Company@X — 2026-06-18#

Signal of the Day#

OpenAI published a study in NEJM AI demonstrating that its o3 Deep Research model helped clinicians reanalyze 376 complex cases, resulting in 18 new diagnoses for previously unsolved rare pediatric diseases. This signals a massive leap in agentic AI capabilities, moving from conversational assistance to autonomously performing expert-led periodic reanalysis on highly complex, fragmented scientific data at scale.

Key Announcements#

[OpenAI] · Source In partnership with Boston Children’s Hospital and Harvard, OpenAI demonstrated that o3 Deep Research can successfully identify leads in long-unsolved neurodevelopmental and rare neuromuscular disease cases. The model synthesized clinical features, inheritance patterns, and the latest scientific literature into testable hypotheses for specialists, proving the viability of AI as a scalable diagnostic research tool.

[xAI] · Source xAI is expanding its footprint in enterprise workflows, making Grok models available on Databricks Agent Bricks to power corporate AI agents with proprietary data. Additionally, the company launched a native Grok plugin for Microsoft Word that allows users to draft documents, summarize web research, and generate diagrams directly within the editor.

[Cursor] · Source The AI code editor introduced /automate, a skill that allows agents to set up developer automations based on plain language prompts. The feature supports kicking off runs via Slack emoji reactions, GitHub triggers for issues and reviews, and includes computer use capabilities for cloud agents.

[Google DeepMind] · Source Anticipating the global scale-up of multi-agent systems, Google DeepMind published its AI Control Roadmap, a framework of structural security protocols used internally at Google. The research notes that the vast majority of agent issues stem from misinterpreted commands or over-enthusiasm rather than malicious intent, emphasizing the need for embedded security constraints.

[AWS] · Source AWS highlighted the enterprise capability of its agentic AI via AWS Transform, showcasing how Bridgestone successfully modernized a legacy mainframe in just seven months. The system autonomously evaluated 1.2 million lines of code to replatform critical applications with human oversight, yielding 90% efficiency gains.

[poolside] · Source AI coding startup poolside released the weights for Laguna M.1, its most capable model to date. Released under the Apache 2.0 license on Hugging Face, the model features a 256K context length and is available in both base and post-trained checkpoints for the open-source community.

[Anthropic] · Source Anthropic shared Phase 2 of Project Fetch, testing the robotics programming capabilities of the Claude family. A team with zero robotics expertise used Opus 4.7 to program a robodog, completing the task roughly 20 times faster than the best human team utilizing Opus 4.1 the previous year.

Also Noted#

• [Retool] (Source): Shipped a governed runtime update to manage the proliferation of “vibe-coded” applications, addressing alarm from 93% of surveyed technical executives.
• [Supabase] (Source): The open-source backend platform surpassed 10 million developers, driven by a hyper-growth phase that added 3 million new users in just the past three months.
• [Ploy] (Source): Launched out of stealth with a $27M seed round led by Y Combinator and First Round to turn websites into autonomous, all-in-one marketing platforms.
• [Liquid AI] (Source): Introduced LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M, boasting ultra-fast multilingual retrieval across 11 languages with enterprise latency as low as 1.5ms.