An Imperfect Canvas ... But An Interesting First Step
Recently, OpenAI announced the beta test for the new Canvas feature in ChatGPT, released in some GPT-4o versions. Canvas partially addresses a glaring hole in the OpenAI ecosystem - how to you use ChatGPT to actually create content, rather than simply regurgitating it.
The idea of being able to create editable documents via Chat-like interfaces has been around for some time - you are after all just modifying a text buffer containing markup, something that a novice HTML developer can write in an afternoon. What makes this more complex, however, is that you need to have a mechanism to persist this information, and a way to dynamically change content that is not, strictly speaking, serial in nature.
The intent of this should be self-evident: At the moment, ChatGPT is essentially a read-only environment and while you can, with some care, add new content, you REALLY need to understand what you're doing in order to not wrap yourself around a tree.
The goal with Canvas is to create just such a layer. It would serve to potentially compete against other web-based platforms such as Google Office or Microsoft Office 365. The goal is laudable, and arguably necessary, as the one thing that ChatGPT by itself is not is an?editor.
What Is An Editor in the GenAI World?
As I have written previously, I see ChatGPT and similar environments as being?Confabulators. Given an input prompt, they make up things, trying to stay as germane to the topic at hand as their training data allows them to do. What they don't really do all that well is provide a vehicle for actually letting writers and coders actually write or code. Instead, you ask the assistant for information, and the assistant makes up a response.
This is not quite the same as editing. With an editor, in general, I'm trying to create something in an environment that is configured to make it easier to write, and I generally do not want the ChatGPT getting in the way.
The closest analogy I can think of is imagining that you are the hero in some late 19th century novel, a more or less penniless lord or lady who nonetheless has his faithful butler Jeeves in the background. In general, what you want Jeeves to do is to provide assistance when you ask him (or her, Jeeves could always be a Lady's Maid rather than a Gentleman's Valet) but when you don't need their assistance, they stay in the background.
ChatGPT by itself doesn't do that. Everything that you write comes back with some form of commentary, often unwanted commentary, as if your personal assistant just doesn't know when to stay quiet in the background.
I've tried using ChatGPT in an editorial role, and even in the most recent incarnations, that can be extraordinarily disruptive to the flow process of writing.
I think this realization finally made its way to OpenAI, because before ChatGPT can move into the role of being a personal valet or secretary, it needs to have that capability. Canvas is a first attempt at doing just that.
A Canvas is in essence a modifiable document (sort of). You can write to the canvas, incorporate headers, paragraphs, and a limited number of other markdown capabilities, and can then both persist it and retrieve it as a document by name or reference.
Writing involves using the prompt window to write new content, which can then be added into the narrative being developed in the Canvas pane. This actually works pretty well for baseline editing work, as it is not?all that?dissimilar from the way that you usually work with an editor. I find that it's worth making the prompt pane and display a little bit wider (you can drag this left or right by hovering over the midpoint and then dragging), so that you can get a sense for the size of paragraphs, but it has the advantage of letting you focus on the writing part.
This can also be done via a voice interface. Some people can dictate directly into a ChatGPT and the canvas works fairly well, though I personally have trouble with voice-based interfaces. Dictation sounds great in principle, but we often speak differently than we write, and that discrepancy can make such voice-based editing systems more trouble than they are worth (they look cool, but as anyone who has attempted to transcribe notes in a noisy coffeeshop can tell you, the reality is seldom quite so forgiving).
What Canvas Does
Overall, neither Microsoft Office nor Google have much to worry about just yet. I found that I spent a fair amount correcting the Canvas output, especially when it was making revisions. From what I've seen, Canvas makes use of its context to provide working memory for the application, and this utilization likely runs counter to other uses of the same context. I made it a regular habit of telling Canvas to create a back up link that I could download (internally it's stored as MarkDown) so if something went wrong, I wouldn't lose everything I'd written. This precaution served me in good stead more than once.
You can add new MarkDown content. I've noticed, for instance, that I can do the following:
领英推荐
However, the longer the document, the more likely it was to make serious errors and delete whole sections, necessitating that very careful attention needed to be made when trying to do novel things such as adding code or specialized modes.
You can go in and manually edit content in the document as well, though currently doing so seems to also toggle the record function (both make use of the space bar), which became a huge distraction when working with the canvas directly. It was simply easier to use the prompt editor.
