I would love to see a text diff in the UI.
2 Likes
Your program is perfect but can I have a list of all the things that it can do? Thanks
My text was basic and my instructions was to “Fix grammar and add business lingo”
the output was the same paragraph repeated many times.
1 Like
We can’t have the same parameters for Insert, we need to break down the TopN and replace it with some Gaussian smoothing function which covalently parses the motivation behind the MOTIF. Same with temperature. Why not have Frozen, Cold, Hot, Null, and LogLogontology in the form of a Bayesian inference exercise. It seems very brute force, and takes in a whole corpus, which is demanding
It would be convenient to have a similar type of “use as input” button for the playground’s insert section which will insert the result into the content in the input box.
It would also be convenient in addition to that if you could use Ctrl+A on the result box without it selecting everything on the entire page and only the content in the result box so that it could be easily selected and copied either to the input box or elsewhere.
Python is completely dependent on indentation, the rest of the lines after the first aren’t interpreted as being part of the reducetext function the way you have it there so why would it add a colon?
Like I’ve seen said elsewhere: You need to still have to know how write code in the first place.
A recommendation for edits: highlight text that the AI editted.
1 Like
I am trying some code completions (insertions) for python code. You may check the attached image which pinpoints the problem.
When asking for an insertions sometimes the model replies with a possible solution given the context, but then continues to return tokens for additional problems by also providing comments that describe the additional problems. In the specific example we ask an insertion in order to get a reply if user can login depending on his username/password, but then we get additional function for “register the user if the username is not taken”.
I can somehow overcome this problem by providing specific stop strings but this is highly dependend to the context.
Is this a prefered behavior of the model? Is there any additional suggestions for overcoming that problem?
Thank you
I think it would be nice to have colours indicating when something has been added, replaced or deleted. Sometimes I have to re-read a generated edit multiple times to finally realise nothing changed.
2 Likes
Some feedback: The token amount is extremely limited for edits. You can only apply edits to trivial files while the ones most tedious to edit does not fit into the limits. Also since the instructions are counted in the tokens, precise instructions reduces the value got out of the edits since it can only edit even more trivial files. It’s a bit counterintuitive since the bigger the file, the more you likely need instructions.
Maybe the endpoint should be able to stream through the file so that it stays within the token limits?
Would be important that there were a structured way to give example inputs & outputs. Now it’s up to guess work and tricks how it should be presented in the optimal manner.
Also looks like if the editable content is a bit longer, like 200+ lines, the Codex is incapable of doing anything meaningful to it. It might add an empty line or change one function invocation, but that’s about it. Feels like the longer the editable code, the less capable Codex becomes. This makes it hard to run any instructions in scale.
1 Like
I posted a separate thread about this yesterday before I found this thread. Is there currently a bug with the two-pane editors, Insert and Edit? I’ve tested on several devices and they all appear to be bugged, replacing the prompt text automatically with the result text. Is this change of behavior intended?
2 Likes
Yea definitely, I’m currently having this strange behavior in edit, where it’s completely replacing the input with the output and not just that it’s like appending it or something…
2 Likes
As someone who is a native english speaker, If someone asked me to “reverse the spelling” I honestly wouldn’t have a clue what they’re expecting me to do. That makes no sense at all, try being more specific like
reversing the order or the letters/words?
reversing the capitalization?
replacing the words with their best antonym?
I ask it to do all kinds of things. I test out so much that i have a bunch of saved personal data sets. Hows your experience with other tests?
I’d love the ability to have a mode kind of like Insert, except it spits out text that is meant to “summarize” the code. You could use that to synthesize large diffs and/or generate PR descriptions, or even just to create general tech documentation
1 Like
The tools works amazingly well, helping to expand on existing ideas
Would be awesome to add a button that copies the response
1 Like
