As Coda Rolls Out its AI Beta, Now May Be the Right Time to Reflect on What a Modern Document Is
In early 2006, just before the iPhone was released, Google Docs became a thing. It was originally called "Writely”, a web-based word processor created by the software company Upstartle. Google acquired Upstartle in 2006 and rebranded Writely as Google Docs. It becme part of the Google Workspace suite of productivity applications, which also includes Google Sheets, Google Slides, Google Forms, and Google Drive. It immediately challenged Microsoft Word because it wasn’t bound to the desktop.
But Docs had one hidden feature that few saw coming; integrated scripting - Google Apps Script. It was the beginning of the age of “smart documents”. Despite this early and clunky mix of scripting and pervasive web access with collaborative editing, it signaled a sea change in the definition of a document.
Coda AI
Fast-forward fifteen years, and like many products in a gold rush to smartify their offerings with AI, Coda dives in with a back three-and-a-half somersault with inward pike entry and barely a splash - a 9.7.
During the alpha phase, I was very critical of Coda AI’s direction. And I also wrote at length about the vision of AI blended with Coda in 2022.
My expectations rose abruptly when they put the stake in the ground beyond the first AI release calling it “the AI Writing Assistant” - a definition that includes many requirements. At the center of the writing assistant’s objective is the ability to perform in-place text creation, corrections, templates, and pattern recognition. For me to consider this a writing assistant, it would need to perform elegant things to mitigate my writing steps without disrupting the writing flow.
While there is no hardened definition of “writing assistant”, there are certainly existing solutions worthy of comparison. A good one is Grammarly. Historically, using AI, Grammarly has performed pattern recognition and corrections quite well for more than half a decade. Last month it added integration with OpenAI, and now it performs in-place text generation and other prompt-based assistive writing tasks. As mentioned in this Substack piece,
Grammarly is to writers what Github CoPilot is to coders.
I fully appreciate the desire to integrate AI writing assistant capabilities for Coda users. The benefits of integrated assistive features would [presumably] span Coda’s object and data types in ways that an external solution could never compete. However, to attract users to Coda’s writing assistant, it must meet a threshold that is at least as useful as an external product could provide. Why is this important?
Any internal AI writing assistant must compete with [and exceed] the abilities of external solutions that work seamlessly across the OS.
As you may know, Grammarly typically installs into the browser. However, it can also run at the OS level. I use the Mac version with Coda and every other app on my Mac. Grammarly provides a seamless and consistent feature set across all work canvases. I’m using it as I pummel you with my thoughts right now.
Unashamed, Grammarly is a lowest-common-feature play. It has no ability (yet) to provide assistive tasks related to canvas-specific tools. But it’s still early, and AI is a clever beast. I anticipate tools like Grammarly will add AI extensibility, and when it does, it will transition from a tool-agnostic assistant to something much more powerful.
Coda’s challenge, therefore, is to create something whose assistive powers excel so greatly over the tool-agnostic assistants, that its attraction is undeniable.
Coda AI Achieves this.
This is the hallmark of disruption. Ironically, Coda must disrupt the disruptors. AGI is advancing at an unprecedented pace such that the definition of obsolescence is almost immediate and certain. Grammarly achieves straight-line productivity, an app phenomenon made possible by AGI itself. I measure AI apps intended to create hyper-productivity by their capacity to transform explicit actions into implicit actions.
Can Coda AI do that? Let’s see.
Subscribe to more content like this (Impertinent)
Implicit AI Actions
Imagine a column of guideposts containing an outline of tasks for a planned press release that needs to also exist as a JSON payload for another automated process.
Until today, this required a complicated formula or a Pack. Now it requires only a prompt.
Automated Transformation of Outline into JSON Action Items
In this example, the AI prompt is established as a column formula and applied to all rows in the table. Examined more closely, the prompt expectedly must reference values from other columns, and this is accomplished with the @ reference, a feature that exists pervasively throughout Coda’s platform. This is cool but also essential if you want to achieve implicit smartness.
Coda AI Prompt
But another capability exists in the prompting feature that insulates document makers from the complexities of AI - selectors that control length , tone , and type of inference. This is a subtle but very powerful abstraction that everyone will enjoy. It streamlines the development of AI-enabled data tables by giving users the ability to apply configuration settings in non-technical terms.
Prompts are First-class Templates
Despite these ideal prompt configurations, there’s one more thing - formulas . At first glance, there’s no indication the AI prompt dialog can access other tables, other fields, or canvas-level placeholders. As is prominent throughout Coda’s user-chummy dialogs, the equals (=) sign is the gateway to advanced formulaic integration, and this is also the case with AI prompts.
You can insert formulas of any type directly into the prompts, thus transforming them into powerful templates. This means that you can also integrate Packs and use their output as prompt inputs.
Prompts can be wholly dynamic, drawing data, content, and values from everywhere imaginable into the AI process.
Writing Assistant - The Usual Suspects
To meet the threshold of a “writing assistant”, we must examine how Coda addresses common writing tasks. Let’s see if it meets or exceeds Grammarly, for example.
This is Grammarly. Not bad. It transformed the base text into three scenes of the story and allowed me to replace it with a single click.
This is Coda AI. Highlight the base text, and select the AI button (or press Ctrl-Space).
Now select the objective.
Complete the prompt and select the Create button. Pretty good output, and the process flow is slightly better than Grammarly. Both AI systems replace the highlighted text. This is a drawback - there should be an option to generate the content and leave the base text in place.
The Unusual Suspects
In-place Summations
This one caught me by surprise. Imagine some data in a table - like customer survey responses. Now imagine a text narrative about that data. In this example, I wanted to create an analysis of customer sentiment about selected maintenance providers from our survey data. Use the /sum canvas command to insert an inline AI block. The prompt is freeform and can reference any Coda objects, including data tables and columns.
The outcome is [seemingly] magical. But that’s just the beginning. As the data changes, so does the analytical narrative. Wait! What? Yeah, this is pretty cool until you realize that LLMs are not very good at analytics, especially where the data may be large or raw. If you refresh this AI block repeatedly, you will get different results. It’s often close but not accurate enough for prime time.
The only way to encourage Coda AI to be precise about data (which is really OpenAI’s LLM) is to perform intermediate computations that give the LLM a chance. Coda’s formulas are exactly the right remedy that can close the gap between errant mathematical outcomes and precise assessments and narratives. Unlike the first attempt, this one generates accurate analytics.
And in true Coda spirit, AI blocks can be arranged to your liking.
Generating Data Tables
Yet another Coda AI surprise is the ability to generate data tables rapidly. Provide a prompt that clearly states the data you want, and a few seconds later, you have the data.
Maybe you need some test data.
Smart Documents
The future of documents is undoubtedly leaning toward smarter, more intelligent works. The examples presented here show some building blocks that can make documents smarter and more useful. There are many more and the possibilities are limited only to your creativity. The ability to seamlessly utilize formulas and table data will be transformative.
Coda seems to have risen to the challenge of integrated AGI provided in a way that blends two key axis - simplifying AI for greater productivity and incorporating AI into all that has previously made Coda special.
