ChatGPT o1-Preview: Great for Content Planning

(Disclaimer: Profanity.)

About a month ago, the nerds at OpenAI dropped their strawberry model called the o1-preview and claimed it could think. Of course, it was a lie to woo the investors and give fodder to the AI influencers who create surface-level content.

In this chapter, let’s debunk that myth by understanding how it works and looking at its applications in the content production workflow.

How the o1-preview model works

Think of talking to multiple GPTs at once.

That’s all.

Let’s consider an example to understand it deeply. 

Imagine you are writing an email to your boss to ask for a week off from work because you have to travel to attend a family function. The message should be professional and concise while containing all the relevant details.

GPT-4o will immediately spit out an email for you. It may miss a few of your instructions though. 

For instance, the email could lack the brevity that is expected in official emails or use flamboyant words or unnecessary literary ornaments. 

(Note: This is a hypothetical example. GPT-4o can do this accurately and easily. I chose this to explain the inner workings of the ChatGPT o1-preview model in a simple way. Please replace ‘writing an email’ with something complicated like ‘create a product roadmap.’)

However, with o1-preview, the chances of the LLM ignoring your instructions or following them in the wrong order decreases significantly.

As mentioned earlier, it is like talking with multiple GPTs.

One GPT interprets your command, another writes a raw draft, and yet another reviews the draft to ensure it adheres to your instructions. There could be even one more GPT that rectifies the produced draft before it’s returned to you. 

How the o1-preview works

ChatGPT o1-preview in content writing

It is crucial to realize that there are no new capabilities. The nerds at OpenAI have found a way for GPT-4o to recursively evaluate its outputs before printing it out. However, this approach does solve a problem with ChatGPT’s flagship model — ignoring instructions.

Problems with multiple parameters (like creating a product roadmap) are seldom done well by GPT-4o. Heck, even while writing content for a specific niche for an audience with peculiar preferences, it stumbles.

Common issues include ignoring instructions altogether, misunderstanding expectations, hallucinating, and following commands in the incorrect order. Consequently, tasks that require you to evaluate multiple parameters are done poorly.

As a professional content marketer, I found this limiting when generating publishing calendars or article suggestions for a particular campaign. The machine simply struggled to analyze all the data I threw at it which included the brand guidelines, campaign objectives, and audience’s needs.

The ChatGPT o1-preview model can analyze lots of commands and deliver an output far more useful than the bare GPT-4o model. I’ve been leveraging it to validate my marketing ideas, get campaign ideas, and more.

Content-specific tasks include extracting details from a large blurb of text, writing promotional posts for a particular article, and editing specific paragraphs or sentences in a draft. Each of these action items requires GPT to review its output and iterate it until perfection.

Limitations of ChatGPT o1-preview

After the o1-preview model was released, folks across the internet hyped it up because emotionally charged content on trending topics delivers great engagement. For a moment, it appeared as if OpenAI reinvented the wheel and everyone in the LLM industry is fucked.

Well, that’s not the case.

While the o1-preview model is a great step forward (depending on how you look at it), there are still some noticeable limitations.

For starters, you can’t upload images or any kind of files for deeper analysis. I frequently used this feature to extract context-specific information in GPT-4o. Perhaps the o1-preview model could do it better. Unfortunately, we will never know. At least for the foreseeable future.

Moreover, it can’t read web links, which means you have to manually provide the latest information, along with the data from the files that could have been uploaded. The image generation capabilities are disabled too.

If you have a bunch of custom GPTs, like I do for my content production process, you can’t use any of them. This further increases the amount of prompting you do.

Finally, the model performs quite slowly, making it apt for high-level ideation-type tasks only. Typing a long prompt and waiting for seconds have to be worth it! 

With all of the advantages and disadvantages of the o1-preview model, it is certain that the upgrade is helpful in only a handful of use cases. For the vast majority of content writing applications, GPT-4o is still the best. Honestly, it’s been a while since I used the o1-preview model and I don’t miss it.

Conclusion: o1-preview is for special cases only

To test whether o1-preview was better than GPT-4o, I ran a bunch of common prompts for both and the outputs from the former were only arguably better. The margin of improvement, especially for general or slightly complicated problems, is quite small.

A general rule of thumb is to use the o1-preview model for problems that have too many instructions or details to consider. Try it out with GPT-4o first and only go to the strawberry model if the output is useless.

Keep in mind that the o1-preview model is still not free from hallucinations. It does spit out garbage on occasion, which I guess, is a feature of AI models anyway. Hopefully, with time, the model will get faster and grow in capabilities.

—

Next Chapter: ChatGPT Canvas and Claude Artifact: Good Assistants

Previous Chapter: Should You Publish AI-Generated Content

Index (with Prologue): Content Writing With AI