OpenAI recently announced *Deep Research*. It searches the web, puts together sources of information and prepares a report. Below, you can find the link to their official release video. This release comes within a week of OpenAI's previous SOTA setting release with Operator - See [[Operator from OpenAI]]. Since Deep Research is available only for Pro customer at the moment, I have not yet had a chance to use it (yes, I am not paying 200$ for the Pro subscription). Hence, I am once again writing this up based of what I have heard rather than tried! In this short post, I have added in some bullets based of what I understood, and then I fed those bullets into ChatGPT to make it "better formatted". I have attached that too beneath. <iframe width="700" height="350" src="https://www.youtube.com/embed/YkCDVn3_wiw" title="Introduction to Deep Research" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe> ### My bullet points after listening to the talk - Deep Research model (yes, its basically a new model) take about 30 mins before coming back with an answer - Searches the internet - output - comprehensive, fully cited research paper - deep research is launching in pro first, then into plus later - when you ask it things, - it asks clarifying questions - it shows on the sidebar all the reasoning it is doing - it even shows you what page it is opening and reading / browsing - there is a status bar that shows you progress of deep research - Powered by fine tuned o3 (which is not yet public), reacts to real time information and back track when necessary - uses a python plot to create images, calculations - resulting model basically does things which would take human beings many hours - deep research model reaches 26.6% on the "Humanity's Last Exam" test, the previous best as o3 mini (high) at 13% - on the GAIA benchmark (for agents), this model also is new SOTA - "the things model finds difficult" are not things that "human beings find time consuming" - this is a very important point in my opinion - deep research works well if - you are very specific including what u want to know, format some personal thoughts - how do you know the research can be trusted? - but hey, how do you know a human being's research can be trusted? lol - but, is it "thinking" or just "aggregating with grammar" ? - how does it verify if a source is "trustworthy" or not? - what does this mean for the future of "academic publishing or research"? ### ChatGPT's (4o) Version of the same , by converting my bullets into a short post **OpenAI's Deep Research Model: A Game-Changer for Automated Research?** OpenAI has unveiled its _Deep Research_ model—essentially a new model designed to produce **comprehensive, fully cited research papers** in about 30 minutes. Unlike standard AI responses, this model: 🔍 **Searches the internet** in real time 🧠 **Asks clarifying questions** before answering 📑 **Displays its reasoning process** in a sidebar (showing the sources it browses and reads) ⏳ **Includes a status bar** tracking research progress 📊 **Generates visualizations & calculations** via Python It’s powered by a **fine-tuned version of o3** (not public yet), significantly outperforming previous models: ✅ Achieves **26.6% on "Humanity's Last Exam"** (previous best: 13%) ✅ Sets a **new SOTA on the GAIA benchmark** (for autonomous research agents) 💡 **Key Insight**: The model doesn’t just “speed up human research” but **tackles different challenges** than humans do. 🚀 _Launching in Pro first, then Plus._ ### 🤔 Some Big Questions: - How do we **verify** the trustworthiness of its sources? - Is it truly _thinking_, or just _aggregating with grammar_? - What does this mean for **academic publishing** and traditional research? What do you think—groundbreaking or just glorified summarization?