The following is a list of personally notable and mildly surprising observations about dating in the age of AI - things that altered my thinking and stood out to me conceptually.
1. The Missing Supervision Layer
At the start of 2025, the "dating stack" for most people looked something like this:
- Pretraining (living your life, accumulating photos ~ongoing)
- Supervised Finetuning (asking a friend "does this photo look good?" ~biased)
- Zero-Shot Inference (uploading to Hinge and hoping ~random)
This has been the stable and proven recipe for staying single for a while. The core issue is that dating is an unsupervised learning problem. You upload photos, you get random feedback (silence or noise), and you try to converge on a "self" that works. There is no loss function. There is no gradient. You are training against nothing.
We have entire industries for skill development. Coding has bootcamps, linters, compilers. Fitness has trainers and heart rate monitors. Investing has advisors and dashboards. But find a life partner - arguably the highest-leverage decision a human makes? You're on your own.
Here's the brutal math. 0.08% success rate - 1 relationship from 1,400 swipes. 1.7 seconds average attention on your profile before the left-swipe. 68% of people lead with the wrong photo - not their worst, their wrong photo. The one that converts at 3% instead of the candid at position #6 that converts at 8%.
Same face. Completely different outcomes.
The apps won't help. A user who figures it out leaves. A user who stays confused keeps swiping - and paying.
2. The Perception Gap
This is the conceptual core. We built a forensic profile audit engine - but it's really a behavioral engine.
You define yourself based on how you think you are. We call these "vibes" - personality archetypes (Adventurer, Intellectual, Athlete, Creative, etc.) weighted 50/30/20. You build a self-model.
Crowds + AI tell you how you actually come across. 40+ strangers vote Pass / Unsure / Maybe / Yes - the actual swipe decision mapped to an ordinal scale. No fake precision. Real intent.
The delta between these two models is where the interesting things live.
You think you're "witty." The crowd reads "irony-poisoned and unavailable." You think you're "adventurous." The photos signal "unstable and never home." This isn't about being more attractive - it's alignment. Your internal model and your external signal are out of sync.
We surface the gap. We quantify it. That's not profile optimization - that's self-knowledge.
3. The Statistical Engine
Here's what runs under the hood:
4-level voting (Pass / Unsure / Maybe / Yes) → ordinal regression with cumulative link. Cutpoints learned offline; latent signal η inferred online. Converts to a 1–10 VCI (Vibe Compatibility Index) via probability-weighted anchors [1, 4, 7, 10].
Vibe priors → each archetype carries a learned offset. No gender in priors - baseline comes from personality markers, not chromosomes. Stabilizes early scores with limited data.
Reliability weighting → time decay (90-day half-life), voter calibration via Brier scores, Effective Sample Size. 100 spam taps ≠ 100 thoughtful judgments.
MRP correction → Multilevel Regression & Post-stratification (borrowed from political polling). If this week's voters skew young when you're targeting 30s professionals, we reweight to your actual market.
Attribution modeling → "Main photo: 62% of appeal." We estimate how much the primary drives the score vs. the full profile. If λ is high and your lead image is weak, one swap moves the whole number.
Polarization scoring → entropy of the vote distribution. High safe score = consistent brand. Low safe score = split audience. Risky but powerful when targeted.
The math runs deterministic. LLM calls are small and surgical - two photo analyses, one evidence-gated synthesis. No hallucinated numbers. Every line in the report chains to drivers, badges, and pre-computed actions.
4. One Move That Moves the Needle
Most profiles don't need a personality transplant. They need a photo order correction plus a bio de-sabotage.
- Swap your true winner to primary. Typical lift: +0.8 to +1.5 VCI.
- Rewrite the first 150 characters. Remove pre-rejection; lead with one specific hook. Typical lift: +0.3 to +1.0 VCI.
Ten seconds of work. Measurable change.
We don't overwhelm with 50 suggestions. We identify THE single highest-impact change: "Swap photo #3 to primary: +0.8 VCI expected." Each recommendation includes expected lift based on historical A/B data.
5. Why This Matters Now
By 2026, experts predict AI "situationships" dominating—your digital concierge "dating" others' AIs, filtering incompatibilities in seconds. Bumble's founder has publicly envisioned this. Searches for "AI boyfriend/girlfriend" are up 525% YoY.
The question: Do we build human-grounded feedback systems now, or let bots fully automate love?
Dating apps are the marketplace - 39% of couples met there. Markets have information asymmetry. The apps have all the data. Top performers cracked the code. Everyone else is playing poker in the dark.
We built the supervision layer that should have existed from the beginning.
TLDR. Your profile is a signal processor in a noisy market. 68% lead with the wrong photo. The Perception Gap - delta between who you think you are and how you come across - is where the insight lives. Zygnal's ordinal-Bayesian, reliability-weighted, MRP-corrected engine pulls truth from chaos and tells you the one move that matters. Stop guessing. Start knowing.
