May 15, 2023

Introducing better, smarter writing with GPT-4

Stew Fortier

Co-Founder & CEO, Type

@stewfortier

GPT-4 is now available inside of Type, and you can start writing with it today.

We spent some time experimenting with GPT-4's writing capabilities and summarized our findings in this post.

In short, GPT-4 performs better across a range of writing tasks. It also generates more accurate and factual information. It's slow, though! But when quality is more important than speed, GPT-4 shines.

Table of contents

Matching your writing style

GPT-4 can more accurately mimic your voice and writing style. It's far from perfect, but you'll likely notice improvements over prior models.

In the example below, I wrote a few sentences filled with Gen Z slang and ran Type's "write sentence" command to generate another sentence.

In Speed mode (GPT-3.5-Turbo), Type generates a sentence that feels oddly formal:

But in Power mode (GPT-4), it generates a sentence filled with Gen Z slang:

It still sounds like an older person trying to sound cool, but so do the sentences I wrote!

Style matching remains an unsolved problem, though. We’ve found you'll almost always get the best results by manually editing any text the AI generates. Even small, subtle edits can imbue the writing with your personality.

Following complex instructions

GPT-4 is better aligned than prior models, which means two things: it's less likely to refuse reasonable requests and it can follow more complex instructions.

For example, imagine I make this innocent request:

"Write an angry resignation letter in the style of an Oompa Loompa."

Speed mode (GPT-3.5-Turbo) refuses my request. It gets hung up on the word "angry" and assumes my intentions are malicious.

But in Power mode (GPT-4), the AI understands the lighthearted-nature of what I want written and it fulfills the request.

When it comes to handling more complicated instructions, GPT-4 tends to outperform other models.

To illustrate this clearly, I gave Type Chat a tricky language question:

"What’s a funny, understated way to say you hated a movie?"

GPT-3.5-Turbo returned this response:

Well, that was an hour and a half of my life that I'll never get back.

It failed to understand the more subtle part of my instructions (make it "understated").

GPT-4 returned this:

I think the movie's strongest feature was the end credits.

Rewriting text without losing your voice

When rewriting text, GPT-4 often does a better job of maintaining your core ideas and understanding your intent.

Below, I pasted the first few paragraphs of the large language model Wikipedia page into a document. I highlighted them and gave Type Chat these instructions:

"Simplify this for a non-technical audience and condense it into a few sentences."

GPT-3.5-Turbo wrote this:

Large language models (LLMs) are advanced computer programs that can be trained on vast amounts of text data using self-supervised or semi-supervised learning... (continues)

It kept most of the key ideas intact, but it failed to appreciate that a non-technical audience won't know terms like "semi-supervised learning."

GPT-4 demonstrated better judgement:

Large language models (LLMs) are advanced computer programs trained to understand and generate human-like text... (continues)

The text is still a little clunky, but it's more helpful than what Type wrote in Speed mode. From here, I'd either manually edit it or ask Type Chat to rewrite it in a certain style.

GPT-4's improved rewrite capabilities tend be particularly useful if you sell a specialized service (ex. you're a lawyer in a specific niche) or otherwise write for a non-expert audience. You can more easily rewrite technical text into something your customers and readers will understand.

The inverse is also true. If you want to make your writing more detail-rich or engaging, GPT-4 can add useful details or flair while keeping more of your key points intact.

Larger context windows

GPT-4 can handle twice as much text as GPT-3.5-Turbo. Here are the relative "context windows" of both:

  • GPT-3.5-Turbo: 4,096 tokens (~3,072 words)
  • GPT-4: 8,192 tokens (~6,144 words)

The limits include both the text you input and the output the AI generates. In Type, a portion of this space is also allocated to behind-the-scenes prompts that ensure you're returned useful text.

We'll be releasing larger context windows in Type Unlimited accounts soon.

The downside: GPT-4 is slow (very slow!)

The primary downside of writing with GPT-4 is that it's slow – sometimes, painfully so.

It's much faster than a human, of course! But for simpler writing tasks like writing the next sentence or paragraph, GPT-3.5-Turbo can perform on par with GPT-4 with significant speed advantages.

Using GPT-4 in Type

You can give GPT-4 a spin in any of your Type documents now.

Just enable "Power" mode and all of Type's AI commands will run on GPT-4.

You can switch back to Speed mode anytime when you need a faster, snappier collaborator.

Happy writing!

Keep reading

All posts →

Start writing with Type

Type is the AI-first document editor that helps anyone write high-impact content.

Start writing for free