Automated Experiments Have a Closed Loop; Compounding Gains in Intelligence Remain Unproven

First, Distinguish Two Kinds of Progress

The WeChat public account proposes that recursive self-improvement (RSI) and continual learning could form a mutually reinforcing flywheel: models help develop better models, while experience from use becomes another source of capability growth. This idea deserves discussion, but assessing whether it holds requires distinguishing automation of experimental workflows from improvements in learning outcomes. autoresearch demonstrates a mechanism for agents to conduct experiments; the continual learning survey calls for examining learning on new tasks, retention of existing capabilities and transfer performance. This suggests that the two could help each other, but the number of automatically executed experiments cannot substitute for learning outcomes.

Evidence: [5] [1] [4]

A Closed Loop Depends on Clearly Defined Evaluation Boundaries

Karpathy’s autoresearch limits each experiment to a single GPU and 5 minutes of training wall-clock time, excluding startup and compilation. The agent may modify only the training file, leaving data preparation and the fixed evaluation unchanged. Changes are kept if the validation metric improves and reverted if it stays the same or worsens, though a simpler implementation may also be kept when results are equivalent. This shows that agents can advance experiments continuously in an environment with clear objectives and permissions. My judgment is that this mechanism can reduce manual intervention in each experiment, but it does not directly measure how many useful discoveries are produced per unit of compute. Even if the metric keeps improving, whether those improvements carry beyond this evaluation setup still needs testing.

Evidence: [1]

Transfer to Production Sets Limits on Optimism

Anthropic’s automated alignment research provides a more concrete test. Agents propose methods, run experiments, analyze results and receive scores through a dedicated server. The best-performing method showed transfer on held-out mathematics and coding data, while the runner-up performed worse on coding tasks. When the best-performing method was tested in Claude Sonnet 4’s production training, it also failed to produce a statistically significant improvement. This does not show that automated research is generally ineffective: the production test covered only one idea, and the scoring was relatively simple. But it is enough to remind us that methods that win on controlled tasks remain candidates. These results suggest that the value of automated research depends on whether screening metrics predict real gains in the target environment. Expanding the search alone does not yet guarantee this.

Evidence: [2]

Online Updates Do Not Yet Mean Long-Term Accumulation

The original TTT paper makes the hidden state itself a learnable model, updated through a self-supervised loss within the test sequence, with each input sequence producing its own hidden-state weights. This provides a concrete mechanism for a model to use the current sequence. The paper mainly evaluates language modeling and long-context perplexity, and the fixed-size hidden state still faces practical constraints on computation time and memory reads and writes. The evidence from this work should therefore remain limited to within-sequence updates; it cannot be directly extended to retention of experience across sessions or general-purpose learning on unfamiliar tasks. Whether local adaptation leaves behind knowledge that remains useful over time still requires separately designed experiments.

Evidence: [3]

Assess Continual Learning by Both What Is Learned and What Is Lost

The WeChat public account treats whether experience can immediately become knowledge as the dividing line for continual learning, but this definition is too narrow. The continual learning survey allows new data to arrive in batches or as complete tasks. It focuses on sequential learning under changing data distributions and performance when old samples are unavailable or accessible only to a limited extent. Under this framework, more useful criteria are whether new experience improves subsequent tasks, whether existing capabilities are retained, whether learning transfers to new tasks, and whether storage and computation costs are acceptable. If a system claims to accumulate experience across sessions, retention should also be tested across that scope of use. My judgment is that only when these results hold together is there reason to interpret a local update as a capability gain that can accumulate; the speed of the update alone is not enough.

Evidence: [5] [4]

What to watch next

• Whether approaches selected through automated research can repeatedly deliver improvements on held-out tasks and in the target production environment, and specify the conditions under which they fail.
• Whether systems claiming to accumulate experience continually can demonstrate gains on new tasks, retention of existing capabilities and transfer across tasks together.
• Whether performance improvements come with acceptable computation and storage costs; when cross-session learning is claimed, whether experience continues to be useful in subsequent sessions.