KISSD / STRATEGIC CONCEPT / MARCH 2026

HONDA AD LAB

RECURSIVE SELF-IMPROVING ADVERTISING

A framework for applying autonomous AI experimentation to Honda's paid social — generating a simultaneous 10x boost in performance and 10x reduction in costs.

Scroll
01 / THE OPPORTUNITY

CREATIVE IS
BECOMING THE
ONLY LEVER
ADVERTISERS
CONTROL.

Meta is systematically automating every lever advertisers once controlled. By late 2026, targeting, bidding, and delivery are fully machine-managed through Advantage+ and its successors. The only remaining competitive advantage is the creative itself — what you show, how you show it, and how fast you learn what works.

In March 2026, Andrej Karpathy released Autoresearch — a system that ran 700 experiments in 2 days, landed 20 optimizations, and achieved an 11% performance gain on a neural network. Shopify's CEO immediately directed his teams to study the pattern. The approach is deceptively simple: give an AI agent something to change, something to measure, and time to run.

Honda Ad Lab applies this exact pattern to Honda's paid social advertising.

02 / THE PATTERN

THREE
PRIMITIVES.
ONE LOOP.

Autoresearch reduces autonomous experimentation to three structural primitives. Each one maps directly to advertising.

Primitive Autoresearch (ML) Honda Ad Lab
Editable Asset train.py — 630-line training script Ad creative package: headline, copy, CTA, image/video, format
Scalar Metric val_bpb — validation bits per byte ROAS, CPA, or composite score from Meta Ads API
Time-Boxed Cycle 5 min per experiment (~100 overnight) 24–48 hour test windows per variation

Karpathy also identified a critical fourth element: the fixed component. In Honda Ad Lab, the fixed components are Honda's brand guidelines, legal requirements, approved vehicle imagery, and campaign objectives. The agent can change everything creative — but it cannot violate the brand.

03 / THE DIFFERENCE

NOT ANOTHER
AD TOOL.

Current Tools
Honda Ad Lab
What they test
Predefined variations you create
Novel creative the agent generates
How they optimize
Rule-based: pause losers, scale winners
Hypothesis-driven: reasons about WHY and designs next test
Creative generation
Template-based or human-made
AI-generated copy, images, video — informed by performance data
Experiment velocity
5–15 variations per cycle
50–200+ per month, continuously
Learning
Each test is independent
Cumulative knowledge base — a proprietary map of what resonates
"Neural architecture search as it existed then is such a weak version of this that it's in its own category of totally useless by comparison." — Andrej Karpathy
04 / THE SYSTEM

THREE
LAYERS.

Layer 1 / The Agent Loop
Strategy → Create → Deploy → Evaluate → Repeat
program.md → Agent modifies ad_creative → Deploy to Meta → Evaluate results → Keep or revert → Repeat
Layer 2 / AI Creative Generation
Language → Image → Video → Format
LLM copy generation → kissdGen / Nano Banana Pro imagery → Seedance / Kling video → Format variation
Layer 3 / Deployment + Measurement
Publish → Optimize → Gate → Review
Meta Marketing API → Advantage+ / CBO → Statistical significance gating → Brand safety review
05 / THE MATH
0×
Performance
  • 50–200+ creative variations per month vs. 5–10 today
  • Compounding learning — each test makes the next one smarter
  • Creative drives 70% of campaign results on Meta
  • Runs continuously — nights, weekends, holidays
0×
Cost Reduction
  • Near-zero marginal cost per creative variation
  • No shoots, no revision rounds, no vendor timelines
  • Human effort shifts from production to strategy
  • 80–90% of current test spend is wasted on losers — agent reduces failure rate over time
06 / THE EXPERIMENT

60-DAY
PROOF OF
CONCEPT.

Phase 1
Build
Weeks 1–3
  • Write strategy brief
  • Build agent loop
  • Establish baseline
  • Define scalar metric
Phase 2
Run
Weeks 4–7
  • Launch with human review
  • Daily creative generation
  • 100+ experiments in 4 weeks
  • Weekly Honda reviews
Phase 3
Evaluate
Week 8
  • Performance vs. baseline
  • Full experiment log
  • Cost analysis
  • Scale recommendation

Self-evaluation mode: Even before real engagement data, the agent can run in self-evaluation mode — generating and ranking hundreds of creative variations against Honda's brand guidelines in minutes, before any ad dollar is spent.

07 / WHY KISSD

THREE
CAPABILITIES.
ONE TEAM.

01
Deep Honda
Brand Knowledge
Established Honda creative relationship. We understand the brand voice, product positioning, and approval culture.
02
AI-Native Creative
Production
kissdGen, Nano Banana Pro, Seedance — custom AI pipelines built for automotive advertising.
03
Social-First
Expertise
We live on Meta platforms. We understand the algorithm, creative fatigue, format and placement.

From creative production company to intelligence-powered creative studio. Honda Ad Lab is the most concrete expression of that evolution.

08 / THE ASK

R&D
INVESTMENT.
60 DAYS.

Honda Provides
R&D budget, Meta API access, an always-on campaign, a designated stakeholder
Kissd Delivers
Working experiment system, 60 days of data, full performance analysis, experiment log, scale recommendation
Success Looks Like
Measurable ROAS improvement, 10x experiments, agent-discovered insights, clear scale path
"The companies that win won't have better marketers. They'll have faster experiment loops." — Eric Siu
"You spin up a swarm of agents… and humans optionally contribute on the edges." — Andrej Karpathy
KISSD / INTELLIGENCE-POWERED CREATIVE STUDIO
Confidential