Warning: ↳ Style tends to drift back to standard Even within a single passage, I found that it would start off with my style directions, but gradually revert to the more standard ChatGPT way of writing. More experimentation is needed here.
LLMS
-
Combining Multiple Styles in AI Text Models Reduces Quality
By
–
Warning: ↳ Too much style doesn’t work very well I tested combing different styles. This a common technique with AI art models to improve the results. But with text, too many instructions resulted in average content. (You can still add formatting instructions.)
-
Using Language Models to Emulate Writer Styles Effectively
By
–
🔸 Emulate a Writer
— Rob Lennon 🗯 | AI Whisperer (@thatroblennon) 13 janvier 2023
For writers who have large bodies of work, language models can emulate their style.
Instead of describing the voice and tone you want, try asking it to write as someone. pic.twitter.com/I0NhBHRhk1Emulate a Writer For writers who have large bodies of work, language models can emulate their style. Instead of describing the voice and tone you want, try asking it to write as someone.
-
Broken English Prompt Makes AI Writing Clear and Readable
By
–
🔸 Conversational English
— Rob Lennon 🗯 | AI Whisperer (@thatroblennon) 13 janvier 2023
I discovered this while trying to get the AI to consistently speak in broken English.
The result, is, perhaps ironically, very clear and easy to read.
Another good alternative to "write like a 5th grader" that's more specific. pic.twitter.com/3bMvX0SFmwConversational English I discovered this while trying to get the AI to consistently speak in broken English. The result, is, perhaps ironically, very clear and easy to read. Another good alternative to "write like a 5th grader" that's more specific.
-
Making AI Writing More Conversational and Relatable
By
–
🔸 Conversational + Relatable
— Rob Lennon 🗯 | AI Whisperer (@thatroblennon) 13 janvier 2023
AI writing can be dry and stuffy.
But it doesn’t have to be.
It knows perfectly well how to talk normally, that’s just not the default.
Here’s a voice/tone/style guide I love. pic.twitter.com/idfaOXebpuConversational + Relatable AI writing can be dry and stuffy. But it doesn’t have to be. It knows perfectly well how to talk normally, that’s just not the default. Here’s a voice/tone/style guide I love.
-
ChatGPT’s User Experience: Why It Feels Like Talking to Humans
By
–
I don't understand why people are so focused on the technical details of ChatGPT. The real value is in the user experience– it's the first chatbot that actually feels like you're talking to a real person.
-
GPT-4 Capabilities: Data Visualization Joke About Future AI
By
–
And for my data storytelling purists out there, this is just for scale comparison. I do not actually think GPT-4 will have the ability to eat hamburgers or run marathons. But GPT-5, however…
-
Correcting misconceptions about GPT model training data amounts
By
–
Also note that other threads have incorrectly stated that 100T is the amount of the data the models were trained on. According to Hyro AI, GPT-3 was trained on 570GB of data. And according to Sigmoid, GPT-2 was trained on 40GB. I have not seen a rumored amount of data for GPT-4.
-
GPT-4’s Rumored 571x Parameter Increase Compared to GPT-3
By
–
This thread is for entertainment purposes only and is based on the rumored 571x gain in parameters for GPT-4 (100T) from GPT-3 (175B). As a point of comparison, the leap from GPT-2 to GPT-3 param count was only 116x. ChatGPT is finetuned from a model in the GPT-3.5 series.
-
GPT-3 vs GPT-4 Parameter Count Calorie Comparison
By
–
You are GPT-3 parameter count and your friend is GPT-4’s. You eat one Big Mac. They eat the daily recommended calorie amount for five and a half months.