DEEP Thoughts into Open AI’s DEEP Research feature

Hey everybody. Today I'm going to be offering my insights into deep research. I'm sure you've seen videos on deep research already. I'm sure you've probably seen examples as well. And while this absolutely will sort of be like a standard review video in some ways, I'm also going to be talking a lot about implications in this video, future prospects of technology like deep research. And also I'm going to be showing you a really good free alternative because I know deep research is kind of locked away behind a pretty high price tag at the moment. Man. Guys, you have to know that when I first tried deep research, I had a moment, the first paper it ever gave me. I read it and I'm like, okay, wow. I feel the AGI. This is PhD level insights. This is PhD level research. In some circumstances we are reaching the cusp with AI. High level customized knowledge. Exactly. For whatever use case you need it for isn't just coming, it's already here and here in ways that previously were not anticipated. With models like GPT4 or even the chain of thought models, Deep research is a whole different beast. Barriers are being shattered by AI left and right. Don't have the time to conduct good research on topics that are relevant to whatever you're doing solved. Don't have the mental prowess or even capability to do the research yourself solved. And there are a ton of other potentialities that might also be solved. And we're going to talk about those concepts in this video. It's just mind blowing it and this video is going to be a really deep dive into this new technology. I'm going to show examples to you guys. I'm going to test it out for myself. You can watch it work. But we're also going to be talking a lot about implications. We're going to be reviewing the research it conducts. I've tried to make an intro for this video a million times over. I don't think there's any perfect way to dive into such a fascinating and multi sided topic. So first off, I want to get everyone on the same page in regards to deep research. First of all, it was introduced earlier this month obviously by OpenAI, the leading AI company in the world. It's integrated directly in Chat GPT as its own autonomous research agent. It looks like no matter which model you choose in ChatGPT, you can just enable it by clicking Deep Research as an option. Obviously it's only available to pro users. Again, we're talking about a great free alternative in this video as well. So for those that don't have deep pockets to buy deep research. I've got a great solution. By design, Deep Research is made to do multi step research. If you choose a reasoning model, it will reason through that research as well. And in minutes you're going to get in depth. Sometimes graduate level, sometimes PhD level reports really depending on your prompts, and it does it all in minutes. I wrote papers in college, I was an undergrad, but I did pretty close to graduate level work and this thing handily beats me out in a lot of scenarios. Comparing it directly to my papers. Mr. Altman, the CEO of OpenAI did mention that it's coming two plus users eventually. It's 200 bucks a month for 100 queries though for that ChatGPT Pro subscription. A lot of people I've seen in the community, including myself, believe it is a stepping stone towards AGI for a variety of reasons, which we'll get into some significant points here. It represents a shift from one off Q& A interactions to a more continuous autonomous inquiry that feels a lot more human and very natural. By the end of this video I really do think you will have to reconsider how high level knowledge work is going to be done in the near future because the usefulness, the value here is so deeply entrenched. Now let's Also Talk Tech OpenAI says that under the hood it's powered by a specialized O3 reasoning model. In ChatGPT though you can choose any model including a non reasoning model to use deep research with. So I'm not sure if the model you pick is completely separate from the O3 reasoning model used for deep research research. It's a little confusing and OpenAI doesn't really give us much clarity surrounding that. If I had to guess when it is conducting research it is using a specialized O3 reasoning model to sift through sources, determine which ones to use and actually go off and read them. And then whatever model you pick inside of Chat GPT is going to do the final compilation and writing. Could be wrong though on that. That's just mere speculation. The model is though apparently optimized specifically for web browsing, text processing and overall data analysis, which makes sense. It's focused and as it works it is an agent. So once you prompt it it is autonomous. Although I have noticed it almost always asks follow up questions, which is crazy. I've never seen a model do that. It asks follow up questions before it begins to conduct research to get some clarification on exactly what you're looking for. What else should I say other than it made me feel the AGI. It begins a search think iterate loop, initiating a cycle of web