ChatGPT 4o with Canvas Review

2024-10-04
#openai, #chatgpt
337 Words
2 min

Yesterday OpenAI’s advanced voice feature was released to free users, and today they released a new feature exclusively for Plus and Team users. Not hard to see that subscribing to OpenAI membership is quite worth it—you can try new features first.

Official introduction: A new way to work with ChatGPT for writing and coding.

Writing Scenario Shortcuts

Writing scenario shortcuts

Notes:

  • You can adjust and modify on the canvas yourself. Each paragraph can have GPT edit, explain, or rewrite.
  • Use “Add emojis” first then combine with “Add final polish” to get a reasonable article with a certain number of emojis. Pure “Add emojis” results seriously affect reading experience, see image below.

Pure “Add emojis” result seriously affects reading experience

I feel this is suitable for writing or paper polishing scenarios—continuously adjusting to achieve expected results. New things are definitely worth trying and exploring, but whether the 4o model with canvas can match or even surpass Claude 3 Opus in writing remains questionable.

Coding Scenario Shortcuts

Coding scenario shortcuts

Notes:

  • You can adjust code on the canvas yourself.

Using my “impossible triangle” pattern from yesterday as an example.

This was generated by Claude after multiple iterations—it kept failing to accurately center the pattern inside the triangle. After several iterations, Claude 3.5 Sonnet’s corrections were minimal.

Claude 3.5 Sonnet couldn’t center pattern inside triangle

This is ChatGPT o1-preview’s first-attempt success—accurately centering the pattern inside the triangle.

ChatGPT o1-preview successfully centered pattern inside triangle on first try

ChatGPT 4o with canvas took 2-3 iterations to successfully fix and place the pattern in the center of the triangle.

ChatGPT 4o with canvas took several iterations to successfully center pattern inside triangle

Overall, personally I think o1 plus canvas would be a better combination. I estimate OpenAI will switch models later—the 4o model is still weak. But this example doesn’t deny Claude 3.5 Sonnet. In my mind, o1 and Claude 3.5 Sonnet are the current closed-source LLM kings—after all, LiveBench’s leaderboard is more trustworthy, while LMArena’s has more bias.

LiveBench LLM leaderboard

Canvas is the first major update to ChatGPT’s visual interface, in early testing stage. OpenAI is also planning to rapidly improve its functionality. I estimate Canvas will later add Claude Artifact-like features for coding.

Appendix

Attaching some user evaluations of this feature I saw this morning.

User evaluations of this feature

via: https://openai.com/index/introducing-canvas/


Emoji Reaction


© 2022-2026 Made with ❤️ By Jiakai