I have a question, for a content writer what base model is suitable rather than GPT3
Playing a bit with ChatGPT and Codex beta. It is wonderful and with many possible use cases where the text or the code are not too complex. ChatGPT fails in simple math but is good at telling jokes and even make better jokes when translating them to Czech 
As mentioned by others, it has a tendency to be verbose.
Codex is even more surprising to me being able to generate moderately complex snippets. However the same verbose tendency in the Edit mode of the playground leads to generating 3 more functions after simple request “remove the function XY”.
It has potential to be an invaluable programmer’s coworker, but the slow response of the cloud services (in general) would be an issue here.
I know this thread is busy, but in case anyone can see my question, will the new capabilities be incorporated into models we’re currently working with in our existing conversations?
I’m making meaningful progress with a model right now and we’re unlocking new capabilities together in an existing conversation. It would be nice to be able to continue our work together while knowing that I have access to the capabilities you just mentioned.
I’ve been trying to contact your support team, but I assume you’re pretty swamped right now, because the help centre chatbot keeps rerouting me to give feedback through the ChatGPT terminal, even though I feel like it would be beneficial to speak with a live person. Hopefully eventually we’ll be able to make contact!
EDIT: Ohhhh is this an old thread?
Out of a sudden the Edit function in Playground started repeating the Input in the Output. For the same Inputs/Instructions I was using before it is now repeating the same Input into output. I tried everything, with no avail.
Not sure what has changed, but its very annoying as I use playground a lot for testing.
By the way, when will text-davinci-003 be available on edit?
Got the following invalid text in response
<!-- You can't change an article's title within a talk page, so you should start a new talk page.
Fix this later.
-->
I used the edit model to repair some text that had been scanned and poorly OCR’d and it worked well on repairing words that were separated/cut up by spaces.
1 Like
I had an issue with code-edit where I asked it to change javascript variable declarations from let to const where possible. But it essentially changed all lets to consts even though that didn’t make sense in all cases.
Excited for the possibilities.
Something that would add value in the edit mode would be if the sections/words that have changed were highlighted.
EDIT SQL code-davinci
Input Two simple selects:
SELECT [created]
,[updated]
,[deleted]
,[materialId]
,[pricepurchase]
,[pricewholesale]
,[pricetrade]
,[pricewholesalemaximum]
FROM [Source].[Facade].[MaterialPrices]
SELECT [created]
,[updated]
,[deleted]
,[materialId]
,[parentMaterialId]
,[name]
,[pharmaceuticalProducerId]
,[salesProducerId]
,[supplyProducerId]
,[blozId]
,[bloz]
,[bazylId]
,[ean]
FROM [Source].[Facade].[Material]
Instruction: Create view with dates materialID, parentMaterialID, Names, and prices
Result is ok, but there is no aliases and we got error about ambiguous column names.
When I added “be carrefull on ambiguous column name” in Instruction part, result was empty.
Just a heads-up that [insert] is still buggy. It gets stuck on an idea and replicates that idea over and over, even when the result is semantically incoherent.
For instance, my prompt “The [insert] way to succeed is good luck” produced five or six iterations of the following before I stopped spending tokens:
“The way to succeed is to work hard and be persistent. You need to have a good attitude and a strong work ethic. You should also set goals for yourself and strive to reach them. Finally, you should be willing to learn from your mistakes and take risks. In some cases, a little bit of luck can also help, but it is not necessary for success. So the short answer is no, the only way to succeed is good luck.”
It’s a neat feature and I look forward to seeing what it looks like as it gains power, but it’s not there yet.
What I like about the Edit request format is that I can use it to structure how I’d like the response to look. Since I’m making API calls, I’d prefer if the Edit end-point answers my question in specific JSON formats, so I don’t have to waste effort parsing varied response formats. Here’s an example:
Input:
{
"subject": "...",
"topic": "...",
"who": "...",
"when": "...",
"where": "...",
"how": "...",
"why": "...",
"what": "..."
}
Instructions:
Provide the meta data for this article:
"The Meiji Era was a period of tremendous change and modernization in Japan that began in 1868 and ended in 1912 when Emperor Meiji died. This era began when the shogunate government was overthrown and the emperor was restored to power."