searches, obviously looking for information regarding whatever you prompted it to do. It will read documents off of websites. It apparently can also look at images as well as read PDFs. It really does prioritize good sources from my testing as well. Lots of direct firsthand sources from research papers or quotes from well established trustworthy researchers in a given field. It also, and I think this is one of the capabilities again that is shocking to me, mimics human research by constantly and consistently updating its approach based on any new data it learns. So if you don't know a lot about a topic and you design your prompt in a way that maybe is going to lead it down an inefficient research path, let's say it's going to be able to learn new things based off of initial research it finds and then change the way that it's going to approach the rest of the research. I mean that's thinking, that's reasoning. Human level autonomy, at least in regards to this specific task. So again, in terms of technicals, it's using end to end and reinforcement learning on browsing and reasoning tasks to accomplish these capabilities and building off of the idea we talked about with updating its own approach while researching. It has the ability to pause research, reflect upon what it's found so far, as well as even backtrack to identify any potential gaps or inconsistencies in its own research. And if you guys have ever written a research paper yourselves, you would know these are things that you genuinely have to do in order to produce a good research paper. You're going to find sometimes conflicting information, inconsistencies or gaps, and they are simply questions that need to be answered in order to flesh out your research paper. It's really truly able to do this. It feels like sheer magic. Currently all you can do is upload text into it and I believe images or PDFs now, but it can only output text. However, OpenAI has made it clear that image output and data visualization output is going to be available in the future. And also guys, all of your reports are going to have inline citations. You'll get access to detailed logs of the research process. I'm sure a lot of you have already seen that. I've also noticed that it does a pretty good job aligning to academic standards within in text citations and citations. At the end of the paper I've had IT ask me to choose APA or MLA for the paper structure before, which is pretty cool. And this is a great note. High computational costs lead to slow response times. Five to 30 minutes per query on average. Mine have been around seven to 15 minutes, I would say for most of them. But yeah, it uses a ton of compute to do this and I think that's a part of the cost. Although I really do believe that the cost to do work like this is going to plummet. This is merely the beginning. And still it's cheaper than hiring a person to conduct research at this level. And it also does it way quicker than a researcher could. I mean, we're talking about papers that would take me well over 20 hours to complete, being done in, you know, 10 minutes. Let's get hands on. I want to show you deep research in action. Like I said, you got to click the Deep Research button in Chat GPT to enable it. We'll start off small. For this prompt. Write me a highly detailed, well researched paper on medical breakthroughs and research related to AI development. We'll send off that chat. It's going to think for a little bit and it's going to now ask us follow up questions to kind of get into specifics on exactly what this paper is going to encompass. Again, guys, I've never seen an AI model just ask follow up questions unprompted like this. I wouldn't be surprised if this was included in the pre prompt. When you click Deep research. Whether that's the truth or not is kind of up for debate. It's asking us though for the specific time frame the paper is going to cover. Recent breakthroughs for the past five to 10 years or a broader historical overview. Provide a brief historical overview, but cover the most recent and notable developments. It's also asking if we want to emphasize specific areas like AI and diagnostics, drug discovery, robotic surgery, etc. We're going to say choose good examples from each subcategory. And finally there goes asking for preferred formatting guidelines. I'll just say APA is fine. Make it short but impactful since we don't have a whole lot of time right, we're making a YouTube video. We'll send this off. By the way, guys, the model I'm using for this test is O3 mini high. Almost all of my testing has been done with this and it's worked great so far. I haven't tried it with some of the smaller models like GPT4 Omni, but I would be very interested to see how the smaller models are able to handle it and especially legacy GPT4 I think that would also be interesting. As you can see it's doing its progress though this one looks like it's going to be completed in a pretty quick amount of time. It's got four sources so far now five and if we click over here we can see the activity on the side. We can also see all the specific sources. A lot of