Today's edition highlights systems that build their own target knowledge, from agents learning new hardware DSLs to models monitoring their own training logs to refresh their task pools.
The loop: A-RSI monitors its own training logs to decide when to refresh a task pool with model-authored programs, enabling continuous multimodal self-improvement without external data.
A multimodal self-improvement framework that monitors its own training dynamics to decide when to generate and introduce new model-authored tasks.
The loop: An LLM agent loop builds a target knowledge base of DSL documentation and microbenchmarks to autonomously optimize decode kernels, which run the models that power the agent.
A multi-agent loop that builds its own knowledge base of a new AI accelerator's DSL and microbenchmarks to autonomously optimize decode kernels.
The loop: An AI agent edits its own harness prompts and rules, using an automated gate to measure and retain improvements to its own operating environment.
A study revealing how automated acceptance gates in self-improving agents can silently discard redundant safety safeguards because overlapping protections mask the loss.
The loop: The SCOPE policy analyzes its own trajectories to generate step-specific hindsight skills, distilling their guidance back into itself at selected error steps.
A framework where an agent analyzes its own trajectories to generate and distill step-specific hindsight skills at critical error steps.
The loop: Sakana Marlin uses multiple AI agents to parallelize hypothesis search and evaluate each other, improving the accuracy and depth of the research reports it generates.
An intelligence report on the US midterms generated by an AI system that parallelizes hypothesis search and uses multiple agents to evaluate each other's findings.
The loop: The auto repair system detects and fixes GPU and fabric faults, which maintains the availability of the hardware used to train and run the system's successors.
Together GPU Clusters auto repair detects GPU, host and fabric faults, picks the lightest fix, and lets you decide per fault and per Xid code which repairs run automatically. Here is how it works.
