模拟取代人工：AI训练管线八大阶段复盘

By AI standards today is a pretty quiet Friday, so it’s time to take a step back and reflect on what is really going on. If you read our 2025 reading list, and followed our coverage of Z.ai GLM, understood the Poolside pivot, been following our AI for Science themes, and tuned in to today’s Simile pod, you not only are one of the biggest readers of Latent Space, you will probably also arrive at this mental model:

Every year since 2022, one more component of the pipeline that produces machine intelligence has flipped from human-made to model-made. Not gradually, and not evenly — each flip has a patient zero, a paper or product where the synthetic version first became load-bearing at a frontier lab, and from there on, the future is simply here but not yet productionized.

And if you squint, what we used to call “synthetic data” and “synthetic rubrics” and “AI researcher” and “end to end RL environments” is just increasingly ambitious human simulation - 10% worse, but 100x cheaper and 10,000x faster.

Stage 1: The reward signal (2022)

The first thing to go synthetic was, counterintuitively, the judge. InstructGPT established the now-canonical trick: collect human preferences once, train a reward model, and let the policy optimize against the model rather than the humans. From the policy’s point of view, the thing dispensing approval was already an LLM. Constitutional AI pushed further and had the AI critique itself against a set of principles (RLAIF), and Lee et al. later showed AI feedback matching human feedback at a fraction of the cost. By the time LLM-as-judge became the default eval methodology (MT-Bench, AlpacaEval), the entire approval apparatus — reward, critique, evaluation — ran on models judging models.

Stage 2: The training data (2023)

Microsoft’s Phi series made the argument in its title: Textbooks Are All You Need. A small model trained on LLM-synthesized, textbook-quality data punched far above its parameter count, and phi-1.5 confirmed it wasn’t a fluke. Apple’s WRAP generalized the move: don’t just generate data, rephrase the entire web with an LLM, and pretraining gets roughly 3x more efficient. From there the pipeline industrialized — NVIDIA’s Nemotron-4 340B shipped with a permissively licensed synthetic data generation pipeline as a headline feature, and by 2025 reasoning-trace corpora (chains of thought generated by strong reasoners) had become a standard pretraining and mid-training ingredient. The corpus, the thing that was supposed to be the irreducibly human input, was now substantially model-written.

Stage 3: The teacher (2023)

Weeks after ChatGPT’s API opened, Stanford’s Alpaca demonstrated that a $600 fine-tune on GPT-generated instructions could clone much of a frontier model’s behavior. Vicuna did it with shared conversations; Orca did it with rich teacher explanations rather than bare answers. The technique matured from imitation into a proper training discipline — on-policy generalized knowledge distillation fixed the train/inference mismatch — and reached its cultural peak when DeepSeek-R1 shipped a family of distilled models alongside the flagship, making “the teacher is a model” the default assumption for every small model release since.

Stage 4: The curriculum (2024)

Stages 1–3 made the inputs synthetic; stage 4 is where the loop starts closing on itself, because the model begins deciding what to learn next. The pieces existed early — Self-Instruct (models writing their own instruction sets) and STaR (models bootstrapping their own reasoning traces) are both 2022 — but the flip came when Meta’s Self-Rewarding Language Models and SPIN showed a model could generate its own tasks, judge its own outputs, and improve past the ceiling of its human preference data. Curriculum design — historically the most artisanal part of ML, the taste-driven choice of what to train on next — became something models do to themselves.

Stage 5: The researcher (2026)

The assistance era (Copilot, then SWE-agents) kept a human choosing the experiments. The discovery era did not. DeepMind’s AlphaEvolve evolved genuinely new algorithms in 2025, and Sakana’s AI Scientist (now in Nature!) sketched the full paper-writing pipeline. The big moment was Karpathy’s autoresearch in March 2026: a deliberately minimal ratchet loop where a coding agent modifies a real LLM training setup, runs a five-minute experiment, keeps the change only if validation loss improves, and repeats overnight. His own extended run stacked 700 experiments into 20 kept improvements, cutting time-to-GPT-2 from 2.02 to 1.80 hours — real, transferable code changes found while he slept.

Stage 6: The environment (2026)

RL’s scaling bottleneck moved from the model to the environment: you need thousands of executable, verifiable, professionally realistic task worlds, and humans can’t hand-build them fast enough. We covered this recently in our z.ai / GLM-5.3 issue: Z.ai built pipelines that synthesize environments end to end — research agents mine real work patterns and convert them into long-horizon environments with hidden state, a judge agent attempts each task to confirm it’s solvable, and verifiers are synthesized without seeing the reference solution, then stress-tested with oracle, no-op, and unsolved-state checks until their binary reward is reliable enough to train on directly. As the GLM-5.3 release puts it, the entire environment, judging, and verification stack is synthetic all the way down. The same week, Ornith-1.5 shipped claiming end-to-end self-improvement — the model proposes its own tasks and generates its own RL rollouts. The gym, the referee, and the scoreboard are all models now.

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. The lineage runs from Joon Sung Park’s Generative Agents (Smallville, 2023) through Generative Agent Simulations of 1,000 People, where 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.

The big hurdle to overcome: frontier models are trained toward being agent models, which makes them bad simulations of real people — so 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. With SimGym at Shopify simulating shopper trajectories and Tencent’s billion-persona approach at the crude end of the spectrum, the focus group, the user study, and the A/B test panel are becoming inference workloads.

Stage 8: The physical world (2026, in progress)

The last row of the grid never quite turns red, and that’s the point. Poolside’s reverse-execuhire letter drew the line precisely: the world’s problems split into intelligence-bound ones (solvable by scaling cognition, soon commoditized by open weights) and experiment-bound ones, where “no amount of intelligence substitutes for real-world experimental feedback — 100,000 brilliant minds won’t cure cancer without a wet lab.” Their bet is that AI’s durable value accrues to whoever owns the experimental loop: AI as “the world’s most valuable scientific discovery engine.”

The bio side is running the same play from the other direction. CZ Biohub is imaging the Human Cell Atlas into a virtual cell — because in silico is roughly 1000x cheaper and faster than in vivo — and extending toward a virtual immune system, with Chai, Xaira, and Lila’s data-center-shaped labs filling in the AI-for-science stack. The physical world is the one component that can’t be fully synthesized — only compressed, cell by cell, into models.

The exponential starts at the diagonal