this guys is going to be related to research. You'll see a lot of Edu, you'll see a lot of.gov sometimes you'll see other sources like Wikipedia. You know some people are not cool with Wikipedia being used as a professional source. It also likes to get a lot of medium articles as well. So it'll go to publications but it'll for the most part make sure that they are trustworthy. Stanford. It also loves to go to Stanford from what I've seen as well. But look at the internal thinking here too. It looks like it began by gathering some good context around AI in healthcare. Famous examples, maybe some older stuff like Google AI, Diabetic 2016 study. It moves on then to articles that are providing detailed summaries of significant advancements. But it'll do things that are not necessarily based on just finding things on the Internet too like mapping out AI's overall evolution over time, covering clinical expertise and diagnostics. And I mean we're starting to get pretty used to seeing this with chain of thought models now but it's still so mind blowing. I knew it was eventually going to happen upon and search for the AlphaFold 2020 protein folding breakthrough. They won a Nobel prize for this. So all great stuff so far. This is that search think iterate loop in action mimicking that human level research updating, changing its approach based on new data it's discovered. At the end of the day we saw 23 total sources which is pret large a lot of dot govs in here. Stanford SGU, EDU, Northwestern. EDU I mean these are pretty solid sources. And finally folks, it does result in the research paper and it reads like one too. Artificial intelligence has been applied to medicine for over half a century evolving from simple expert systems to advanced machine learning models. It starts off here talking about early milestones dating back to the 1970s. It mentions Internist 1 from 1971 that was one of the first computer programs to exist in diagnosis serving as a medical consultant. And you can click on the source and will actually pull up the highlighted text where it pulled it from. Looking back here it seems like it was a simple search algorithm to arrive at clinical diagnoses Based on symptoms. Now, I'm not going to dive obviously too deep like into this research paper. There's so much to talk about even with just this. It's providing you so much information from 23 sources. The point I want to make is that it's choosing to start off with early milestones, just like I prompted for kind of taking us through, taking us on the journey up to present day. That's exactly how I would introduce a research paper. It talks about other landmarks, like IBM's Watson outperforming humans in jeopardy, bringing it even closer today. By 2017, Watson was being used in medical research to identify proteins linked to diseases like ALS and then starts to break out into different areas of expertise. It's very well formed, formatted and organized. We start off with diagnostics, we talk about deep learning for medical imaging. We talk about early disease detection, a deep learning system by Google that analyzes 3D Lung CT scans meeting or exceeding that of expert radiologists. It moves on to drug discovery, yet another category, and it breaks it down all with these sources. Obviously it's going to have to talk about the incredible DeepMind AlphaFold2AI system. It was something like 200 million proteins were processed by the AI. Thanks to AI, researchers can now target proteins and design inhibitors that were once undruggable due to lack of structural data. It's really well structured. It's really well structured because this is exactly how I would write my research papers and the best way that I found to write them when I was in college, where you take it by category, link the sources that are related together, give a brief summary and move on to the next category. You know, it talks about robotic surgery, it talks about personalized medicine, and then it finally comes to a conclusion. Now, I did ask it for a briefer paper just so it's a little bit easier for us to take a look at the structure here. But when we look at some of the more advanced papers that it's written for me, you'll see that it does more than just summarize information it's gathered from the Internet. It can also draw its own conclusions. It can make novel connections and really think about how different perspectives can interact or different research perspectives. Take a look at this example. I had IT do a comprehensive analysis of AI growth trends here. The prompt was we need to conduct research into the growth of AI. I want to understand its rate of growth and capability development over time. I want to understand if evidence exists to suggest that we are nearing the tipping point of exponential growth and how or if that could be made possible beyond what is currently speculated by potential emerging capabilities such as self improvement and autonomy. Again, it's asked a bunch of follow up questions here. We had six I think, which I went through and then it went ahead and began the research. This paper is a little bit