Google AI Updates: October 18, 2026

1. Google Research Publishes Three Papers on Connecting AI Agents to Physical Wet-Lab Experiments

Google. Google Research described three projects that link AI agents to real lab work, each with an accompanying paper. VERID is a semantic digital twin of a lab workcell built on the open-source PyLabRobot framework, providing live 3D state tracking, virtual dry runs, and hardware safety checks that block invalid commands. In an aptamer design sandbox, agents ran three closed-loop design-test rounds of about 7 to 10 days each, reaching a 42.1% confirmed on-target hit rate at 98.5% specificity for alpha-thrombin by the third round. A third project on binder design for the tumor-associated protein FAP sped up each structure prediction cycle by 8.6x by excising the local binding pocket, and a human-agent hybrid nearly doubled the in-silico hit rate of a scientist working alone at equal compute. Source

2. Google Cloud Publishes Serving Recipes for Reflection’s Beam on llm-d and NVIDIA Dynamo

Google. Google Cloud announced day 0 support and ready-to-run deployment recipes for Reflection’s Beam, a roughly 501B-parameter (about 23B active) Apache 2.0 open-weight model for coding and agentic tasks. The recipes cover NVIDIA Dynamo with SGLang on A4X Max (GB300 NVL72) instances and llm-d with vLLM on G4 VMs (RTX PRO 6000), using KV-cache-aware and prefix-aware routing, prefill/decode disaggregation, speculative decoding, and tiered KV cache offloading to host DRAM and local SSD. Based on customer benchmarking, Google reports up to 5x higher throughput and 8x lower latency on A4X Max versus a standard Kubernetes deployment with stateless round-robin routing, and up to 2x throughput on G4. Beam is also available as a managed option in Model Garden. Source