Today's edition highlights the diversity of self-improvement loops, from optimizing distributed training infrastructure to rewriting the fundamental numerical algorithms that agents rely on.
The loop: AI agents perform post-training on a base model to improve its benchmark average. The resulting model is evaluated on a benchmark specifically designed to measure AI driven engineering tasks.
Claude Code achieved a new high score of 44.6% on PostTrainBench by successfully post-training a base model.
The loop: Micius-Flash generates synthetic data and reasoning sequences for hardware coding and semiconductor design. These outputs are used in a multi-teacher on-policy distillation process to train a unified successor model.
A 27-billion-parameter model generates synthetic data and reasoning sequences to train its successor via multi-teacher on-policy distillation.
The loop: The InfraRSI agent co-optimizes launch knobs and code patches for distributed LLM training. This improves the underlying training infrastructure the agent relies on.
An agentic harness turns human commands into verified training optimizations by co-optimizing launch settings and code patches for distributed LLM training.
The loop: An AI agent rewrites and redesigns numerical algorithms against a pre-specified metric in a guarded loop. The resulting efficiency gains improve the underlying mathematical tools the agent uses for optimization.
A coding agent autonomously rewrites and redesigns numerical algorithms in the ASE atomistic simulation library to run faster and converge with fewer evaluations.
The loop: A multi-agent loop generates, evaluates, and refines custom CUDA kernels to replace PyTorch operators. These kernels improve the execution speed of the model on its own hardware.
A study evaluates how different interaction strategies affect a local 7B parameter code model's ability to write custom CUDA kernels that outperform eager PyTorch.
The loop: The chain lens extracts multi-token concepts from a model's latent representations using contrastive decoding. This surfaces specific computational intermediates that can be used to oversee and refine the model's reasoning.
A new interpretability technique extends single-token methods by using contrastive decoding on an MTP head to recover multi-token concepts from model activations.
![The Chain Lens [interim]](/media/r-cca9717349b6fefe.webp?v=1794464581 "Image: lesswrong.com")
The loop: An AI system automates empirical alignment experiments, producing data and insights that are used to make the next generation of AI systems more aligned.
I used to think of "automating alignment" as a natural category. But when I tried to think through the strategic implications, I realised that it depends a lot on the type of work. I'm now thinking of it in three buckets: 1. Empirical...
