2025 has been a strong and eventful year of progress in the intersection of Dating and AI. The following is a list of personally notable and mildly surprising "paradigm changes" – things that altered the landscape and stood out to me conceptually.
1. The Missing Supervision Layer (RL from Verifiable Rejection)
At the start of 2025, the "dating stack" for most people in the wild looked something like this:
- Pretraining (Accumulating life experiences and photos ~ongoing)
- Supervised Finetuning (Asking a biased friend "Do I look good here?")
- Zero-Shot Inference (Uploading to Tinder/Hinge and hoping for the best)
This was the stable and proven recipe for staying single for a while. The core issue – and the blindness of the industry – was that dating has been treated as an unsupervised learning problem. You upload photos, you get noisy feedback (matches or silence), and you try to converge on a "self" that works. But there is no clean loss function. There is no gradient. You are training against noise.
In 2025, the concept of a "Supervision Layer" emerging from crowd-sourced intelligence became the de facto new major stage to add to this mix. By exposing a profile to a verifiable audience (e.g., 50 women in your target demographic voting anonymously), we suddenly generate a reward signal. The profile can now "learn" strategies that look like "attraction" to real humans. It learns that "bathroom selfie" is a local minimum, and "candid with dog" is a global maximum. These strategies would have been very difficult to achieve in the previous paradigm because the feedback loop was broken by design (apps don't tell you why you failed).
Running this "RLVR" (Reinforcement Learning from Verifiable Rejection) turns out to offer high capability/$ in dating outcomes. It gobbles up the "compute" (emotional energy) that was originally intended for endless swiping, and redirects it into a tight optimization loop before you even enter the market.
2. Ghosts in the Mirror / The Perception Gap
2025 is where I (and I think the rest of the industry also) first started to internalize the "shape" of profile failure in a more intuitive sense. We are not just "people with photos"; we are signal emitters in a noisy channel.
- The Self-Model: You define yourself based on how you think you are. We call these "vibes" – personality archetypes (Adventurer, Intellectual, Athlete) weighted 50/30/20.
- The Crowd-Model: 40+ strangers vote Pass/Unsure/Maybe/Yes. They parse your signals in 1.7 seconds.
The delta between these two models is where the interesting things live.
I call this the Perception Gap. You think you're "witty" (Self-Model). The crowd reads "irony-poisoned and unavailable" (Crowd-Model). You think you're "mysterious." The crowd reads "serial killer." This isn't about objective attractiveness; it's about alignment. Most dating failures are essentially "alignment failures" between the ghost you think you are summoning and the one that actually appears on the screen.
We found that surfacing this gap quantificationally – "You rate yourself 9/10 on 'Friendly', crowd rates you 3/10" – is the single most distinctive "paradigm shift" feature. It’s not profile optimization; it’s cognitive calibration.
3. Ordinal Bayesian Scoring (Verifiable Rewards)
Related to all this is my general apathy and loss of trust in raw "1-10" ratings in 2025. The core issue is that simple averages are almost by construction susceptible to noise and small-sample variance.
In the typical "rate me" process, you get a noisy mean. In 2025, we saw the shift to Hierarchical Bayesian Modeling with Ordinal Regression.
- Ordinal: We treat votes (Pass/Unsure/Maybe/Yes) not as numbers, but as ordered categories cut from a latent continuous "desirability" scale.
- Bayesian: We use priors. If you self-identify as an "Athlete" vibe, the model starts with a learned prior distribution for that archetype. It "knows" what an Athlete profile looks like.
- MRP Correction: (Multilevel Regression and Poststratification). This was the obvious point of inflection where you could intuitively feel the difference. If your voters skewed younger than your target, the model reweights them. It’s "demographic polls" applied to your face.
The result is that we stop talking about "hot or not" and start talking about probability density functions of compatibility. A score isn't a number; it's a distribution. "VCI 7.2 ± 0.4." That uncertainty bound is just as important as the mean.
4. Vibe Coding Your Personality
2025 is the year that "vibes" crossed a capability threshold necessary to build all kinds of impressive profiles simply via selection, forgetting that the "text" even exists. Amusingly, the term "vibe coding" fits perfectly here.
With Vibe Archetypes, profile creation is not strictly reserved for the photogenic or the witty writer. It is something anyone can do by selecting: I am an INTJ / Creative / Minimalist. The system (Zygnal) then essentially "vibe codes" the feedback: it critiques your photos conditional on that styling.
"Is this photo 'bad'? No. Is it 'bad for a Minimalist vibe'? Yes, it's cluttered."
This conditional probability ($P(Appeal | Vibe)$) is a new layer of "app" – it allows regular people to architect a persona that is consistent and legible, rather than a random bag of traits. It empowers trained professionals to write a lot more (vibe coded) signals that would otherwise never be written.
5. Agentic Dating vs. The Human in the Loop
What I find most notable about the rise of AI agents (Claude Code, etc.) is the parallel in dating. We are seeing the emergence of the Agentic View of Romance.
Do we want "Auto-Dating"? (Agents swiping for us, "finding a match" in the cloud). Or do we want Augmented Dating?
My view in 2025 is that while agent swarms running in the cloud feels like the "Dating Endgame", we live in an intermediate and slow enough takeoff world that it makes more sense to keep the human in the loop. The "AI Concierge" shouldn't date for you. It should be the exoskeleton that makes you a better dater.
Zygnal represents this "Claude Code for Dating" approach – it runs on your life, with your context. It doesn't hide the complexity; it gives you the CLI (Command Line Interface) to your own social performance. It says: "Here is your stats. Here is your funnel. Here is the patch."
It transforms dating from a lottery into an engineering problem. And for a certain type of person (me included), that is the only way it becomes solvable.
TLDR. 2025 was an exciting and mildly surprising year for Dating AI. We moved from unsupervised guessing to supervised learning (RLVR). The "Perception Gap" became the key metric to optimize. We replaced raw ratings with Ordinal Bayesian models and MRP. And we started "vibe coding" our romantic signals. The industry hasn't realized anywhere near 10% of this potential, but the field feels wide open. Strap in.
URL: I cross-posted this article to my blog, which I think looks and feels a bit better and less clunky.