The entire pipeline for building artificial intelligence has fundamentally shifted: what used to require human judgment, data collection, and manual curation now runs on models training models, with synthetic simulations replacing real-world experiments at every stage. This isn't happening gradually. Each year since 2022, another critical component of AI development has flipped from human-made to model-made, creating a cascade where judges, teachers, environments, and even the people being studied are increasingly synthetic.

What Exactly Is Happening to AI Training?

The transformation started with something counterintuitive: the judge. InstructGPT, released by OpenAI, established a now-standard approach where human preferences are collected once, then a reward model learns to predict what humans would approve of. From that point forward, the AI being trained optimizes against another AI's judgment, not against humans directly. Constitutional AI pushed this further by having models critique themselves against a set of principles, and research showed that AI feedback could match human feedback at a fraction of the cost.

The next domino to fall was training data itself. Microsoft's Phi series demonstrated that a small model trained on high-quality, AI-synthesized data could outperform much larger models trained on raw internet text. Apple's WRAP approach took this further by having an LLM (large language model) rephrase the entire web before pretraining, improving efficiency by roughly 3 times. By 2025, reasoning traces (chains of thought generated by strong reasoners) had become a standard ingredient in pretraining, meaning the corpus of training examples was now substantially written by other models.

DeepSeek-R1 accelerated this trend by shipping distilled models alongside its flagship reasoning model, making "the teacher is a model" the default assumption for every small model release since. The technique matured from simple imitation into a proper training discipline, with on-policy generalized knowledge distillation fixing the train-inference mismatch.

How Are Models Now Training Themselves?

The loop is closing on itself as models begin deciding what to learn next. Self-Instruct and STaR, both from 2022, showed that models could write their own instruction sets and bootstrap their own reasoning traces. But the real flip came when Meta's Self-Rewarding Language Models and SPIN demonstrated that a model could generate its own tasks, judge its own outputs, and improve past the ceiling of human preference data. Curriculum design, historically the most artisanal and taste-driven part of machine learning, became something models do to themselves.

The breakthrough moment arrived with Karpathy's autoresearch in March 2026: a 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 extended run stacked 700 experiments into 20 kept improvements, cutting time-to-GPT-2 from 2.02 to 1.80 hours, finding real, transferable code changes while he slept.

Steps to Understanding the Synthetic Simulation Pipeline

• Stage 1 - Synthetic Judges: Reward models and AI feedback systems replace human evaluators, allowing models to optimize against other models rather than waiting for human approval.
• Stage 2 - Synthetic Data: LLMs generate training examples, rephrase web content, and create reasoning traces, making the corpus of training material substantially model-written rather than human-collected.
• Stage 3 - Synthetic Teachers: Distillation becomes standard, with large models teaching smaller ones, making "the teacher is a model" the default assumption for all small model releases.
• Stage 4 - Synthetic Curriculum: Models generate their own tasks, judge their own outputs, and improve past human preference ceilings, automating what was once the most artisanal part of ML training.
• Stage 5 - Synthetic Environments: Z.ai built pipelines that synthesize entire task environments end-to-end, with research agents mining real work patterns and converting them into long-horizon environments with hidden state.

The environment bottleneck is now the critical constraint. Reinforcement learning (RL) needs thousands of executable, verifiable, professionally realistic task worlds, and humans cannot hand-build them fast enough. Z.ai's approach synthesizes environments entirely: research agents mine real work patterns and convert them into long-horizon environments, a judge agent attempts each task to confirm it is solvable, and verifiers are synthesized without seeing the reference solution, then stress-tested until their binary reward is reliable enough to train on directly. The entire environment, judging, and verification stack is synthetic all the way down.

Ornith-1.5 shipped claiming end-to-end self-improvement, where the model proposes its own tasks and generates its own RL rollouts. The gym, the referee, and the scoreboard are all models now.

Where Does This Leave Human Input?

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 is the layer Simile is replacing. The lineage runs from Joon Sung Park's Generative Agents research 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.

Frontier models are trained toward being agent models, which makes them poor simulations of real people. Simile post-trains on interviews, transaction data, and registered randomized controlled trials (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.

What Does This Mean for the Future of AI Development?

The trade-off is clear: synthetic approaches are roughly 10% worse in accuracy but 100 times cheaper and 10,000 times faster than their human-made counterparts. This is not a temporary optimization; it represents a fundamental restructuring of how machine intelligence gets produced. Every year since 2022, one more component of the pipeline has flipped from human-made to model-made, and each flip has a patient zero, a paper or product where the synthetic version first became load-bearing at a frontier lab.

Poolside's strategic positioning reflects this shift. The company's reverse-execuhire letter drew a precise line: the world's problems split into intelligence-bound ones, solvable by scaling cognition and soon commoditized by open weights, and experiment-bound ones, where "no amount of intelligence substitutes for real-world experimental feedback." 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 biological sciences are 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 1,000 times cheaper and faster than in vivo. The physical world is the one component that cannot be fully synthesized, only compressed, cell by cell, into models. This convergence suggests that the future of AI development is not about building bigger models, but about building better simulations of the systems those models need to understand.