Are we closer to super intelligence? It has become so commonplace to look for an answer on the web today that we have a verb for it – why not try and 'google' that? Now, a newly dropped Large Language Model tool takes that idea and supercharges it.
Users describe Open AI’s Deep Research tool as like having a specialised assistant that can analyse hundreds of websites, textbooks and academic papers, compile a detailed research report, and even identify knowledge gaps. And it does this all in about the time it takes you to make a cup of coffee, drink it, have a stab at the crossword and think, 'Gee, I really should start on that research report'.
Released at the start of this month, Deep Research seems to represent a significant leap forward in AI capabilities. It's powered by OpenAI's o3 model, which has been specifically optimised for web browsing and data analysis. But what makes it special isn't just its speed – it is the model’s ability to think through complex problems step by step, much like a human researcher would.
What are users saying?
Some early adopters are big fans – Professor Derya Unutmaz from the Jackson Laboratory reports that Deep Research showed impressive sophistication when he asked it about CAR-T cell therapy. "OpenAI's product was better at prioritising the most relevant pieces of information," Unutmaz noted, comparing it favorably to Google's competing tool Gemini.
However, one of the world's most respected scientific journals, Nature, strikes a more cautionary note. Their analysis points out that while Deep Research can generate reports in "tens of minutes" that would take humans many hours, it still comes with significant limitations - including the potential to hallucinate facts or fail to distinguish between authoritative information and less rigorously established thoughts.
The smartest AI ever?
What's particularly fascinating is how Deep Research has performed in what's being called "Humanity's Last Exam" - a collection of some of the hardest reasoning problems known to science. While its claimed score of 26.6% might not sound impressive (let's face it, when most of us scored 26% in something in school, we weren’t racing home to tell mum and dad!), and while all claimed test results by AI platforms have to be taken with a degree of scepticism, if true this result represents a massive improvement over previous AI models. By this metric, it could lay claim to being the smartest Large Language Model ever!
Worth the money?
The tool isn't cheap - at $200 per month for Pro tier access with a limit of 100 searches, it's clearly aimed at serious researchers and professionals rather than casual users. It's also quite compute-hungry, sometimes taking up to 30 minutes to complete complex queries. But as I mentioned with Unutmaz earlier, the people commissioning these reports often suggest it would have taken them weeks to do the same.
At this early stage, it might be best to think of this as less of a replacement for human researchers and more as a powerful first step. Back to my original assessment, perhaps it's like having a smart research assistant who can quickly pull together an initial report - but one who occasionally needs fact-checking and supervision. A powerful tool in the kit rather than a substitute for human analysis. After all, even the smartest AI still needs someone to ask the right questions - and more importantly, to know when to question the answers.
Where to from here?
When assessing first generations of these models, it's important to remember how much our ability to write, code and generate graphics with these tools has advanced in just the last two years. If Deep Research continues on anything like that trajectory, it and similar projects could have a tremendous impact on the field of scientific literature and even one day complete independent research without human intervention.
As I regularly find myself saying regarding AI... strap yourself in for the ride!
In other news: scientists have recalculated the flight path of Asteroid 2024 YR4 (which I wrote about here) and decided it is almost definitely not going to smash into Earth seven years from now! Read more
That’s all from me for now. If you'd like more geeky fun, please check out my other newsletters below, or connect with me on LinkedIn and/or X.
Yours in nerdiness,
Adam
Comments