
I pitted DeepSeek and ChatGPT against each to see which app is best at helping people brainstorm, think critically, and plan. This kind of thought partnership is one of the greatest values of AI. Let's see which of the two platforms provided the most value!
Table of Contents
The Task
I asked both DeepSeek and ChatGPT to help me come up with new topics for posts on the Elevate AI Coaching Blog that were likely to engage readers.
To maximize thought partnership, I used the "conversation sculpting" method popularized by Chris Lele, Elevate AI Coaching's CEO. Instead of trying to craft and execute a single "perfect prompt," I conducted a longer "prompt-versation," taking an exploratory, open-minded approach as I partnered with ChatGPT and DeepSeek to optimize the EAIC blog.
The Models I Used
ChatGPT 4o
For my work with ChatGPT, I prompted using version 4o. I did this because 4o has a unique advantage over DeepSeek: it's multimodal, able to analyze photos as well as text. DeepSeek, on the other hand, can only work with text uploads.
DeepSeek R1 (web & offline)
For DeepSeek, I used its new R1 model in two versions: the web-based version and the downloadable version that can run locally on a PC. DeepSeek’s unique strength lies in its ability to operate offline. My plan was to use the local version if I encountered access issues with the web-based version—and boy, did I encounter some issues!
(Note: I chose to run the offline version of DeepSeek in a "novice-friendly" way that requires Ollama and Chatbox. See this Reddit thread for instructions on the approach I took.)
DeepSeek Web is SLOW right now!

At the moment, DeepSeek's web app appears to be a victim of its own sudden fame. It’s not just getting coverage on the usual AI news sites like TechRadar and Tom's Hardware—it's also being written up in The Atlantic and The New York Times while trending on Al Jazeera English and BBC World News. This surge in interest has caused DeepSeek’s web app to run very slowly.
While I did have to repeatedly hit "regenerate" to bypass error messages about busy servers, the actual results were worth the wait once they finally arrived. Let’s take a closer look.
My Conversation with DeepSeek Web
I opened the conversation by priming DeepSeek to think carefully about the topic at hand. I’ve already discussed priming in my article about LLM hallucinations. To recap from a recent article of mine: priming involves getting the LLM to do some general thinking about a topic before you prompt it on the actual task.
In this case, I primed DeepSeek by providing links to the two most visited posts on our blog at the time of writing—AI Hallucinations vs. BS and OpenAI Strawberry: What We Know So Far. I asked DeepSeek to analyze why those two posts were the most popular.

The Journey
DeepSeek Web was an excellent thought partner as the conversation unfolded. It surfaced aspects of my writing that I’m not usually conscious of, such as my use of storytelling, visual elements, multimedia, and SEO structure. DeepSeek called out these elements with specific examples, helping me clarify and expand on both my thoughts and its suggestions.
It also proposed some useful best practices for writing blog posts targeted at Elevate AI’s audience. Notably, it included multiple "Why these ideas work" sections where it defended and evangelized its suggestions.
Ultimately, DeepSeek helped me identify two main categories of blog posts that I can focus on to replicate the success of the EAIC blog’s most popular posts.
The Final Outputs
The goal here was to come up with a truly great list of topics for future Elevate AI Coaching Blog posts. One cool thing here is that I didn't have to explicitly ask DeepSeek for that output—as the conversation progressed, it was able to infer my intent and offer me blog post topics unprompted.
The first three topics Deep Seek Web suggested were:
Aside from the dated "Strawberry" wording in the first suggestion, these are all excellent, intriguing pitches. From there, however, DeepSeek Web's blog post suggestions were a mixed bag. There were some some really bland ideas, such as "The Art of Prompt Engineering." Other ideas were simply repeats of the sample posts I'd given it—"AI Hallucinations: Why They Happen and How to Fix Them" is a particularly egregious example.
Grade:
B+
This could have been an A-, as the sophistication and exemplary thought partnership DeepSeek showed was so valuable that it largely made up for the mediocre final outputs. However, I dinged the Web version of DeepSeek a few points due to its slowness.
My Conversation with Locally Run DeepSeek
I had originally planned to only use this model as a backup in case DeepSeek Web got so slow it stopped working entirely. But while I waited for DeepSeek Web's responses to load, I had to do something to kill that time. So I concurrently kicked off the same prompt-versation in the offline version of DeepSeek R1.
The Journey
I opened up with the exact same prompt I'd given DeepSeek web. I said:
I am the editor for the Elevate AI Coaching blog. Elevate AI Coaching is a company that helps businesses leverage platforms such as ChatGPT, Claude, and DeepSeek into their business operations. The homepage for Elevate AI Coaching is here: https://www.elevateaicoaching.com/ . And the page for our blog is here: https://www.elevateaicoaching.com/blog
The two attached PDFs show our most popular blog posts to date. Please give me your thoughts on what those posts have in common, and how I can replicate the success of those posts in future updates to our blog.
And of course, I attached the abovementioned PDFs.
What a mistake! DeepSeek R1's offline incarnation is really slow and clumsy at evaluating attached files. I had to slowly, agonizingly coax the locally run DeepSeek into truly analyzing my blog posts rather than just summarizing them. I also had to frequently remind DeepSeek about what we had been talking about because it kept full-on forgetting. Just as annoyingly, I had to reupload the PDFs multiple times, because it seemed to forget that it had been given them.
Deep Seek Local's advice was overly simple, often inadequate, and occasionally off-topic. And it was slow too, but for different reasons than Deep Seek Web. Deep Seek Local writes out the thought processes that lead to its output. While this is an interesting look at the "black box" part of LLM response that's invisible in most apps, it also slows down the whole prompt-versation.
I decided to experiment with non-PDF sources to see if that would help. I gave the offline version of R1 a URL of the EAIC Blog's third most-popular article, our write-up on ChatGPT 5.0. It still struggled to analyze URL, although it was clearly able to access it. Finally I copy-pasted the text of our 4th most-read blog post, which compares the use of AI in Star Trek to real-life prompt engineering. DeepSeek Local is definitely faster and more confident if you copy-paste context right into the window.
The Final Outputs:
I was pleasantly shocked at the end. DeepSeek Local seemed like a dumbed-down version of DeepSeek Web, but its suggestions for blog post topics were undeniably smart. Here are some samples:
Unfortunately, the rest of DeepSeek Local's suggestions, while also smart, were off-topic. The rest of its pitches focused on AI applications that had nothing to do with LLMs, such as analytical AI in medicine, and AI automations in fintech.

