4/12 OpenAI Canvas is Disappointing...
... for coding, but amazing for writing!
Good morning!
Today's OpenAI release is all about Canvas. OpenAI has rolled out a new version of Canvas to everyone.
So, what’s new (specific to Canvas)?
Side-by-Side Editing: You can edit AI-generated text while ChatGPT offers helpful suggestions (if you’re lucky). OpenAI is pretty much trying to replace Notion with this one.
Leave Comments: ChatGPT can now comment on specific text parts. It will only delete or edit your text without you knowing it, thus avoiding having to reread it all. It can leave comments, as we can on Google Docs or Notion.
As per GPT’s comments, it seems like it didn’t read the rest, as we mentioned the comparison with Google Docs for comments. Regarding Notion, it is true that we were assuming you knew about the tool. It’s basically a very useful text-editing tool that can do almost anything, from writing to editing, commenting, adding images, formatting, tables, etc., similar to a very powerful Word document.
Regarding the Skepticism, that’s what we were going for, so skip on that suggestion! This is a good time to remind you not to automatically accept all GPT changes, which is nice here as we can just close them rather than playing with the new generations and replacing lines while going through all of the newly generated parts vs. the old version. In the specific case of editing a newsletter, Canvas saves us a lot of time!
Custom GPT Integration: There is not much here, but GPTs can now use Canvas.
Run Code in Canvas: You can now write, debug, and generate graphs directly within ChatGPT. We are not convinced how useful this is as most IDEs can already do this quite without much trouble and have access to your files (e.g., Cursor), but it could be helpful to for one-off quick tasks you'd usually do on Colab.
About Coding with ChatGPT Canvas…
After some testing of coding with Canvas, here are some key points…
Pros
It can use the web search tool in the chat session and apply changes to the code, but you must ask it in the chat; it does not have an “apply” button.
It can translate code to PHP, Python, C++, Java, Javascript, and Typescript.
You can ask it to add comments to the code (explanation and details for each line), which can be useful to learn.
Cons
Canvas can run only Python code (for now). Running HTML/Javascript code (React components) or SVG like we can with Claude Artifacts would be a welcome improvement.
Third-party libraries are not installable (since you do not have full control of the environment).
You cannot use o1 with Canvas.
The code execution in Canvas cannot read files (CSV or other file types uploaded to the chat), while the old code interpreter tool feature can.
You cannot share the Canvas and collaborate with others.
We tried to train a basic model on the MNIST task and were disappointed. We could not successfully run the code once because of missing libraries. We tried to prompt the model to use other popular libraries (Tensorflow, Pytorch, Numpy), but nothing worked. Even if it would happen to work, we couldn’t download the model weights or anything from it.
In short, the code feature is still in early development. It can be helpful for small tests and demonstration purposes, but for now, we believe IDEs + AI, such as Cursor, are much more efficient for internal code development. ChatGPT remains useful for debugging and quick questions in the regular chat or quick one-off tasks requiring basic libraries.
Unfortunately, no updated Bingo slots today, but here’s the sheet so far—a blank day doesn’t mean we aren’t watching:
See you tomorrow for day 5!