Artificial Improv

As most of those who know me are aware, I've been acting and otherwise performing on stage for the majority of my life - from school plays to after-school groups; university theatre companies and little theatres. I've tried my hand at standup, singing and improvisation.

That last one has always been something of a safe space for me. As someone who is historically averse to artifice - I hate "role playing" exercises for example; I struggle to really define why - improv feels like a second home to me. It provides an easy space for me, one where I can relax and belong, where I can show off and fail.

I would hazard a guess that most people in the UK and USA will know of improv through Whose Line is it Anyway? - we call this "short form" improv, as it's mainly about short sketches, scenes or theatre games. As the name suggests, there also exists "long form" where whole plays are acted out, based on audience suggestions. Here in the UK one of the more famous examples is Showstopper! The Improvised Musical; in the US troupes such as The Upright Citizens Brigade regularly perform "Harolds" as part of their act.

But at the core of it, it's all the same - performers come into the show with nothing more than their own life experiences and knowledge, and with some structure in mind for their act, but ultimately get their performance seeded with suggestions from the audience, which we call prompts.

Not only do I enjoy performing, I have also found improvisation incredibly useful in all walks of life; it builds confidence in front of others, it teaches the ability to accept and adapt to the situations one is in, and it helps you explore your own creativity.

There are two core tenets of improv that are oft repeated and are nigh cast in stone as inviolable rules: "Yes, And" and "No Blocking."

"Yes, And" - variously called "embrace and endow," or "accept and extend" - is key to the whole act, helping it flow and ensuring that improv is a team sport. You don't build a house by having a number of builders each doing their own thing. Instead, you work towards a common goal - the only real difference is that said goal is rather emergent during a scene, with the performers figuring that part out during the course.

"No Blocking" is the complement to "Yes, And." It mainly exists to remind the improviser not to put up dead ends for those performing with them. In our house example above, it's akin not only to each builder constructing their own thing, but interfering with the work of others, such as removing a brick they just placed. Blocking is essentially the act of denying the input of your fellow participants; it not only interrupts the flow of a scene, but invalidates the other.

A lot has been written about how these rules are also inherently useful in business and team environments; indeed, expanding on my TEDx talk on Team Building Through Deception we can also see Bruce Tuckman's Stages of Group Development - or "Forming, Norming, Storming and Performing" - model of team building in every scene we construct. In fact, I highly recommend everyone in business check out Improv4Business, conceived of and run by my good friend Dr. Alexander McWilliam for more information and training on how to use improv in your day to day activities at work.

These two rules, plus a lot of practice, lead to a glorious parlor trick. Between the performers, a brand new shared reality is created - one that's not just for them, but for the audience. The "Yes, And" philosophy helps enrich and enhance the reality; the "No Blocking" rule helps to stop the shattering of the group illusion that's being spun. People watching suspend their disbelief, aided by the consistency and coherence of the tale being told, and the interactions being observed. This happens, of course, in all sorts of art, but the added knowledge of the core conceit - that those telling the story are making it up on the spot - helps draw them in.

With all of the experience I've had over the past three decades of doing this, I've inevitably done some philosophizing of my own, and often try to impart this piece of advice or verity to those who doubt that they can be improvisers:

Every conversation we have is improvising. Just like improv, we don't know what the other person will say; we often take it at face value, and we don't know what we are going to say in return. There are thousands upon millions of responses we can give that not only accept their contribution but enhances it, and gives them something to respond with in return. We exchange ideas and emotions; we represent ourselves as well as our reality. In short, the trick to improv is solely to come back with something that doesn't deny the premise, that sounds right and makes sense.

Once you've mastered that, the next step - if you want to, of course - is to practice until you get better at making that selection; like exercising wit. As with any exercise or skill, if you keep practicing, keep putting yourself on the line, and you'll find yourself improving and getting more and more flexible with how you perform, and can tailor your act to the audience or those with you on stage.

Before and during the COVID19 lockdown, I was a member of The Free Radicals, a Reading UK-based improv troupe that based its shows on STEM concepts. The first half of our act would be some well-worn theatre games that we rebranded with more scientific names or concepts - my favorite was playing the Alphabet Game with the chemical symbols of the Periodic Table of Elements, in order or atomic weight.

At the beginning of the second half, we would have a scientist, engineer or other such STEM speaker come out and give a short talk on their field of expertise. We had talks ranging from fusion in acrylics to satellite imagery and photogrammetry, to contraception in elephants. The performers would listen to these talks for the first time, pick up ideas and themes and then launch into… something! Sometimes it turned into a coherent story; other times it would be a collection of sketches, all to a varying degree of success.

Our (mostly) amateur troupe was made up of a real mix of experience in STEM subjects, but on the whole those from humanities backgrounds. I, of course, had a leg up on them with my software engineering expertise, but generally the topics kind of went above the heads of the improvisers. And yet we soldiered on, and built our own little realities with the audience watching on - realities that they could buy into. And this gets me to the lede of this article...

With a modicum of a background in machine learning - I covered it a few times at university, as well as giving talks featuring ML concepts (Simple Machine Learning and Graphs are Everywhere!) - I've observed the meteoric rise of Large Language Models (LLMs) with some scepticism. Those I've worked with and for over the past couple of years know me to be a cynic when it comes to what people are currently calling "Generative Artificial Intelligence." They are not AIs, instead they are examples of machine learning; it's mostly a very impressive parlor trick.

A parlor trick that suddenly came fully into focus to me today: LLMs are improvisers. They don't think (I hate that they're labelled as AI), they don't understand what they've been taught, they don't even understand what they're told when users interact with them. Instead, they just know how to "Yes, and" you, and to "not block" you... They know how to respond with something that makes sense given the reality that has been built for them through their prompts. They apply some randomness to construct what works in reply, figure out which is the best of those choices and reply back to you.

Think about it - when interacting with a tool like ChatGPT, you're actually improvising with it. You enter your prompt to seed the shared reality - like an audience making a suggestion - and it replies with something that fits that reality. Now, just like any good improviser, if you "block" the LLM by fact-checking it, it will "Yes, And" you and thus change the course of your constructed reality by trying again with something else that fits.

And herein lies my thesis: just like the improvisers in our STEM troupe, who are mostly just regurgitating stuff they've heard without necessarily understanding it, in a way that ostensibly only makes sense for the context they are working in you should not trust the output of an LLM when asking it questions without double-checking what it says against a resource that represents real knowledge.