more advanced than the last one we looked looked at. It does though begin again with a contextual background for us, taking us through how AI has been progressing and it found some really intriguing insights. I think this prompt is a little bit of a better example to show off the power potential this has. It was able to sift through and figure out that from 2012 to 2018 this is pre chat GPT we're still talking about. We still saw a 300,000x increase in computer for training AI models and it notes that this was an exponential trend as well. Even all the way back then, a 3.4 month doubling time far outpacing Moore's law, which I'm sure that you famously have heard at this point. This growth was underpinned by advances in algorithmic innovation, massive training data sets and specialized hardware like GPUs and TPUs. It's so cool to read that too because we were on the cusp of this. It talks about more recent development in the past few years because, well, a lot of AI development has been in just the past few years. We're really getting on the tipping point potentially of it, which is kind of what this paper is talking about. It brings up the foundation models we all know and love. GPT3, GPT4 Bert Palm vast Internet scale data. GPT4 demonstrated human level performance on numerous benchmarks, passing a simulated bar exam in the top 10% of Test makers famously. Oh, that's a classic one. I remember covering that. I remember being on the cutting edge there. Again brings up AlphaFold2 solving the 50 year protein folding challenge. I really like this too because this is kind of like the thesis of the paper. These milestones illustrate AI's crystallized intelligence. LLMs and related models act as repositories of human knowledge, excelling at tasks that draw on accumulated information. What a great, what a fantastic way to distill what we're talking about here. But yeah, as you can see it follows again the structure of breaking it down by category, talking about computing, scaling trends. We talk about GPUs and TP, A100 GPU, the H100 GPU obviously very influential high compute processing that has taken us to where we are today with AI. Talk about cloud computing, Azure for OpenAI Google Cloud for DeepMind, around a trillion dollars a year in hardware spending, AI model performance benchmarks and scaling laws. It talks about the SAT AP exams. It talks about models like Clip, talks about investment trends in AI development as well as in just 2021, investment was at $93.5 billion, which was double the previous year of 2020. 2022 saw a small dip, but a strong rebound in 2023, focusing in generative AI. What a great insight. I mean, we were on the precipice and the cusp of all this stuff, so we know all about it. But yeah, that is extremely accurate. That is a very accurate summarization. In 2023, funding for generative AI octupled. Octupled. And I think the real kicker here is the volume of AI research has exploded. The insights it was able to get here. Between 2010 and 2022, the number of AI publications nearly tripled 88,000 to 240,000 per year. And I'm sure it's grown a lot since then. Probably doubled again. Yeah, doubling roughly every two years. So we're probably above that now. They talk about breakthroughs and innovations, transformers, generative models, reinforcement learning. It moves on to talking about self improvement and autonomy, which is a critical aspect of AI's potential growth, especially if it's going to be exponential. And you have to think about, when it's writing and researching this paper, when it makes statements like this, a critical aspect of AI growth is the emergence of systems that can improve themselves or design other AI models. It makes a note that it's a stepping stone towards autonomous AI development. It makes sure to preface this because we are about to actually dive in to that specific topic. Even below AI systems that are, you know, capable of some level of self improvement, it knows it needs to talk about the role of autonomy in accelerating progress. Agentic systems, recursive self improvement, theoretical constructs in line with I.J. goode's 1965 idea of an intelligence explosion. I mean, can you imagine being that much of a forward thinker all the way back in 1965, suggesting that once AI reaches a certain capability threshold, it could start improving itself without human intervention. Finally, it gets down to the part again, which I think is the most telling and the most valuable from this whole paper. An analysis of the tipping point. We ask the question, is AI nearing a phase of exponential acceleration beyond the already fast pace it's at? We introduce again that original idea from 1965, an intelligence explosion of self improvement. Expert opinion from Masayoshi Son From Softbank, he had to accelerate his AGI timeline prediction, suggesting AGI could arrive within two to three years as opposed to the very late 2020s. Sam Altman and other AI lab leaders hinting at current models being on the cusp of greater reasoning capabilities. Based on all of its research, it was able to novelly come up with potential triggers