… what DinnerClub taught me about DeepSeek (well, sort of)
As someone who has spent years thinking about intelligent systems, first in human behaviour, now in AI, I had one of those moments where my brain short-circuits because two completely unrelated things suddenly make sense together.
This happened when I was reading about DeepSeek and the Mixture of Experts (MoE) architecture. Turns out Dinner Club, my pandemic dating experiment, wasn’t just a desperate attempt to keep love alive during lockdown, it was accidentally pioneering AI design.
Who knew playing digital cupid would one day help me understand machine learning?
At first glance, designing a dating experiment and designing AI systems seem unrelated. But both require understanding incentives, optimising decision-making, and building feedback-driven intelligence. Both involve trial, error, and the occasional catastrophic failure. That’s the heart of my work.
The Accidental AI Architect
It was 2020. The world was in lockdown, dating apps were flooded with bored people who “wanted to see where things go” (nowhere, my friend, nowhere), and I launched Dinner Club, essentially speed dating for the apocalypse, minus the awkward silences, plus some human intelligence.
Unlike dating apps, where everyone swipes incessantly into a void, I set strict limits, including three potential matches max, mandatory feedback forms (yes, homework), and a social credit system that rewarded kindness.
Because apparently, adults need credit to remember basic courtesies.
For those unfamiliar with AI terminology, a Mixture of Experts (MoE) system is an efficient architecture where a “router” directs inputs to specialised “expert” models rather than running everything through one massive system. It’s like having specialized doctors instead of making everyone see a general practitioner for every ailment.
This is similar to how LLMs like DeepSeek utilise MoE architecture as part of their design. I had accidentally built a human version of a Mixture of Experts system, complete with routing algorithms (me, playing matchmaker) and specialised experts (highly illiquid people in the dating market who excel at specific types of connections).
The Anti-Tinder Manifesto
Traditional dating apps are like those massive LLMs everyone’s obsessed with, that burn through resources like a tech bro burning through his Series A funding.
Every profile could potentially match with every other profile, creating an inefficient system. Dinner Club took a different approach. Like an MoE system’s router, I acted as the gatekeeper. I directed each person to a limited number of matches based on both obvious and subtle compatibility patterns.
Sometimes these patterns were unconventional. “Both similarly weird” turned out to be a surprisingly successful matching criterion, though I highly doubt traditional matchmakers would approve of this.
This routing efficiency was only part of the system; equally important was how we collected and used data to improve matches over time.
When Feedback-Forms Met Feelings
Those mandatory feedback forms after every date weren’t just bureaucratic exercises, they were valuable data collection tools. Each date generated quantitative ratings on niceness and compatibility, plus qualitative feedback to refine future matches.
It was basically A/B testing for hearts, using guided preference evolution. This dating approach resembles techniques already used in AI development like preference learning and RLHF, but applied to human relationships.
Take the woman who insisted on dating men who “who loved Eckhart Tolle and lived in the present”. After I matched her with exactly that, a wanderer who travelled the world with a satchel and no savings, her tune changed swiftly. Suddenly, “future-oriented” didn’t sound so bad. Funny how that works.
The Art of Being Wrong (gracefully)
When participants clung to rigid preferences (looking at you, “must be a CEO of a funded startup” person), I didn’t just shrug and move on. Instead, I developed a three-tier approach, courtesy my inner therapist:
- Self-discovery exercises (people prefer to realise they’re wrong on their own)
- Pattern-based insights (12 years of matchmaking teaches you that “must love dogs” is rarely the real deal-breaker)
- Experiential learning (you have to let people date the wrong person to appreciate the right one)
This is where AI systems could actually level up. Imagine an AI that doesn’t just nod along like a sycophant but subtly nudges users to expand their horizons, like a trusted adviser. It’s the difference between “I understand your preference for emotionally unavailable partners” and “Have you considered therapy?”
The Genius of Social Credit
The social credit system in Dinner Club started as a way to gamify good behaviour, but it revealed something deeper. When we reward the right behaviours, we don’t just get better dates, we build a better ecosystem.
It’s like training a puppy, if the puppy had an MBA and unresolved issues. The genius wasn’t in the points themselves but in how they rewired behaviour. Kindness became its own currency, which is probably the most capitalist approach to decency ever attempted.
This is essentially the reward function for reinforcement learning in LLMs, where models are trained to maximise positive feedback while minimising negative outcomes.
Just as my system encouraged daters to be more considerate and responsive by rewarding those behaviours, RLHF shapes AI responses by reinforcing helpful, harmless and honest outputs while penalising problematic ones. Both systems evolve through iterative feedback, gradually aligning behaviour with desired outcomes.
A Systems Thinker’s Evolution
My work now isn’t matchmaking. It’s designing intelligent systems, whether human or AI. Because at the core, I’ve come to realise that they both require structures that don’t just guide behaviour but create better decision-making, better relationships, and ultimately, better intelligence.