OpenAI Launches New 'Canvas' ChatGPT Interface Tailored To Writing and Coding Projects

OpenAI has introduced "canvas," a new interface for ChatGPT that provides a separate workspace for writing and coding projects. "Canvas is rolling out in beta to ChatGPT Plus and Teams users on Thursday, and Enterprise and Edu users next week," reports TechCrunch. "Once canvas is out of beta, OpenAI says it plans to offer the feature to free users as well." From the report: In our demo, [OpenAI product manager Daniel Levine] had to select "GPT-4o with canvas" from ChatGPT's model picker drop down window. However, OpenAI says canvas windows will just pop out when ChatGPT detects a separate workspace could be helpful, say for longer outputs or complex coding tasks. You can also just write "use canvas" to automatically open a project window. Levine showed TechCrunch how ChatGPT's new features could help write an email. Users can prompt ChatGPT to generate an email, which will then pop out in the canvas window. Then users can toggle a slider to adjust the length of the writing to be shorter or longer. You can also highlight specific sentences, and ask ChatGPT to make changes such as "make this sound friendlier," or add emojis. Users can also ask ChatGPT to rewrite the whole email as-is in another language.

The features for the coding canvas are slightly different. Levine prompted ChatGPT to create an API web server in Python, which spawned in the canvas window. By pressing an "add comments" button, ChatGPT will add in-line documentation to explain the code in plain English. Further, if you highlight a section of code that ChatGPT created, you can ask the chatbot to explain it to you, or ask questions about it. ChatGPT is also getting a new "review code" button, which will suggest specific edits for the code in the window, whether generated or user-written, for them to approve, edit themselves, or decline. If they press approve, ChatGPT will take a stab at fixing the bugs itself.

Many Hello Worlds Will Be Written

[thesjaakspoiler]

and an occasional DOOM for that Samsung microwave.

Re:

[Tony Isaac]

This.

We are a long, long way from an AI tool that can do more than write sample source code, that one must then fix and enhance.

With that said, I find that level of assistance to be highly valuable and a significant time-saver.

Re:

[Aristos Mazer]

Obviously we aren't there yet, but "long long"? Based on distance from ChatGPT 1.0 to now, I'll give us three years to get to advanced code. Is that long long in your measure? Code is much simpler than general language. With a clear goal for the LLM to work toward (resulting code either works or doesn't and can be fed test data quickly), with so large a corpus of non-trivial code available, and with structured language...I think getting LLM to produce high quality advanced code from fairly basic inputs is j

Re:

[Tony Isaac]

In my experience, the 80/20 rules applies here. We have 80% of the functionality that LLMs can produce. We can see it, almost taste it. It does some really cool stuff. But there are a million edge cases that have to be addressed. So we've spent 20% of the time it will take to get to that lofty goal of "advanced" programming from simple prompts. And the timeline doesn't start 2 years ago, I'd go back to 2004, when IBM started the Watson project.

Why aren't self-driving cars everywhere already? They are far, f

Worse than static code completion

[Martin S.]

Having tried various AI programming models, not one of them has even come close to the reliability of template driven static code completion.

Re:

[dinfinity]

You're using them incorrectly. You should approach them as a (sometimes incompetent) peer (not an oracle!) who has experience in the task you're trying to accomplish and whom you are asking for advice on that task.

With that frame of mind your expectations and way of dealing with the responses are far more constructive: when asking a peer for advice/tips you already accept that some of what they say may be incorrect or incomplete, but can still find value in what they say. You also understand that if you hav

Re:

[Tony Isaac]

What AI is most useful for, is to suggest code that does something you are not personally very familiar with. For example, if I want to perform an XSL transform in JavaScript, and that's not my best language, or I don't really know how to go about this in that language, AI can be a big help. Or if I've never written a CTE or a MERGE in SQL, AI can help. And it's better than a template, because it names things based on your actual request, not just generic names.

Re:

[destined2fail1990]

I don't know, I hit a record like 10 tabs in a row with Cursor and it basically wrote what I was trying to do for me 100%. I didn't have to do anything but write out my comments explaining what I want to code before I hit tab to code it.

The real value is when you have to type 30+ something parameters from a GET or POST request and parse them all out and do filtering on them. AI will basically autocomplete all of it and you don't have to deal with it. Saves a ton of time.