AI Visual News
Deep Research using ChatGPT 'o3' model
Season
2025.02
Summary
[ GPTers OpenAI Workflow Study Case ]
I recently tested the updated 'DeepResearch' feature on the o3 model.
($200 Pro Plan)
.
I recently tested the updated 'DeepResearch' feature on the o3 model.
($200 Pro Plan)
.
Category
- AI LLM
- OpenAI
Tag
- Updates
Dates
2025/02/06
Created by
- mintbear
GPTers OpenAI Workflow Study, February 2025 case study.
Hello! This is Mint Bear.
We tested the DeepResearch feature, which was updated this week, on the “o3 model”.
Objective 1. Evaluate the value and performance of Agent AI by determining how much in-depth search and analysis is actually performed.
Objective 2. We wanted to obtain data or insights on the ‘AI image tool analysis’ derived from the performance results.
How to proceed
We leveraged the DeepResearch feature in ChatGPT's 'o3-mini-high model'.
($200, using Pro Plan.)
Step 1. Select GPT model & In-depth research & prompting
I did this by selecting [Model] in the upper left, clicking the [Deep Research] button at the bottom, and entering the prompt below.
Currently, check out the best AI image generation tools available for general public use.
Also check and update for recent changes in various other tools.
There are currently several versions of the 4o, o1, and o3 models available (the 4o and o1 models are the ones I have been using).
This time, we decided to test the Deep Research feature on the o3 model.
This time, we decided to test the Deep Research feature on the o3 model.
We ask that you proceed by referring to the information on the image creation tool site operated by Mintbear.
I thought it would be useful to provide users with guidance and intent about their data.
I thought it would be useful to provide users with guidance and intent about their data.
Step 2. Ask o3 questions.
He didn't just give me an answer. He asked me a question.
So I answered as follows:
All factors are important, but since each item is weighted differently depending on the user, I think we should go with an integrated evaluation.
However, since I have set up a general user, please suggest something that excludes learning curves that are too difficult.
After answering as above, we started conducting full-scale [in-depth research].
Step 3. In-depth research - inference activities and sources
We made inferences in various steps from 19 sources over a period of about 5 minutes.
Here's a quick look at the meaningful in-depth research process:
Here's a quick look at the meaningful in-depth research process:
[ Search and Inference Process ]
1.
Review Mintbear's list of AI image tools at the provided URL.
Read more about the image tool on the Mintbear website
> Open the slash page and check the contents of Delft.
> I think the slash page contains Midjourney, Dreamina (free tool?), and Sora related items. I keep scrolling and looking.
> I'm looking at the list of AI image generators. This list is divided into various categories, so I'm looking into it more.
2.
I understand that anything highlighted as 'free' in the list of tools is 'recommended'.
> This page lists various image generators. I recommend "Dreamina" as a free option.
> I'm looking at various image generation tools, and there are some great options out there, including Midjourney and DALL-E 3. The Stable Diffusion-based tools are also interesting.
> We are looking at the latest trends in Midjourney and DALL-E 3, LeonardoAi based on Stable Diffusion, "DreamStudio" and "Adobe Firefly".
3.
After that, we continue our inference by searching related websites.
> The features and pros and cons of the recommended tools are compared in a table format, and the final recommendation is comprehensively presented.
> We compare Leonardo.ai, Midjourney, Ideogram, DreamStudio, Adobe Firefly, Canva Magic Media, and more to analyze which AI image creation tools are suitable for general users.
> I'm scrolling down further because it seems like there's more content down there.
> DALL-E 3 focuses on ease of use, while Midjourney delivers the best results. Ideogram handles text in images, Stable Diffusion supports user customization, and Adobe Firefly integrates into creative work. Other tools like Canva are likely to be included as well. > Referring to the Zapier article and TechRadar report, DALL-E 3, Midjourney, and Adobe Firefly are expected to be in the spotlight in 2024. Each of these tools offers unique features.
> Every time I use it, I think of a new way, and I'm looking for the optimized results of Zapier DALL-E 3 and Midjourney. > Well, looking at Zapier's list on those platforms, DALL E 3 is easy to use, Midjourney gives the best results, Ideogram is useful for precise text writing, Stable Diffusion is ... (suggesting the possibility of customization and control or open source) "Leonardo for something" or "Adobe Firefly for something" or "Canva for quick social or integration".
Etc.
O3 Deep Research continued to do some very interesting searches and inferences.
[ Source ]
Although the sources were listed as 19, which is quite a lot, I was disappointed to see the results as a uniform source.
So, I realized the need to request a variety of search targets.