that would lead to an exponential phase. So essentially based on what it's seeing right now in the space from its research, it's able to reason about and consider these, these triggers. It's like a human man, but yet massive compute breakthrough. Obviously, you know, quantum computing or a new paradigm like analog AI chips or optical computing would obviously allow us to test architectures we can't run today. It also talks about the idea of novel architectures or algorithms themselves. You know, something similar to the discovery of the transformer unlock these new efficiencies and capabilities, bringing back around that self improving AI agent idea. AI that can rewrite its own code to run faster on given hardware or use resources more efficiently. A sustained positive feedback loop, AI making better AI, which in turn makes even better AI, etc is essentially the singularity scenario that, that people bring up and talk about a lot these days. Collective AI systems that are working together, coordinating could better AI, et cetera is essentially the singularity scenario that people bring up and talk about a lot these days. Collective AI systems that are working together coordinating could yield an AI scientist that accelerates research leading to more powerful AI. Again, the self improving agent scenario. And by the way, this sounds very similar to what Sam Altman recently brought up regarding what GPT5 is going to be. That's an important note. He was recently quoted saying that hey, GPT5, it's probably going to be a system that knows when to use reasoning, knows when to use a lighter weight model, knows when to use tool use or image recognition, etc. We also see that if AI ends up significantly boosting economic output, which I think it definitely will, more investments and more data are on their way to produce better AI. I really like the note down here too, that any combination of these factors might usher in exponential phase beyond the already fast current growth that we're seeing. Even pointing out the key notion of recursive acceleration. When AI starts to meaningfully contribute to improving itself, even indirectly, that growth can compound. And then it finally gives, you know, a nice analysis here, really great insights. It makes note that things we might have forgotten about, such as GPT3's few shot learning capability or GPT4's ability to do multi step reasoning are phase changes that hint at non linear scaling of capability with size. And it really seems to only be growing from there with things like chain of thought models or even internal inference reasoning models, which is a brand new thing. Haven't talked about that on this channel just yet. I did talk about it on my Twitter though. So you guys are aware. Go check out my Twitter. I'm always talking about the latest. As long as funding keeps at pace, increasing costs won't deter scale. Overall, the investment data suggests a willingness and the market as a whole to push towards that edge. Where something truly novel might occur, such as unprecedented growth, it comes to a final verdict as well, which I really like. I like that it comes to like this final like smack dab point at the end of the paper saying this is where we're at with this, this idea, with this concept, bringing all these concepts together into one statement, it is plausible that AI is nearing an inflection point. While it's hard to predict exactly when exponential becomes super exponential, the convergence of scaling, funding and early self improvement hints at a coming acceleration Expert sentiment overall though is mixed. I'm glad that it brought that to the table because it is. Some see AGI in years, others see it in decades. But the risk of underestimating AI's momentum is widely acknowledged because it's happened several times over. It's happened to me. I've seen the power potential of AI and actually jumped on it early and then still underestimated it. So that's, that's the thing with AI. That's why a lot of people are saying like buckle up because we're nearing exponential growth. It makes predictions saying, you know, if GPT5 or analogous systems emerge in say 1.5 years. By the way, Sam Altman said that such such a framework would appear in weeks or months, not 1.5 years. Keep in mind, with far superior capabilities, that shrinking interval would signal acceleration. The qualitative nature of AI is also changing. Another really solid point it's bringing up here from passive prediction models to active agents, which could mark a phase shift. In summary, the overall analysis suggests AI is approaching a tipping point, though it might manifest as a rapid series of advances rather than a single moment. Traditional scaling laws which assume diminishing returns eventually are bypassed by new dynamics. AI optimizing itself, for example, yielding faster than expected gains. And then it even breaks down into the implications here, how it's going to affect various industries, societal impact. I get we're kind of running out of time here. We could we could dive into this in just its own separate video? Which I would