Bonus Feature:
DeepSeek Local came through as a backup to DeepSeek Web after all! At one point, while I was jumping back and forth between the Local and Web threads, DeepSeek Web slowed down so much that I was worried I wouldn't be able to get the intermediate output I was looking for. At that point in the conversation, I wanted to identify two broad categories of blog post.
So I copy-pasted the entire DeepSeek Web thread into Deep Seek Local, and had Local pick up where Web had left off. Deep Seek Local gave me a comparable (but far less detailed) response to the one I eventually got from the web version.
Grade:
D
Although DeepSeek Local is better than nothing when I use it as a backup, it's still slow, and weak in all of its operations. It gave a handful of fantastic final outputs, but I could have gotten outputs that good with far less effort in other apps.
My Conversation with ChatGPT
ChatGPT got the same initial "priming" discussion question as DeepSeek Web & Local. I primed OpenAI's LLM by inviting it to look at Elevate AI's two top blog posts and analyze the reasons for their popularity.
The Journey
ChatGPT had ideas that were fairly similar to the ones from DeepSeek Web. But its approach was a lot less nuanced. It identified the importance of key writing skills such as storytelling, just as DeepSeek did. But it did far less to give examples of why those things are valuable, or how the writing in the two most popular blog posts drove those important points home.
In addition, ChatGPT did very little to defend and evangelize its perspectives. Instead, it had a very lightweight recurring series of analyses of its observations labeled "What it suggests."
The Final Outputs:
ChatGPT had the best mix of solid blog post ideas. Here are just a few of the great pitches it tossed my way:
GRADE:
B-
I'm grading ChatGPT on a curve, relative to how well DeepSeek Web did.
On the one hand, DeepSeek's overall observations and initiative were amazing. This provided a value that outweighs the ChatGPT prompt-versation experience, even though ChatGPT had some strong final outputs. On that basis alone, ChatGPT might have gotten a C.
But we also have to consider the fact that-- for now at least-- ChatGPT gives answers quickly, without server issues or loading errors. And for that, I can bring ChatGPT up to a B.
In a different use case, where the ability to comprehend images matters more, ChatGPT's multimodal nature might have brought it all the way up to a B+, relative to DeepSeek.
Final Thoughts
While DeepSeek Web wins this round, it's important to remember that the online version of DeepSeek R1 only proved the strongest for this exact use case. Chris Lele, has identified another use case-- AI-written articles-- where DeepSeek is not the clear winner. Check out Chris's Medium article for details.
You are also free to judge the outputs for yourself. Access the full threads from my experiments below:

Comments