We actually did a post-test to see if we could use a variety of references.
We actually did a post-test to see if we could use a variety of references.
But the reasoning process itself was very reasonable,
Despite the sparse evidence base, we have reached generally useful conclusions.
Despite the sparse evidence base, we have reached generally useful conclusions.
Step 4. Draw the first conclusion
O3's primary in-depth research was based on inferences drawn from a variety of sources, resulting in very useful and objective information.
It was more satisfactory than the existing 4o, o1, but it seemed that biased information was dealt with in general insights.
It was more satisfactory than the existing 4o, o1, but it seemed that biased information was dealt with in general insights.
This is something I see because it is my area of expertise.
To a non-expert, this may seem like useful and solid information.
To a non-expert, this may seem like useful and solid information.
The upgrade of the function has been confirmed, but
Mintbear, not satisfied with the conclusions drawn, continues with further inquiries.
Mintbear, not satisfied with the conclusions drawn, continues with further inquiries.
Step 5. Secondary in-depth research
We gave the answer a light evaluation and requested a second in-depth research.
The general evaluation online may be as above, and the evaluation or criteria for each tool are agreed to be very good.
However, in Mintbear's personal opinion, which is checking the recent real-time changes, it is better than Ideogram, Leonardo, and DreamStudio.
Krea, Recraft, Dreamina, Freepik, and video creation tools, but I also think image models like KOLORS 1.5 supported by Kling are useful.
Maintain the basic evaluation above, but reflect Mintbear's current opinion, search and review again, and then draw a conclusion based on that evidence. (In-depth research)
Step 6. Ask additional questions to O3.
When I asked for [in-depth research], I received follow-up questions this time as well.
ChatGPT o3: Thank you for your good opinion. Reflecting Mintbear's latest opinion, instead of Ideogram, Leonardo, and DreamStudio,
We take a deep dive into AI image generation tools, including Krea, Recraft, Dreamina, Freepik, and Kling’s KOLORS 1.5 model.
We take a deep dive into AI image generation tools, including Krea, Recraft, Dreamina, Freepik, and Kling’s KOLORS 1.5 model.
Are there any key considerations that you would like to focus on in this study? For example: Emphasizing specific features (e.g., style control, resolution, editing capabilities) Comparing paid vs. free models Testing image creation speed and quality Analyzing recent updates and technological changes Including these points in your analysis will allow you to draw more specific conclusions.
The attitude of checking the user's intention and specific information,
It was like looking into the eyes of a really smart and passionate new employee. 😺
It was like looking into the eyes of a really smart and passionate new employee. 😺
The questions were very specific and seemed to hit the nail on the head thanks to previous in-depth research.
I was happy to provide feedback and guidance.
For user areas of expertise, the quality of the conversation improves as you can provide specific categories.
We will have to do a similar integrated evaluation as before. Instead, the mentioned style and resolution are not a big issue, and it is very important to provide editing functions for the generated images.
In the case of paid/free, the utility and value are evaluated accordingly, and the speed of image generation is rarely an issue these days. So, rather than the speed of image generation, you should comprehensively evaluate how useful the generated image is to the user by evaluating the aesthetic beauty, realistic or factual image presentation, and responsiveness to prompts in terms of quality.
The last recent update is likely to be a criterion for selection, as it is continuing at a rapid pace and has recently been updated with useful features. I hope you will proceed by setting this complex criterion well. Thank you.
Step 7. Secondary search and inference
After that, very effective analysis and inference continued. (Inference summary omitted)
Results and Lessons Learned
O3 & Deep Research were very powerful.
Not the level of reasoning experienced at level 4o, o1
O3 had its own unique level of differentiation,
O3 had its own unique level of differentiation,
We have identified a style of powerful self-directed learning AI that has reached the agent level.
Definitely a colleague I would like to work with.
Definitely a colleague I would like to work with.
Sometimes there may be limitations due to reference issues, etc.
Given clear guidance from the user, we've implemented some really powerful inferences and approaches.
Given clear guidance from the user, we've implemented some really powerful inferences and approaches.
It has continued to be used on other topics since then.
I am having a very amazing experience in more specific and professional discussions rather than superficial ones.
I am having a very amazing experience in more specific and professional discussions rather than superficial ones.
Going beyond o1, which showed limitations due to various functional restrictions including the past search function
Now o3 feels like he's freed from the chains that bound him, and is running out into the world.
Now o3 feels like he's freed from the chains that bound him, and is running out into the world.
It was also a valuable experience to be able to share valuable insights about current image generation tools.
2025.02.06 OpenAI WorkFlow
AI Visual Director, Mintbear
