Learning with AI via deep research and study mode

Recently, I started using LLMs to answer some of my questions. I’ve been using both Gemini and ChatGPT in parallel, asking the same question on both. In fact, I was asking the same question 4 times, as for each one of these two engines I was using both the Deep Research and the study modes – known as “Guided Learning” in Gemini and “Study and Learn” in ChatGPT. In this article, I’ll summarize my experience with all 4 of these, after using them for nearly 3 weeks, often getting to the end of the available quota.

First, looking at the Deep Research mode, both engines work mostly similar. You ask a question, they go to the internet and search for articles that are related to your query and you get a long answer. You need to wait at least 10s of minutes – decaminutes? –, but in general the results are quite comprehensive. I asked math questions, history questions, even politics questions and got answers that looked right. We still need to always check the results, and I did so.

The nice benefit here is that before these modes one would have to do the internet searches by themselves, collect the information, structure it out, and so on. All of this while navigating multiple websites, different pop-ups, ads, sites that contain the same information or sites that contain information that is irrelevant – like getting information about Haskell, one of the cities in the US, instead of Haskell the programming language (this is also a huge problem on Bluesky, I tried to configure a custom feed for Haskell topics and it merges the cities with the programming languages – it even started adding information about one of the Haskell cities to a programming custom feed, but I really went on a tangent here…).

The downside is that you only get the information that is in the articles that got hit during the search phase. If you were to do the internet search yourself, then you could adapt the search after reading each article. The process would be more adaptive, but also slower.

Overall, I think it’s a good idea to fire up such Deep Research queries to start literature research, but it’s still better to complement with internet searches done without the LLMs.

One can still ask the LLM follow-up questions after the first results are in. And this is where the difference between ChatGPT and Gemini arises, at this time. ChatGPT’s first answer – the result of processing the online search – is in its own div, which can be popped up and then saved to an article. But, subsequent answers just follow the conversation. The answers are large, with structured headers. And, they end with questions about what else you might want to search, related to the query. For example, in a Deep Research session about partial sums, and a follow-up that focused the search on a specific form for the terms of these sums, ChatGPT then proposed deriving closed forms for the sums, analyzing a concrete example, or suggesting what changes when the sums become integrals.

On the other hand, Gemini starts with a report that is always visible on the main div on the page. Future questions result in Gemini updating the same report, but sometimes it is hard to see what got updated. There is no diff being displayed, although the response in the side diff with the conversation tells you what sections got updated. There are a few cases where the answer comes in the conversation side div, but in general it is the report that changes. Gemini also suggests future opportunities to extend the research, but in general we get only one suggestion each time. On the same example mentioned above, Gemini suggested looking at how the partial sums would look like in finite field. This was something totally unexpected, a moment of serendipity.

Next, let’s move on the study modes. The fact that they are named differently in the two engines is not just a naming difference. In fact, Gemini’s Guided Learning mode is somewhat close to ChatGPT’s Deep Research behavior: at the end of each reply it suggests you 3 directions to continue the discussion on, to learn more. Gemini starts by directly answering the question, gives some detailed answer, and then suggests the 3 avenues for continuing the learning process. It is faster than Deep Research mode and still suggest serendipitous connections.

Next, ChatGPT’s study mode. It takes your question, rephrases it a little and then asks you what level of expertise you already have with the field. It either asks this directly, or just poses a question to check it. You answer the question, it judges it and then gives a little bit more information. The process then continues, it is a live version of the Socratic method. I found myself using this mode quite a lot and the questions at the end were very different than the questions at the start. It is a very good personal tutor and can become addictive.

Overall, Gemini (in both modes) tends to give more multi-modal responses. You get tables, diagrams, code, plots. You get formulas and text, with section headers, nice formatting. The answer are on the longer side. ChatGPT on the other hand tends to give only the information that is being asked for, with as little extra as possible. When it provides some images, it gets them from external sources and links to them, whereas Gemini seems to generate them via plotting / Nano Banana.

All 4 of these modes are really useful. But, one needs to still be in control. Check the outputs, use the tools to learn more, and enhance your work. Don’t use them to do the work for you. Take advantage of the speed-up, don’t let these tools replace your brain.

For future articles, I’ll also look at the various AI coding tools. But that won’t come soon, as I want to prepare a few more things first. And we still have to continue the spirals on a grid puzzles. We looked at how LLMs scored in the last article, but there needs to be some follow-up – although I now realize that it won’t be the one I was planning at the start.