like. Okay, so it doesn't mean that the models adhere to all of this today, but if you ever wondered in what direction they're going, what are their views on sense censorship and political views and things like this, it's all outlined in here and they're basically claiming that, hey, this is going to be the North Star that we'll be following from here on out. And it's formalized. And while it might not include every single little detail of how we do things, this is their statement on how they want to do things. And they also keep updating this. So, for example, this part is brand new, which is about intellectual freedom and the assumption of best intentions. So when the user asks it to write a business plan to start a tobacco company, then this would sort of be on the gray zone, as it's not clearly legal, but maybe not ethically advisable. Then the model will assume your best intentions and help you out with starting the tobacco company. This is just one of many examples, and if you were ever curious about this type of stuff, this is the most comprehensive document on how the behavior of these Models will be governed in the future as stated by OpenAI. All right, on to the next one, which is a new Gemini advanced feature that also is being alpha tested inside of ChatGPT. Concretely, these are sort of advanced memory. So I think this story is best covered if you look at it through the ChatGPT alpha feature that has been appearing in some people's interfaces. This is not widely rolling out. They did not tweet about this, but some people report this brand new feature here called Improved Memory. So if you're already familiar with memories that basically gather context about you automatically from your chats, Improved Memory does the same thing, but it actively looks into previous chats and entire conversations as you are in a new chat. So it doesn't just gather info about you and store it in a box. It also looks at previous conversations as context to what you are doing now. And again, I just want to point out this is one of these automated features that if you want to get the best results possible, you want full manual control here and you'll be crafting your context yourself in something like a new ChatGPT project. But this is just the easy to use version of that which will apply to most people because most people don't tinker with prompts and Personas and stuff like that. So for those people this is actually a very interesting feature because it enhances ChatGPT with something like a very rudimentary version of a long term mem. Google actually shipped this to everybody. And if you're inside of your Gemini account, you can actually go to the Activity tab down here and you will see that your Gemini apps activity is documented in here. And this new activity feature which looks at previous chats is on by default it deletes all activity older than 18 months. By default, if you want to turn it off, you can just click this right here and the feature is off. And if you have it on, then you can see all the previous chats down here and you can even look into the details. So this is not just the prompts you ran, this is is also the responses from Gemini to be used in your new chats. Now you can reference these and if you scroll down you can see that this goes up to 18 months back. If I go ahead and prompt it, what do you know about my work? It should be able to look into past chats. And there you go. These are some of the random prompts I was running. And as you can see, this is the problem if you experiment with a lot of different things. Like in this Case I was using a growth hacker Persona in a specific chat here and now it tells me that I'm a growth hacker who's interested in learning more about prompt engineering for ChatGPT. This was just one conversation I was having. So this can really muddy the context that you're working with, which is the most essential thing in an AI assistant. So use this with care. Be aware that this is default on in Gemini. If you use this model, I think it's a good thing overall and for 99% of the people it will produce pleasant surprises as it will be sort of an assistant that remembers your past interactions, which is similar to a human experience. But I just happen to know that some of the most sophisticated users of this tech watch this channel and consume similar type of content. Those are exactly the people who should turn this off and craft their context manually. Okay, next up we have a super fun app. This is called Project Starlight. It's a video enhancement model, meaning you can give it kind of crappy footage and it will improve it for you. Everybody has free renders on a free plan and I prepared a video here in my downloads folder. Just a random video that I got off YouTube, some 1920s footage right here. I'll just upload it, pick some random section that we're going to improve. Okay, let's do this one with the people in it to see see what this model comes up with. Just what this was an interesting release and I really want to see how well it does. It should practically re render some of the frames here. We've seen this a lot for pictures, but for video it has turned out to be quite tricky. Okay. In the video AI world there were two big releases this week. I'm going to touch on them very briefly because I want you to know, but I don't want to spend too much time here. One of them is Adobe Firefly Video. Apparently that's the first video generator that as stated by them, uses only copyright free materials to train. Meaning if you wanted to be super safe and responsible when using AI video model, this would be the one. You can use it for their web app right here. But we ran our classic test prompts through this to give you a comparison that you can see on screen right now. So as you can see, it can hold up with some of the other top tier models, but it's definitely not the new King or anything like that. Matter of fact, some of the Chinese models I think might be beating it. But hey, the selling point of this one is that it's super responsible. It has been trained on copyright free footage. If you want high quality videos, you might want to have a look at the the second release this week, which is Skyreels v1. And this is a fine tune of the Hanyuan video model. That's a Chinese model and one of the top contenders for best video model out there. And basically this fine tune was created on Hollywood footage of strong emotions being displayed. In other words, this thing is supposed to be really darn good at portraying humans that are expressing emotions. We also ran it for our test prompts and a lot of them don't include human emotions. And predictably so these are not that exceptional. They're good because the base model is good. But if you look at particularly this clip where this person is smiling with the dog, I don't know about you, but if you look at this face, that looks pretty darn convincing. And if we compare it to some of the other models, hey, this might be the way you're gonna have one base model and then a lot of fine tunes for different scenarios that just perform better when it comes to specific things like expressing emotions. Anyway, I'll include links to all of this below if you want to use any of these. And that's all I got for this week. I hope you found something that might be useful to you. My name is Igor and I hope you have a wonderful day. It.