love to do. By the way, though the point here is that it's not just regurgitating things from its training data or regurgitating things from the Internet. It does so when it needs to, like in the beginning of the paper, introducing topics and facts and research that we know about and evidence. When it comes to the part where the rubber really meets the rope, the implications, conclusions, this is applying the knowledge. If we were, you know, a teacher grading this as a student's research paper, this is the part that we would mostly be thinking about because this is application of the knowledge in a creative and novel way. And that is what the, that is what deep research is doing here. I mean, we knew AI could do this before, but now it's so much easier to get something so high quality and well researched. Good data in, good data out. I mean, look, look at some of these impacts. AI assistants could become used, ubiquitous, handling everyday tasks like scheduling information queries, even complex errands via robots. Not like a wild idea at all. We're already seeing humanoid robots being developed rapidly and competitively. Scientific advancement as we already spoke about personalized AI tutors revolutionizing learning, tailoring education to students need. It also brings up safety considerations talking about bias, privacy and misuse. If agents become autonomous, ensuring that alignment with human values exists is obviously a necessary outcome. Society might need new institutions or personalized AI tutors revolutionizing learning, tailoring education to students need. It also brings up safety considerations talking about bias, privacy and misuse. If agents become autonomous, ensuring that alignment with human values exists is obviously a necessary outcome. Society might need new institutions or international agreements focused on safety and governance. It also talks even about existential perspectives. In the extreme case of international agreements focused on safety and governance. It also talks even about existential perspectives. In the extreme case of super intelligent AI, the implications are existential. Potentially, it itself writes a 10 to 20% chance of AI posing an existential threat means that society must treat AI development with caution. Doesn't stop progress, but but implies parallel investment into safety research. And fair enough, especially for a research paper like this. It even dives into preparing for the future. Embracing AI's benefits like improved productivity, scientific discovery and quality of life, while mitigating the risks such as potential job displacement, ethical pitfalls and autonomy. Without alignment, it does conclude that AI is growing on an exponential curve with hints of an upcoming tipping point point. Let's talk about why I think this is absolutely a significant step towards the ever big Idea of AGI or artificial general intelligence. I'm going to provide a bunch of examples down in the description, full papers that it's written. But it absolutely excels in a wide diverse range of fields like linguistics, rocket science. Deep research in my mind really does suggest a leap towards general intelligence. It really does also go through through very human esque problem solving. It can gather the resources, trace back and realize when it needs to maybe disregard a source or worry about overlapping information or contradictory information and how it should present that to the viewer in an honest way. It really mirrors a human researcher's process. It's just like when I used to write research papers in college. It actively seeks out the best sources, very specific specialized information. It can adapt and change its research strategy on the fly, refine the end output. That's all human like problem solving skills. Obviously, you know, we're seeing some practical challenges with it right now. It requires a ton of compute. OpenAI is charging a lot for it right now. Again, free competitor. We're going to talk about it very soon, don't worry. And it does still obviously remain completely dependent on human prompting and human oversight. I've seen it get things wrong before, ever so slightly. I mean, when you're writing a research paper, sometimes wrong can be up for debate. Even like I saw it, talk about some ideas from a certain philosopher and I'm like, I don't think that philosopher actually believes those ideals as much as it was displaying them. Semantical. Semantical though yes, it does face practical challenges very much still relying on us for oversight. But I do think it's a wake up call. This is rapid pace AI development. It's rapid pace AI evolution. I don't think any of the models from Google OR DEEP SEQ R1, for example, have a close enough competitive edge to deep research, but I wouldn't be surprised if they put something out pretty soon, thanks to the free market. And let's also talk implications. Specifically, I want to focus on future research and potential future breakthroughs. As I mentioned in the beginning of the video, barriers are now smashed. They're broken. We can now compile and review thousands