Synthetic data's core flaw: embedded biases speed up overfitting and drive mode collapse

gerardsans · x · 2026-10-21

Responding to the synthetic data debate, gerardsans argues: synthetic data is the product of an inference pass, so outputs carry biases embedded from the training corpus. Fine-tuning on sampled data accelerates overfitting — a strict loss of information that leads to mode collapse.

This is a core technical objection to the "models training on model outputs"路线 in current scaling debates.

Related event: Researcher argues synthetic data accelerates overfitting and can't enable recursive self-improvement(3 posts)→

Original post →

More from Research

• Recursive generalization makes neurons 'clean' — agent swarms could crash alignment costs to 7 figures — daniel_mac8 · 2026-10-21
• OXR: 1.3B Open-Source Parser Tops Three Doc Benchmarks at 3.05x GLM-OCR Throughput — rohanpaul_ai · 2026-10-21
• Python 3.15 ships Tachyon profiler with 1ms sampling and live flame graphs — KhuyenTran16 · 2026-10-21
• Stanford's full CS336 LLM course is on YouTube — homework includes FlashAttention in Triton — techNmak · 2026-10-21
• PhysCanvas world model predicts vision and touch, lifting sim-real correlation to 0.89 — stepjamUK · 2026-10-21
• SCAMPER: a virtual rat with realistic motion for RL-based neuroscience experiments — tyrell_turing · 2026-10-21

Research channel →