Asari AI’s co-inventor agents optimized the full vLLM serving stack for DeepSeek v4 Pro and GLM 5.2 on NVIDIA B200s, improving throughput and interactivity by up to 16% across concurrency levels while preserving model behaviour through distribution-level correctness checks (Asari blog). In the same issue Intology’s Locus took the top of PostTrainBench by post-training models unsupervised in ten H100-hours and beating human tuners on the harder variant (Intology blog), overtaking the Sol post-trains Luna result of July. Together they extend the loop from production kernels and self-evolving harnesses to the serving stack and the post-training pipeline, two more of the layers between a model and its own next version; four days later Poetiq’s self-optimizing optimizer makes the same claim for the harness layer.
Development
Self-improving agents rebuild the inference stack serving two open models
The Singularity filing its own optimization tickets: the loop reaches the serving layer, the last piece of infrastructure between a model and its users, and the post-training pipeline in the same week.
Claim
Asari AI's self-improving co-inventor agents rebuilt the inference stack for two open models on B200s, lifting throughput and interactivity up to 16%, while Intology's Locus agent led PostTrainBench by post-training models unsupervised in ten H100-hours and beating human tuners on the harder variant.
Reported metric: up to 16%
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to August 4, 2026 (theinnermostloop.substack.com)
- Speeding up end-to-end inference with self-improving agents (asari.ai)
- Scaling Automated Post-Training (intology.ai)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.