OpenAI's Deep Research Agent in 8 Minutes

--- type: transcript date: 2025-02-03 youtube_id: cR6WxQvMrHM --- # Transcript: OpenAI's Deep Research Agent in 8 Minutes openai has just announced deep research which is their second AI agent after operator and with this deep research feature this is an AI agent that plans and executes multi-step research workflows they mentioned that it could take anywhere from 5 to 30 minutes for the full research report to be generated how this works is you can ask for what you're looking for it will ask some clarifying questions just to make sure before it goes off that it does have what you're asking for correct and any clarifying detail that it might need and then it will go off and begin researching so you can see here it starts researching and then step by step as the agent goes through the different tasks it's going to highlight what it's doing in this we can see it's searching different websites it's Gathering particular details from each site and going through what potentially could be dozens and dozens of websites before it Returns the report back to you let's just quickly go through some aspects of the blog post today we're launching deep research in chat GPT a new agentic capability that conducts multi-step research on the internet for complex tasks it accomplishes in tens of minutes what would take humans many hours this is going to be included in chat GPT and you'll be able to find analyze synthesize hundreds of online sources to create a comprehensive report at the level of a research analyst one of the interesting pieces within the blog post that they mention is that this is actually powered by one of the upcoming openai 03 models that's optimized for web browsing and data analysis this deep research model is actually the first publicly available model where we're going to be able to access that full 03 model and not just these 03 mini models that came out on Friday so they mentioned that this is optimized for both web browsing but also data analysis within its abilities it can reason for what it searches for it has a code interpreter where in the announcement they mentioned that you'll be able to create visualizations from that interpreter you can imagine if you have a data heavy task that requires it to plot some sort of data they that this feature will be able to analyze massive amounts of text like hundreds of pages potentially but it can also analyze images as well as PDFs on the internet as they described in the announcement it's going to go through a page and crawl all of those different pieces on the page whether that's an image or potentially a PDF within the website it looks like it's going to have the capability to reach all of those different assets just as if you were researching this yourself they mentioned that deep research marks a significant step towards our broader goal of developing AGI which we have long envisioned as capable of producing novel scientific research in terms of why they built deep research they mentioned that this is for people that do intensive knowledge work in areas like Finance science policy as well as engineering and need thorough precise and reliable research but in addition to that they mentioned that even for things like shopping if you want hyper-personalized recommendations on purchases that typically require careful research like cars appliances and Furniture it will be a ble to do that type of task as well in terms of how to use deep research now this is going to only be available in Chad GPT Pro this is their $200 a month tier at time of recording it will be coming out to chat GPT plus but the timeline for that wasn't yet announced if you are a pro user you'll be able to select deep research in chat gpt's message composer input your query with optional file attachments if you have any and then receive a detailed report complete with a sidebar summarizing the research as well as all of the different cited sources all of these are going to be delivered asynchronously and the process generally takes like I mentioned 5 to 30 minutes or so within the blog post there's a great comparison section where it shows gepd 40 as well as deep research with the same queries here's a good example under ux Design Within their blog post find evidence that shows that buttons icons and labels are more usable than buttons without Labs or labels without icons I know there are a lot of user studies on it would love to see a detail report within here we see gp4 O's response is relatively brief it really doesn't give us enough information but if we just look at the Deep research example I can just scroll through this and continue on and on now the other thing to know with this that you can't really see within the UI is once it's within chat GPT you will be able to see all of the different sources that it references for whatever task it might be here we can see the references at the bottom here just like you would have with a professional report there's some other examples in here that are really good as well you can take a look at the business example where a similar case here it will give you a very detailed Report with tables as well as detailed metrics on particular things that you ask for they mentioned that deep research was trained using endtoend reinforcement learning on hard browsing and reasoning task across a number of different domains another interesting piece with the announcement was they mentioned that it learned to plan and execute multi-step trajectory to find the data that it needs but one thing that stood to me was that it backtracks and reacts to real-time information where necessary that is very helpful cuz a lot of these early tools would go and try create a research report oftentimes the problem that you would see is you would let it go and do its own research but then by the time it actually gave you a report it often times would deviate from that original goal keeping it on track is obviously something that's really important another interesting thing with this and I mentioned this already but it can create graphs within python it will embed those within the report but it will also gather images from websites as well within its responses now another impressive metric is on Humanity's last exam this came out as we started to see a lot of these benchmarks starting to become saturated we have more and more of these models getting closer and closer to that 100% doing things perfectly across some of these benchmarks that we were previously looking at on how to test a lot of these models but what's interesting with this is that this test consists of 3,000 multiple choice and short answer questions across a 100 subjects from Linguistics to rocket science Classics to ecology as we can see on this we can see that this scores a 25.3 in terms of accuracy when we compare that to Open Eyes 01 that is only added 9.1 and with opening eyes 03 mini which just came out on high mode that scores a respective 133% there are some other benchmarks within here if you're interested in checking them out now another interesting piece with this is they mentioned that the more the browses and thinks about what it's browsing the better it does which is why giving it time to think is important that's one of the big differences with this model it's not really focused on trying to get you that response as quick as possible it's really focused on trying to give you a report that's highly detailed as well as specific and accurate where this is really great is you can see how much time is saved across a number of different disciplines and this is honestly going to be something that I think a lot of people will have a lot of value using effectively what you can do with this model is give it really hard tasks and just send it on its own and you'll be able to save a ton of time by leveraging it there are obviously still some limitations these models can hallucinate facts in responses and they mentioned at launch there might be some minor formatting errors and things like that that over time obviously these things will improve but time of recording it's only available to Pro users with only 100 queries per month plus and team users will get access next followed by Enterprise they're still working on bringing it to the UK Switzerland as well as the European economic area another piece within this all paid users will soon get significantly higher rate limits when we release a faster more costeffective version of Deep research powered by a smaller model that still provides high quality results in terms of being able to access this if you subscribe to the pro tier you will be able to get this today on chat jpt web and they are going to be rolling this out to both mobile as well as the desktop app within a month they mention that in the future you'll be able to connect to more specialized data expanding its access to subscription based or internal resources to make its output even more robust and personalized but overall that's pretty much it for this video let me know your thoughts on something like this do you think it's worth the $200 a month for a 100 queries if not how much do you think something like this is worth is this valuable in the type of work that you do let me know your thoughts in the comments below otherwise if you found this video useful please like comment share and subscribe otherwise until the next one