news

when i woke up early in the morning, without any warning, i saw that openai had launched a new feature: canvas.

the layout of left and right columns reminds me of claude, but it is completely different from claude's artifacts. there is no preview function and more local modifications that are more in line with the needs of human-machine collaboration.

after using it overall, my feeling is: canvas has taken a praiseworthy step in the direction of human-machine collaboration, but it is only a small step.

let me illustrate with an example.
(the text in the example is not my own. it is just a simple example that has no cognitive load and makes it easier for everyone to focus on understanding the canvas function. ）

editing based on ai generated content – ​​very useful

i took a few photos and asked ai to write a circle of friends for me.

it helped me generate the text below, which accurately captures the essence of the photo,references to books and the cat under the chair。

the three paragraphs are quite clear from summary to details to opinions. however, this writing style is too unfriendly to the circle of friends. i tried to change it directly, but the effect was average. return it, modify the first paragraph on the original text, and then ask it to help me modify the remaining two paragraphs.

knock on the blackboard: modify directly on the original text, very convenient!

it said it was fixed. when i looked at it, why was it still a bunch of text?

ask it and it appears innocent.

then i realized that it was actually capable of forming single sentences into paragraphs, but there was a bug in the canvas display, and some line breaks were not recognized. so i manually changed the lines and deleted three long-winded words. let's compare it with the original version, and the improvement effect is very obvious.

so far so good. being able to edit on ai-generated content makes me feel really good.

knock on the blackboard: examples matter. most of the improvements were made by ai based on the example in my first paragraph. in addition, don't compete with ai, it is helpless when there are bugs in the code.

the general modification function is not suitable for personalized needs - i can’t complain

next try the suggest edits function. i fell apart as soon as i tried it.
it gives a lot of advice in english. hello, what does a chinese article mean when you give me advice in english?

i immediately knew i couldn't expect anything more from suggest edits -i can’t even personalize it in chinese and english. it’s impossible to talk about my preference for words and what kind of values ​​or philosophy of life i hope to be reflected behind the words.
i thoughtlessly clicked apply to accept all the suggestions, and sure enough i got a sad result.

i expressed my dissatisfaction to it. it has iterated one edition.

i gave further suggestions and told it that some sentences were "too pretentious", and it improved it a few times. but, after so much effort, it’s still a cliché.

in fact, it is not surprising to get such a bad result. because after the suggestions edit button is clicked, it only uses the user role to send a generic and shockingly poor quality prompt:

in my imagination, this functionat least the direction of modification must be determined based on the entire previous conversation record. if memory is turned on, it should also generate more personalized suggestions for improvements.
and for general improvements like this, the suggestions you get can only be rubbish.

knock on the blackboard: want to improve your article? select text and make a personalized modification request via askchatgpt. don't rely on the suggest function that comes with the product.

add emoji - such a simple function is actually not done well

i have no choice but to go back to the previous version (by the way, the undo/redo function in the upper right corner of the canvas is very useful).

instead of using suggestions edit, try a simple function: let ai help add emoji to text. (after all, even ultraman is asking everyone to vote.)

but in fact, this function is not very good: the generated content seems to have been replaced based on the vocabulary list, which feels very stiff.

i checked it with the developer tools. no wonder...it actually just issued such a simple prompt word in the role of user:

i still have to patiently train myself:

the results obtained were much better. 🤯🧘‍♀️⏰✨🧡🛋️ the use of emojis is very appropriate and something i wouldn’t have thought of myself.

let's change it to several emojis and use them together to make it more interesting.

such a small example actually fully illustrates the need for personalization - the large language model is a weighted average of human intelligence. without any adjustment, the results it gives in line with the average preference are usually mediocre results.human-computer interaction is required, users are required to show their own aesthetics, and ai is required to capture and apply it in order to obtain results that are more in line with user expectations.

we are still far away from the ultimate vision of human-machine collaboration

i originally wanted to evaluate the encoding function. but when i glanced at the function keys in the lower right corner, i saw that there were only toy-level functions such as adding annotations and logs (and the functions were implemented using prompt words from the caotai team), and i immediately lost interest. friends in the ai ​​programming group asking about the comparison between canvas and cursor, just use cursor honestly. when you see big models that no longer use snake or tetris as examples, it’s never too late to pay attention again.

canvas allows users to make modifications based on ai-generated content, which is great. but the button in the lower right corner is obviously too superficial at present——it should at least first generate personalized prompts based on the conversation history or even the user background in memory., instead of using rough instructions written down by the grass-roots team.

even,i hope that the entire canvas can dynamically generate different buttons based on the current state., instead of always having toy-level buttons to add emojis, adjust length and language difficulty, etc. occupying the space.

i've been looking forward to the ability to dynamically generate buttons for over a year. last year in ren xin’s podcast we imagined the future of ps:

i said: my hair doesn’t feel right, i need to adjust it.
a bunch of controls immediately appeared next to the hair: change the color, change the volume, change the degree of elegance...
i clicked on the "change the flowiness" slider until my hair was raised at an angle i felt comfortable with.
modification ends.
this will be a perfect combination of lui and gui.

it’s unrealistic to expect ps to follow his words and immediately change the angle of raising my hair to the way i expected after i said my hair didn’t feel right——a thousand hamlets, big models as averages not sure which one i want.

the tweet from the creator of canvas also seems to be talking about such a future - a blank canvas that evolves with human preferences.

but, this is still very far away.

even, i'm not sure whether openai will consider "giving users more opportunities to provide information."

because even o1 is currently not able to interact with people. itwith obsessed self-confidence, just want to complete the task by bootstrapping。

i often look at o1, this stupid young man thinking all over the place and spitting out a bunch of incomprehensible words, and i just want to shout - what are you thinking about? why don't you stop and ask me?

many gpts that my friends and i have written before will sort out the questions before generating answers, proactively ask users questions, and then start working after collecting enough information. the effect will be much better than making blind guesses and drawing blind boxes.

i don't know if there is any special reason why o1 doesn't do this. is it because you want ai to do things that humans cannot direct it to do?

questions such as math questions that have exact methods and answers and can be verified by themselves and get feedback may indeed not require humans to do anything (most people, including me, really don’t have the ability to express any opinions on advanced mathematics). however, if it is designed to collect information, stop and ask the user first, whether it can also perform more humanistic tasks much better?

after all, this world is much richer and more exciting than large models.