#optimization Tech News.

Compilers translate high-level model code into hardware instructions. For decades, they relied on hand-crafted heuristics written by experts. Machine learning is now replacing those manual rules, automating decisions like function inlining to make AI models run faster.

Researchers applied LLM agents to multi-objective SMILES optimization in evolutionary chemistry. The approach uses agents to guide molecular generation, balancing multiple chemical properties simultaneously. This method improves upon traditional rule-based techniques for designing molecules with desired characteristics.

Datadog built DODO (Datadog Observability-Driven Optimizer) to ground AI code optimization benchmarks in live production telemetry. DODO uses CPU profiles and real production call samples to generate Go micro-benchmarks with ≥98% similarity to production. Three optimizations cut over 8% of a critical service's total CPU cost, saving O(10k) cores around the clock.