infinite action spaces

At the moment, every morning, I wake up, sit at my desk, and think about what I will work on for the rest of the day. This is actually pretty hard—this is because there is such a large action space. This is in contrast to what I’ve historically been accustomed to in school/engineering jobs—there, you have defined tasks to do, and all you have to do is figure out the sequence of actions to take to accomplish said tasks. For ideating in startups, nobody is telling you what to do, and you can literally do anything that you can think of (ok, within the realm of sane human being, legal stuff). The thing that makes this tricky is that after you pick something to do, you are lowkey locked into working on that for a while, during which you realistically are not going to be thinking extensively about what the best use of your time is. But the longer you spend just deliberating what to work on, the longer you are spending on…not working.

Figuring out how to balance this is super hard, and I feel like I’m doing a bad job at it. When I do pick a task to do, it feels like a suboptimal use of my time. When I sit around thinking about what to do, I am painfully aware that I’m not getting anything done either.

landscape

Wih the release of GPT-3 and the public’s realization of its shockingly good results, the first-order application that immediately comes to everyone’s mind is an AI-powered writing tool/text editor. It’s quite possibly the lowest-hanging fruit imaginable when it comes to generative AI, and there are several startups that do precisely that (with quite large valuations to boot). However, I’ve never personally used any of the tools, nor do I know anyone who uses them either. A brief tour of AI Twitter reveals countless demos of AI writing tools—why aren’t any of them adopted by the mainstream public, like how Github Copilot has proliferated the dev community?

A lot of the best AI writing tools, like Jasper and CopyAI, are built specifically for copywriting. This includes situations like generating an Amazon product description for your item, writing a blog post about a certain topic, optimize SEO for your content, etc. The product’s entire identity is centered around AI, but there’s zero AI tech built in house—most of these products are simply a wrapper around GPT-3 with a UI/UX tailored for specific copywrite functionalities. These products are entirely input-output driven: you type in a prompt, and the LLM-wrapped product spits out words for you to copy-paste and use. As AI models improve, anyone can easily imagine this replacing humans in copywriting positions. Even today, many people make a living off of being an expert user of Jasper/CopyAI and simply taking contractual, copywriting work. Yet, these products are often advertised as creative inspiration, something to help with writer’s block, rather than a fully automated copywriter replacement. The models simply aren’t good enough, the UX is separate from people’s text editor and is thus clunky, and, apart from the copywriting use case, essentially nobody utilizes these tools.

UPDATE 10/18: Jasper, an AI copywriting startup, raises $125M at a $1.5B valuation

This goes back to my long-running thread that, while AI is still developing/improving, it makes more sense to focus on augmenting skilled individuals rather than entirely replacing low-value workers. There is a lot of context stored in a human’s brain that cannot be captured by a prompt/preceding sentence, and models currently are simply not good enough to predict brand new content without said context. Furthermore, while AI models can imitate good writing, people have distinct preferences when it comes to writing styles—writing is often an art that’s unique to the individual, and having wholly generated content fails in most personal situations.

ai writing

For most people who write not for a job, writing is often a form of thinking. This means that an AI autocomplete model is not only difficult to execute well, but also destructive in the sense that the writer doesn’t get to think through the ideas to generate said sentence. To that end, a suggestion model here makes a lot more sense as the writer can focus on the idea generation itself, rather than painstakingly editing their writing style to make their ideas intelligible to the reader.

What I’m imagining here is a workflow similar to how I used to edit my friends’ essays back in high school. They would typically share a Google doc with me, and I would leave edit suggestions on said doc. Then, they can just go through and click accept or reject on all the suggestions. This sort of editing is nice because the original author still has ultimate discretion over what the final document looks like—if an edit is bad, then they can simply choose to ignore it. Similarly, an AI writing tool could make edit suggestions for the human writer to easily accept/dismiss—this has the massive additional benefit of being easier to make. The model doesn’t need to create great sentences all the time—it can simply choose to only display edit suggestions that it is highly confident in (it knows almost for sure that the writer would want to fix typos, grammar mistakes, and poorly written sentences). What takes away from the magic of most existing writing tools is simply that a lot of the content generated is really far off from what you intended on writing—in this edit suggestion model, only content that the user would find high quality will be displayed.

This scales not just for people who are bad at writing, but even experienced writers will dislike having to do a lot of busy work proofreading/editing long passages that they just wrote. Thinking back to my high school/college days, I always had the philosophy that putting down words on a piece of paper and then editing was much easier than trying to construct perfect sentences from the get go, and the same is true for LLMs. The AI model here simply augments this workflow by streamlining the tedious editing process, while also avoiding the hard work of generating the content itself. For many other applications, like art/design, the hard part is creating the content itself, but most people are able to formulate ideas into words—the trick is to turn those words into high quality pieces of writing.

I think edit suggestions are simply the beginning—the lowest hanging fruit would be fixing really poorly written sentences, grammar, and spelling mistakes. My friend sent me an article before written by a humanities person, and I noticed that the writing style in the article was very different from the tech articles I’ve been accustomed to writing (in this case, the humanities article was also about AI, which is why I noticed). It would be interesting to see if humans would be able to learn/incorporate writing styles that they admire/respect into their own writing by telling the the model to emulate specific authors. Writing is also a perfect medium to fine-tune with personalized AI—since writing taste is unique to the individual, for an AI writing tool to be truly good, it would have to learn what each individual likes. An edit suggestion model is perfect for this, because every time a user accepts an edit, the model can utilize that decision and surrounding context to personalize that model for the user. This is all in the realm of what people are used to with current text editors—as society becomes more and more accustomed to using AI-embedded products, you could probably get away with even crazier UX’s that unlock even more functionality.

Anywho, I haven’t fully committed to working on this because we’re pretty deep into AI design—working on this would feel like a lot of effort that’s tangential to our central goal, and I’m not sure about how efficient it would be. However, on a personal sense, it would be interesting to work on, and I would derive a lot of utility from it.