Sam Altman Confirms OpenAI Has Experimented with Pure Synthetic Data Training Models
OpenAI CEO Sam Altman stated in a podcast with The Atlantic's CEO Nicholas Thompson that GPT-4 is the last model trained with almost no AI-generated data.
Thompson pointed out that the internet is flooded with AI content, to which Altman nodded in agreement. When asked if they had run a model trained entirely on synthetic data (AI output to train the next generation), Altman hesitated and hinted that experiments had been conducted.
Altman believes that the core of the model is reasoning ability, and pure synthetic data can fully achieve tasks like mathematics, even surpassing humans; however, understanding human values and culture still requires real human data.
Source: Public Information
ABAB AI Insight
Sam Altman has previously advocated for the scaling of synthetic data. Earlier, OpenAI had used models like GPT-4o to generate training data. This podcast indirectly confirms experiments with a closed loop of pure synthetic data, continuing the shift from "human data dependence" to "synthetic data dominance" in training, especially in the context of increasingly scarce high-quality human data.
From a capital perspective, OpenAI reduces reliance on external data collection and labeling through synthetic data while retaining core human cultural data for alignment training, forming a mixed cost structure of "massive synthetic reasoning data + selected human values data," significantly compressing the training costs of the next generation of models and accelerating iteration cycles.
Similar to how Google DeepMind surpassed humans with AlphaGo through self-play using pure synthetic data, OpenAI is in the mid-stage transition from "human data dominance" to "synthetic data dominant reasoning + human data for alignment." Altman clearly states that tasks like mathematics can be purely synthetic, while values require human data.
Essentially, this represents a technological substitution: the traditional training paradigm relying on massive amounts of real human text data is being replaced by a closed loop of synthetic data. Altman's experiments show that reasoning ability can be achieved through AI self-iteration, using human data only for cultural and value anchoring, reconstructing AI training from "data hunger" to an efficiency mechanism of "synthetic bootstrap + a small amount of human refinement."