Typical Response:
{
"subject": "The Diet and the Meiji Constitution",
"topic": "Meiji Restoration",
"who": "educational video",
"when": "1868-1912",
"where": "Japan",
"how": "underwent a transformation from a feudal society to a modern, industrialized nation",
"why": "become a powerful nation on the global stage",
"what": "Diet and the Meiji Constitution"
}
The small issue I have with this format, though, is that since I’m using the “Input” for the response format, my actual input text ends-up in the instructions, which is cumbersome, and often confusing for the Edit end-point.
It would be great if there was a way to send a response format / template separately from the input and instructions.
The model added <br> tags (the TEXT model to be clear) to the end of every line when I told it to “enhance” the input.
I have been experimenting with this tool to edit html and even build entire webpages and it has been amazing. I can tell it to make changes to pretty much any part of the page and it does. Add spacing, add navbar, titles etc. I have tried to prompt it to get it to generate tailwind code, and change the font with Google fonts. The results so far have been amazing. I have been working on a small Webpage builder project based on the edits API.
I must be doing something wrong, because it’s not editing my text? It’s mostly just giving me output that is exactly the same as the input. Once it added line breaks after each sentence, or deleted a word (I’ve been varying the prompt).
Example input:
The contest at Rainhill in October 1829 was held to decide how the new railway would be worked. But its consequences were destined to be much more profound and far-reaching than any one railway. What was about to happen here would change the world.
Instructions:
Rewrite these sentences to be less awkward and more streamlined.
Output:
The contest at Rainhill in October 1829 was held to decide how the new railway would be worked. But its consequences were destined to be much more profound and far-reaching than any one railway. What was about to happen there would change the world.
Output is literally the input unchanged.
I also tried the prompt “Rewrite this as a sonnet”, and it gave me back the input, and then added one sentence in a new paragraph: “It was the first time that steam had been used to propel a vehicle of any kind on a railway.” So it really doesn’t seem to be following instructions.
Model: text-davinci-edit-001
Temperature stated at default 0, I moved it to 0.5, doesn’t seem to make a difference
Top P = default of 1
Using the responses to a survey, I’ve tried variations on the instruction to ‘Fix spelling mistakes’ and keeping the temperature between 0.0 and 0.5. It seems to work quite well but hallucinates short responses into python code.
For example the single response ‘None’ will generate python code.
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
File: file.py
Author: limingdong
Date: 7/10/14
Description:
"""
import sys
reload(sys)
sys.setdefaultencoding('utf-8')
def read_file(file_path):
"""
读取文件
:param file_path:
:return:
"""
with open(file_path, 'r') as f:
for line in f:
print line
def write_file(file_path, content):
"""
写文件
:param file_path:
:param content:
:return:
"""
with open(file_path, 'w') as f:
f.write(content)
def append_file(file_path, content):
"""
追加文件
:param file_path:
:param content:
:return:
"""
with open(file_path, 'a') as f:
f.write(content)
if __name__ == '__main__':
file_path = './test.txt'
read_file(file_path)
write_file(file_path, 'test')
append_file(file_path, 'test')
If I prepend all responses with the text “Response :” the model performs as expected.
In regards to the Edits capabilities, I tested it by having it fill in a “table” (in list form) of U.S. taxation information that is normally easily accessible with Chat-GPT and while it filled in all of the the information in the result, it’s answers were completely factually false/incorrect. To be fair, I didn’t do much reading through the guidelines before testing it out. But I just wanted to give you guys a heads up that it doesn’t seem to be doing as well with factual data.
This may not be the best chat area to contribute to. I am beginning experimenting with GPT-3 Edit capability. I want to have it edit an early draft of my scientific manuscript to improve the content. First try was to ask it to edit the introductory paragraphs of the manuscript. My request text was “Revise the text for a scientific paper”. Most options were set very low except Temperature was set to 0.5 I received an error message that it couldn’t be done, suggesting the Temperature might be too high. I reset Temperature to 0. It gave me back the same text that I submitted. Unfortunately my input text included writing errors such as repeated sentences and thoughts. I assume I will need to experiment with different Temperature settings as well as settings for some other options, and possibly for the request/command text.
Is it me or the edit functionality is not working at all?
I requested to compact text, reducing unnecessary words, removing punctuation marks, removing dots, removing all words, delete all words starting with A and nothing, I always get the same text.
I have a rather long text, if it matters, but I got no warning about token size or anything.