Being Smart About AI
Jul 11, 2023
Recent Advances in AI
Unless you’ve been hiding in a damp log for the past several months (if you have, that’s fine, I don’t judge these things) you’ve probably heard quite a lot about how AI is poised to change everything, everywhere, for all time. Between large language models (LLMs) like ChatGPT writing text, and Stable Diffusion systems making images, a number of people have gotten very excited, with lots of think-pieces predicting everything from utopia to apocalypse.
Recently, some people have started to display more nuanced views, amidst the tumult. So this is probably a good time to examine at least a few of the things these systems do well, where they can let you down, and some strategic ways businesses can use at least some elements of AI well. Bear in mind that this post can only scratch the surface of a complicated topic, and that some elements of this may be rendered obsolete within a year or so.
Strategies and Pitfalls
Let’s start with the Stable Diffusion part. Leaving the possible ethical and legal issues to one side, as somewhat out of the scope of a blog post, what are the benefits and what are the potential pitfalls?
On the benefit side, people clearly like being able to enter a prompt as text, and get something that might actually correspond with their request, i.e. “Richard M. Nixon as a circus ringmaster.” Once we get past the novelty value, however, these images often have a characteristic look that makes them easy to spot, even when they don’t give people extra fingers. The recent kerfuffle over the just-released (as of the time of writing) trailer for Marvel’s Avengers: Secret Wars movie, and its obvious use of AI visuals made this quite clear.
So where would be a smart and strategic place in an organization for the use of AI visuals? The most obvious is for quick first-draft storyboarding images for ads or video planning, or placeholders for final photos, illustrations, etc. in print, web, and presentation materials. These sorts of uses allow people to ideate and iterate quickly, and to communicate ideas to a range of stakeholders without committing time or resources up front.
Why not use the generated images as your finished art? Some people are certainly trying this approach, and in some niche cases, they’ve even had some successes. However, if you are in a corporation of almost any size bigger than the classic “five seed-round-funded programmers in a startup space,” you are giving up one of your biggest advantages: scale. A tiny startup can’t afford to hire creatives or even license artwork, but for even a mid-sized company, the costs are almost vanishingly small if these resources are well-managed.
At the same time, generated images are already getting tagged as a kind of cheap, lowest-common-denominator look, and that means that if you use them for finished art, your billion-dollar company looks indistinguishable from that tiny startup. That’s a huge loss of strategic advantage.
A similar set of cases occurs with using LLMs. The strengths these systems demonstrate include summarizing text, suggesting concepts, pulling keywords and phrases from a body of writing, and assisting with searches.
By now, most of us have also seen not-so-great use cases, often revolving around the “hallucination” problem. LLMs are designed to produce well-formed verbal responses to queries, but they lack our experiences of the world, so they might respond plausibly, but with wild inaccuracy. The recent case where a lawyer ended up in major difficulties after having an LLM write a response in a court case is illustrative. The response looked like solid written legal prose, but was found to include issues like citations of non-existent prior cases, even if those citations were correctly styled.
Beyond such obvious pitfalls, there are less overt issues. The writing these systems produce is grammatical, and follows the mix of written and unwritten rules for English quite well. They can, however, produce prose that sounds like a sleep-deprived college freshman hammered it out from some notes during an all-nighter, where it’s correct but full of generalities.
Part of the problem is in specific expertise. These are systems that, however cohesive their statements, only know what they know, and have no way to assess accurately the things they don’t know.
That can cause trouble for people who try to, for instance, have ChatGPT write a whole website. The results often convey fairly superficial and general ideas or information. If too much of a site is anodyne and low in focused information, it’s not just boring to read, but search engines tend to treat it as filler, and downrank it. This is not speculative on my part: a quick readthrough of Search Engine Optimization forums over the past few months will turn up too many people who used LLMs to build a 60+ page site, with lots of good keywords, and now can’t figure out how to get it to rank anywhere for their clients.
Also, good corporate communications have a “voice,” whether that’s casual, formal, academic, or some combination, that is part of the company brand. It can be difficult to develop, but we notice when it is missing and we, humans and not LLMs, are the audience. And while ChatGPT can rewrite all your communications into, say, cinematic pirate speech, that’s probably not the best idea.
So aside from the earlier-mentioned cases, where can LLMs actually help you? Again, ideation is a good point in your process.
Have a fuzzy wad of ideas, it’s Monday, and you need to get things a bit more organized for a report or white paper? Some reasonably good queries can turn that into a better-focused set of topics and even sample thesis paragraphs, and save you a lot of stress and time.
I have actually gotten decent results just by asking for arguments around a one-sentence concept, which returned several solid, if not particularly stylish paragraphs. A bit of fact-checking and editing could actually turn these samples into useful text.
If you have SEO tasks, LLMs can extract keywords and phrases from text, whether yours or a competitors, which can help you to align content with search-results goals. They can even produce good starting lists from a topic description. Again, they can be very helpful, sometimes exhaustively so, when you need to locate concepts quickly.
Summarization, which is sort of the reverse of expanding out a single topic, can get you from a complex paper to something like a set of slideshow topics faster, to cite a problem many of us will face in a workweek. Again, it’s good to think of it as a way to get started on a project, not as a way to complete it.
Here at Graph, we’re working on ways to incorporate LLM input that aligns with the strategic philosophy I’ve been discussing here. We think it can help to summarize a video from the video’s script or transcript, suggest chapter topics, and generally provide an assistive “co-pilot” role.
In short, you still have your job. AI is there to make it easier, not to take over. Because it might help you to navigate your workday, but we don’t think you should give it your car keys.
David Seaver is a co-founder and the Director of Design at Graph, Inc.