Jason Says
Cloudflare's security audit skill is the real deal—not a demo but a production-tested agent skill, finally giving the Skills ecosystem something with genuine teeth. Meanwhile, Astra for Law signals OpenAI is treating every high-value vertical as its own product line, bad news for legal SaaS incumbents.
OpenAI launches Astra for Law, a purpose-built AI platform offering frontier models, custom firm workflows, connected legal data sources, and confidentiality-grade controls. A significant vertical push into legal tech that will pressure incumbents like Harvey and Clio.
Cloudflare open-sources a coding-agent skill that orchestrates multi-phase security audits—recon, coverage-led hunting, candidate validation, and machine-readable findings—used internally in their own vulnerability discovery harness. A rare production-grade security skill for the agent ecosystem.
As AI agents handle longer, faster, higher-volume tasks than humans can review, companies are turning to AI supervisor agents as the oversight layer. This marks a fundamental infrastructure shift from 'human-in-the-loop' to 'AI-in-the-loop,' raising new questions about accountability chains.
Google DeepMind launches a dedicated institute to take big AGI questions—definitions, timelines, risks, governance—into broader public discourse. A strategic move to shape the AGI narrative and gain societal legitimacy for DeepMind's research direction.
Voicebox is an open-source, fully local AI voice studio supporting voice cloning, speech generation, dictation into any app, and agent conversations with owned voices. Positioned as a privacy-first, zero-cost alternative to ElevenLabs—highly relevant for indie devs and SaaS builders.
CERA-MoA fixes a critical flaw in Mixture-of-Agents systems: routing strategies and agent fine-tuning have been siloed, preventing adaptive specialization. By co-evolving the router with agents iteratively, it unlocks synergistic performance gains—directly relevant for anyone building multi-agent pipelines.
Edge0 introduces a 'prerouter' that predicts which MoE experts layer N+1 needs before layer N finishes, hiding SSD read latency behind compute. This makes local inference of 35B+ MoE models on consumer hardware genuinely practical—a game-changer for indie devs running models locally.
Crusoe closes $3.9B at a $30.9B valuation to build both massive data centers and small modular AI factories. The dual strategy targets the full compute stack from hyperscale cloud to edge inference, signaling continued accelerating demand for dedicated AI infrastructure.
Investors: Roo Capital、Grand Ventures
AI 驱动的网络安全与合规平台，主打持续 Agent 化的安全审计与合规自动化，目标替代传统人工合规流程。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free