Having worked on dialogue since I was basically five years old (I wrote my first play in kindergarten, you can imagine!) — linguistics, theater, screenplays, discourse institutions, now Interintellect and conversation AI — it still feels crazy to me that we invented LLMs before we could fully understand how humans communicate.
Of course playwrights always knew. But do the people building these dialogue systems know? Doesn't feel that way.
Interintellect, my teacher
After building AI interventions for the first few years of the company (for 1:1 texting, group chats), Interintellect has now run human-only for 7+ years and we figured out a ton about what enables people to have good conversation at scale. (My background in language + theater helped a lot, but there still were surprises.)
Some of these are:
1) Deep conversation is a ritual.
Aristotle's unity of time and place applies. Make dialogue intentional, with a beginning and end. Humans prefer to make an effort, including cognitive and emotional, when they know where it ends, what the end and/or goal looks like. This is hard to do in the continuous texting era, but small rituals (welcomes, room style UI) are doable even within apps.
2) Real encounter requires identity.
Whether you're a well-known (!) anon or present with your real name and face, being identifiable and a recurring character will make your conversations so much better -- you'll have a context.
Listening and remembering is relational investment: if the other participants know they might see you again, you will make and receive more such investment. Invest in your character and you have invested in your conversation outcomes.
3) LLMs still don't understand time.
But humans do. We are mortals with always limited timeframes whatever we do, our communication is t-axis first, escalations arise from repetition, we are animalistically aware of gaps, beats, slowdowns, urgency, pressure, and time passing. Like in a dance, the physical form of language (and its physical effect) can be much more important than the content of what is being said. The human animal is a subtext reading machinery: we're rarely talking about what we are seemingly talking about, and yet we somehow end up discussing what we wanted.
On Interintellect we solve these problems with a mixture of directness and buffers:
All online and offline spaces on Interintellect are intentional, paid, real name, real face, real time. Come as you are.
Most topics have "containers". Interintellect hosts invite attendees and members to discuss Madame Bovary, Tocqueville, Dostoyevsky, and the First World War, and from there they can jump into sex, freedom, death, and fathers as an option but not obligation.
We are an institution and so we move people away from the immediate urgency of markets. You are given "decadent" time to immerse, discuss, understand, and bond. To grow your self and your relationships. The space is itself the ritual.
Different organizations and products will have different solutions for these needs, but having a solution in place is crucial.
When people ask me how I built categorically better conversation as an artform at scale, I always say there is no “one thing” — it’s the many small things and the constant learning and iterations. LLMs would greatly benefit from such design solutions as well.
The carnality of shared intention
We should think of conversation as a lump of Play-Doh which doesn’t exist unless we, together, simultaneously all decide to create something out of it. Its form will be ephemeral, its effect is forever. It’s the original, purest act of genesis, of generation. (Logos becomes reality, the word becomes world.)
LLMs are such conjurers too: whatever they will say doesn't *exist* until the words of a human make it coalesce out of the Collective Information of humanity. It comes out of nothingness for us.
But can LLM language take a shape, the way human language does when we use it with each other?
What will it be like when the LLMs will be looming in all human conversation that we’re having with each other — the way I predicted 10 years ago — if they lack all intuition of this material effect, all grasp of temporality, the severe physical limits within which human beings unfurl their lives and manage their relationships?
Who will build the ritual? How will humans have Good Conversation?
More on LLMs:
Kundera, Hegel, and the AIs
I was sitting in San Francisco last night with a book and a glass of wine, and I got into a conversation with the LLM Claude. I had been meditating on a passage in Kundera — the book I was hosting an Interintellect reading club about, a book I was re-reading after a very long time. In
Funny how we now have to consider slowing down as a gift of decadent time. I immediately thought of The Gift of Time, which is a collection of tributes to the work of software engineering/systems development pioneer Gerald Weinberg. Largely written by engineering types who'd worked on organizational development efforts, the essays share a common theme of the need to create space in which to grow.
Tom de Marco, another software pioneer, argues in his book Slack that organizations and individuals cannot achieve substantive change without accepting that it inherently requires a willingness to waste time — or rather, to spend time without a guarantee of immediate results.