of different research papers filled with knowledge, as well as look at the hard data and the science behind them and make new discoveries and novel thoughts about these in a fraction of the time it would take a person to do so. It's also worth mentioning that researchers now can grab their information a lot more quickly and focus their efforts better on analysis, coming up with new hypotheses and coming up with new creative Ideas. And again, the AI, like I said, I think the a AI is capable of all these things too. And I, I think you just need to make sure you are prompting for them. And still you never know when a human is going to come up with an idea that the AI would never think to consider. But that also is a two edged statement. Same thing could be said for the AIs I think. And let's talk about prompting deep research too. Out of the box. It's like a research assistant, it can gather great data. The examples I showed you were pretty straightforward. Not really pushing the model to do anything that's a little bit more strange, we could say, but you absolutely could prompt it in that way. And I'm going to be doing more testing with that. Think about taking two completely separated fields, like, I don't know, I'll come up with a quick example. Biology and space exploration. We could use deep research to gather good information on both and then bridge a gap in between them, creating a new novel connection saying, you know, what kind of biological life forms could potentially exist on Mars or have existed on Mars in the past? Or how about Neptune or another planet or another solar system? It's just an example. You can bridge different scientific communities together by comparing and contrasting their research and information. Gain really valuable new insight. That's creativity and that is novelty. And you can pretty much prompt it to do exactly that. You can just instruct it. Traditional research settings might overlook things like that. They can end up in their own bubbles for sure. The AI is not constrained by such things. I mean there are a million ways you could think of to prompt it in order to identify specific gaps in existing research or suggest new avenues for research and exploration. Now, despite all of these awesome capabilities and implications, we do still need to be somewhat cautious. As I mentioned, it can still hallucinate potentially very, very low error rate, but it's definitely possible. Take its paper and run it through again and say like hey, double check this. Probably a good idea. Alright. Finally, before we get into conclusion time, let's talk about the free alternative that just released today. Check it out, check it out. Perplexity AI comes to the table with their own deep research on Perplexity. IT scores a 21.1% on humanity's last exam, which is a very difficult benchmark. And Perplexity's deep research here has five prompts for free each day, so you can use it entirely free. Very cool. We're gonna run a quick prompt in a minute. But yeah, right now on humanity's last exam a score of 21.1% for perplexity deep Research is really impressive. I mean you can see how low models like GPT4, O Grock 2, even 3.5 Sonnet are scoring on this benchmark. Even Gemini Thinking is nowhere close to this. It actually outperforms the quite newly released O3 mini high. But at the end of the day the top model still is OpenAI's variant of deep research which gets a 26.6 I believe. So totally crazy. It makes it stand out as a really good competitor. Here again, free access to Deep Research for those who don't want to pay the $200 a month price tag for the OpenAI variant, here's another like face melting benchmark from this new perplexity deep research. 93.9% accuracy on SimP QA you can see 01 preview is pretty far down the list here. Not even half as good as Perplexity Deep Research and it just stomps over everything from 01 mini to GPT4, O 3.5 sonnet, etc. Also worth noting that Perplexity's Deep Research does seem to complete tasks a little bit quicker than OpenAI's variant, saying most tasks are completed in under 3 minutes minutes. So here we are on the free plan on the Perplexity website. If we click this button right here that says auto we can scroll down. We see a pro search. They've got reasoning with both O3 mini and R1, but they've also got their brand new Deep Research that does in depth reports. So click on that here. Let's give it the same exact prompt that we did for the AI tipping point 1 and this was a really good paper again. I wonder if it'll produce a worse or better paper than this. All right, I sent it off. It is now taking a few minutes to complete. It's preparing a research plan. I don't think this one's going to ask any follow up questions. Summarize some of the search results. It's gotten eight sources so far apparently. Definitely different sources than I'm seeing from the likes of OpenAI here. Hey, it is trying to address certain gaps from the last research it completed. Now we're up to 26 sources man. It's just gathering more and more information I guess. Well we'll let it do its thing. Alright. It looks like it's.