Can AI Self‑Improve? Inside a Stanford PhD Defense on Continually Self‑Improving AI

Yesterday Zitong Yang, a PhD student at Stanford University, defended his dissertation titled Continually Self‑Improving AI . The defense video quickly went viral, and the committee included Stephen Boyd, Percy Liang, Emmanuel Candès, Tatsunori Hashimoto, and former Meta researcher Pang Ruoming.

Definition and Desired Properties

The author defines a continually self‑improving AI system as one that, once created, can autonomously and continuously improve itself, achieving better performance than any further improvements made by its human creators. The definition rests on two assumptions: the system is a parametric neural network whose knowledge is stored in explicit weights, and it has undergone a resource‑intensive pre‑training phase.

Three target properties are identified:

After the initial pre‑training, the system can acquire new knowledge without catastrophically forgetting old knowledge.

The system can generate its own training signal, and learning from this self‑generated signal yields larger gains than learning from human‑generated data.

The system can autonomously design its own learning algorithm.

Motivation: Human Limitations

The motivation stems from three human‑centric limitations:

Model weights become static after training.

High‑quality human data are scarce and will be exhausted as models grow.

Discovery of new algorithms relies heavily on human effort, limiting the algorithmic search space.

Research Direction 1 – Synthetic Continual Training

The first direction introduces a synthetic continuing paradigm that uses entity‑graph synthetic data generation to keep a model learning niche‑domain knowledge after pre‑training while avoiding catastrophic forgetting.

Experiments use the QuALITY dataset (265 books, ~1.8 M tokens, ~4 000 multiple‑choice questions). Baselines include a static Llama‑3 model (39 % accuracy), GPT‑3.5 (44 %) and GPT‑4 (45 %). Simple rewriting of source documents yields modest improvements, but the EntiGraph method—extracting entities from source texts and prompting the model to describe relationships—produces larger gains (up to ~56 % closed‑book accuracy and ~60 % open‑book accuracy).

Research Direction 2 – Synthetic‑Guided Pre‑Training (SBPT)

The second direction proposes Synthetic Bootstrap Pre‑Training (SBPT) , a three‑step pipeline:

Pre‑train a language model from scratch on a fixed amount of real data.

Fine‑tune the checkpoint to become a synthetic data generator without adding new text.

Combine real and synthetic data for a second round of pre‑training.

Comparisons are made against a baseline that simply repeats the same data and an “Oracle” that has unlimited real data. Across token budgets (200 B and 1 T) and model sizes (3 B and 6 B parameters), SBPT consistently achieves lower OpenWebText2 loss and higher downstream QA accuracy. Fact‑error rates drop from ~50 % at 200 B to < 10 % when scaling compute or model size, showing that higher‑quality synthetic data can approach real data performance.

Research Direction 3 – AI‑Design‑AI

The third direction builds a research environment that supplies the model with (i) a code‑base context and (ii) a value function that maps a generated idea (a string) to a numeric reward.

A “researcher” agent follows four steps: (1) ingest the environment, (2) generate an idea via a “thinker”, (3) produce a code diff, (4) execute the diff on allocated GPU resources and receive the reward. Demonstrations include:

A mathematical‑work‑memory buffer idea for a GRPO‑based math‑reasoning task, improving few‑shot accuracy by ~10 %.

Test‑time search methods (serial vs. parallel) that further boost performance, achieving 69 % accuracy on a post‑training task—surpassing the previous human benchmark of 68 %.

Philosophical Reflection

The author draws an analogy to Einstein’s field equations, arguing that just as a theory can out‑predict its creator, AI systems have the inherent potential to evolve beyond human intelligence. Empirical results provide concrete pathways—synthetic data, self‑generated training signals, and autonomous algorithm design—while acknowledging that current advantages still rely heavily on scale and relentless experimentation.