OpenAI alumni launch $100M Zero Shot fund

Substack

Ex-DeepMind researchers launch Elorian to build advanced visual AI systems, claiming current models lack sophisticated image understanding needed for real-world applications in automotive, architecture, and robotics.

Substack

Former OpenAI executives have launched Zero Shot, a $100M venture fund that has raised $20M and made early investments in AI automation and robotics startups, leveraging insider knowledge to identify promising opportunities while avoiding overhyped AI sectors.

Anthropic

Anthropic's Managed Agents decouples agent architecture into modular, independently-replaceable components (brain, hands, session), enabling more reliable long-horizon AI task execution while reducing latency by up to 90% and improving security and flexibility for enterprise deployments.

UC Santa Barbara, MIT CSAIL, and MIT-IBM Watson AI Lab

This comprehensive study reveals that LLM agent skills are fragile under realistic conditions: when agents must retrieve from 34k skills rather than receive curated ones, performance degrades significantly and approaches no-skill baselines. Query-specific refinement can recover much of this gap when relevant skills exist, but cannot compensate when they're absent entirely.

UNC-Chapel Hill, University of California, Santa Cruz, and University of California, Berkeley

ClawArena is a new benchmark evaluating AI agents in realistic, evolving information environments with conflicting sources and implicit user preferences across 64 scenarios. Experiments show model capability matters more than framework design (15.4% vs 9.2% performance range), though self-evolving frameworks can partially close capability gaps.