Stage 7 — The Human Subject (2025)

If models can be the judge, teacher, and environment, the remaining human

role in the loop is subject — the source of preferences, behavior, and

demand. That's the layer Simile is replacing.

• Generative Agents (Joon Sung Park et al., Smallville 2023) — a town of

LLM agents producing believable daily lives.

• Generative Agent Simulations of 1,000 People — digital twins built

from two-hour biographical interviews reproduced their source humans'

survey and behavioral responses 85% as accurately as the humans

reproduced themselves two weeks later.

• Simile — post-trains on interviews, transaction data, and registered

RCTs from the Open Science Framework specifically to recover human bias,

inconsistency, and causal texture, and reports early scaling laws for

simulation quality. The hurdle: frontier models are trained toward being

agent models, which makes them bad simulations of real people.

• SimGym (Shopify) — simulating shopper trajectories; the follow-up

("2,000 robots walk into a shop", Feb 2026) validated real serving

optimizations — partitioning, speculative decoding — from simulated

shoppers alone.

• Tencent's billion-persona approach — the crude end of the spectrum.

Mid-2026 was the stage's money quarter:

• Simile's valuation went vertical — $100M Series A (Feb 2026) to a $2B

valuation by July 2026 (Greenoaks-led Series B, ~20x in five months,

revenue up 5x). Synthetic subjects became an enterprise budget line.

• The competitor split — Simile requires real interviews to calibrate

each twin; rivals (e.g. Minds) generate personas from descriptions alone,

trading fidelity for instant availability on any audience.

• SimAB (arXiv, March 2026) — reframes A/B testing as fast,

privacy-preserving simulation with persona-conditioned agents.

• Delegation, not just measurement — Habermolt sends AI representatives

into public deliberation standing in for humans; Moltbook is a

Reddit-style platform populated entirely by autonomous agents with humans

only observing. The subject is simulated even when nobody asked a

question.

The focus group, the user study, the A/B panel — and now the town hall —

are becoming inference workloads.