AI Daily — June 18, 2026
2026-06-18
openai.com
A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry
OpenAI and Molecule.one deployed a near-autonomous AI chemist built on GPT-5.4 that successfully improved a difficult synthetic chemistry reaction relevant to drug development. The system operated with minimal human intervention, iterating on reaction conditions in a closed-loop fashion. This is one of the more concrete demonstrations of AI driving real wet-lab chemistry outcomes rather than just predicting molecular properties.
openai.com
Introducing LifeSciBench
OpenAI released LifeSciBench, a benchmark designed and reviewed by domain experts to evaluate AI on realistic life science research tasks and decision-making. Unlike existing biology benchmarks that lean on factual recall, this targets applied research workflows and judgment calls. It is intended to track progress as models are increasingly deployed in scientific contexts.
huggingface.co
GLM-5.2: Built for Long-Horizon Tasks
Zhipu AI released GLM-5.2, a model explicitly optimized for long-horizon agentic tasks requiring sustained multi-step reasoning and planning. The release is hosted on the Hugging Face Hub, making weights accessible to the open research community. It positions itself as a competitor in the growing segment of models targeting complex, multi-turn agent workflows rather than single-turn instruction following.
huggingface.co
From the Hugging Face Hub to robot hardware with Strands Agents and LeRobot
Amazon and Hugging Face published a joint integration showing how policies trained and shared via the Hub can be deployed to physical robot hardware using AWS Strands Agents and the LeRobot framework. The pipeline bridges the gap between model sharing and real-world actuation with minimal custom glue code. This represents a meaningful step toward standardizing the robot policy deployment stack.
rss.arxiv.org
JetFlow: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting
JetFlow proposes a speculative decoding approach that addresses the scaling ceiling where increasing draft budget stops yielding speed gains due to the causality-efficiency tradeoff in existing drafters. It introduces parallel tree drafting that combines path-conditioned acceptance quality with lower overhead than autoregressive tree methods. The work is relevant to anyone deploying large LLMs where inference latency is a primary constraint.