Early Days for Canvas
I don't think that either Microsoft nor Google have any real reason to fear that Canvas will completely eliminate their respective office suites. Like all editors, people use editors that they are comfortable with, and the capabilities that Canvas provides, even if they worked reliably, are a far cry from any commercial editing system.
Another major issue was the comparatively slow speed of update when adding new content, as the entire document periodically had to be refreshed. This suggests that, at least at the moment, canvas is probably pretty good for writing email, taking notes, and composing relatively short blog posts. To be honest, its biggest utility right now is as a potential substitute for Microsoft OneNote, which I've long felt was a brilliant idea implemented badly.
Yet for all that, Canvas is an interesting (and necessary) experiment. I was prepared to be disappointed with Canvas, and the bugs here are such that I'd probably have weighted to provide an open beta for a bit longer until they are resolved, but the paradigm that ChatGPT introduces, one built around documents that are produced without the formal rubric of "applications" is actually one I find quite impressive.
Canvas and its imitators will play a significant role in collaboration (which I see as simply a succession of edits by different users in a multi-user chat) and it has the possibility to finally overturn the hoary paradigm of files and folders. I'm sixty years old, I remember both of these in their physical incarnations, but as an organizational paradigm they are definitely subpar in comparison to a decent multidimensional taxonomy system, something I see Canvas being particularly well suited for.
By the way, this article was initially composed on OpenAI Canvas. It was maddeningly imperfect at times, but there is clearly potential there if OpenAI spends some real development time in user interface rather than just trying to make this week's iteration of AGI. I'm not necessarily interested in god-like general intelligence, but a truly useful text editor is something that I think has real value. Hopefully Canvas or something equivalent proves to be that editor.
In media res,
Editor, The Cagle Report
If you want to shoot the breeze or have a cup of virtual coffee, I have a Calendly account at https://calendly.com/theCagleReport. I am available for consulting and full-time work.
Everywhere, knowingly with the bG-Hum; Crusties!
1 个月L'Editeur Chef has been rather prolific this past week. I was still thinking over this Canvas missive when the political deep fakes one crossed my roving desktop and so rather than attend to that tortuous topic I return to the Canvas one. Kurt, What you are describing canvas to be doing is to get into your face and interfere with the normally seamless flow of one's frontal lobe.? I see it here with LinkedIn messaging insisting on being the first at thinking your thoughts to respond with. In so doing it captures your screen and mouse cursor and holds you hostage for it's offered answers.? It gets worse when it freezes your edit controls and offers it's one two or three Pong like quip buttons to hit while you wait for it's permission to say something. It's been a few days since I began penning this and have since noticed that what I describe here gets severe when one of the chat participants actually passes your chat dialogue through an AI chatbot. A two-way full-duplex conversation becomes half-duplex on coherent rules of parliament and then quarter-duplex when yo think you are communicating to a full-duplex human. And Chatty in-spam does its clever Type II errors on its thoughtful lookahead hallucinative counter-responses.
Aspiring Data Scientist/#BuildwithAI 2024 Hackathon with GenAI Ambassador for Europe and Africa/#BuildwithAI 2024 Hackathon with GenAI Mentor/GenAI Pioneer/
1 个月Insightful.
Security hardware and software architect
1 个月Right, order, or knowledge, is entirely a figment of our collective imagination. This shows that the net information content of our scribbling is essentially nonexistent.
Bible lover. Founder at Zingrevenue. Insurance, coding and AI geek.
1 个月So now we are increasing the demand load on AI vendors' GPU clusters. In the early days of ChatGPT, serving a 7 billion parameter LLM to a million users consumed over 50 MWh of energy. That's around between one to two hours of peak energy supply roughly 80,000 homes from a large scale battery energy storage project (BESS). Now as we are in the era of trillion parameter LLMs, and with waves of edits hitting said GPU clusters, those with critical life support equipment, hospitals, malls, schools etc will have to ration their power. I mean we already have homes and EV charging stations maxing out the grid. Note that the problem here is not the electricity supply, it's the demand. ??
Bible lover. Founder at Zingrevenue. Insurance, coding and AI geek.
1 个月Editing in Canvas is a tiring experience I can tell you that. ??